fs/btrfs/extent_io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include <linux/bitops.h>
   4 #include <linux/slab.h>
   5 #include <linux/bio.h>
   6 #include <linux/mm.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/page-flags.h>
   9 #include <linux/spinlock.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/swap.h>
  12 #include <linux/writeback.h>
  13 #include <linux/pagevec.h>
  14 #include <linux/prefetch.h>
  15 #include <linux/cleancache.h>
  16 #include "extent_io.h"
  17 #include "extent_map.h"
  18 #include "ctree.h"
  19 #include "btrfs_inode.h"
  20 #include "volumes.h"
  21 #include "check-integrity.h"
  22 #include "locking.h"
  23 #include "rcu-string.h"
  24 #include "backref.h"
  25 #include "disk-io.h"
  26
  27 static struct kmem_cache *extent_state_cache;
  28 static struct kmem_cache *extent_buffer_cache;
  29 static struct bio_set btrfs_bioset;
  30
  31 static inline bool extent_state_in_tree(const struct extent_state *state)
  32 {
  33         return !RB_EMPTY_NODE(&state->rb_node);
  34 }
  35
  36 #ifdef CONFIG_BTRFS_DEBUG
  37 static LIST_HEAD(buffers);
  38 static LIST_HEAD(states);
  39
  40 static DEFINE_SPINLOCK(leak_lock);
  41
  42 static inline
  43 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  44 {
  45         unsigned long flags;
  46
  47         spin_lock_irqsave(&leak_lock, flags);
  48         list_add(new, head);
  49         spin_unlock_irqrestore(&leak_lock, flags);
  50 }
  51
  52 static inline
  53 void btrfs_leak_debug_del(struct list_head *entry)
  54 {
  55         unsigned long flags;
  56
  57         spin_lock_irqsave(&leak_lock, flags);
  58         list_del(entry);
  59         spin_unlock_irqrestore(&leak_lock, flags);
  60 }
  61
  62 static inline
  63 void btrfs_leak_debug_check(void)
  64 {
  65         struct extent_state *state;
  66         struct extent_buffer *eb;
  67
  68         while (!list_empty(&states)) {
  69                 state = list_entry(states.next, struct extent_state, leak_list);
  70                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  71                        state->start, state->end, state->state,
  72                        extent_state_in_tree(state),
  73                        refcount_read(&state->refs));
  74                 list_del(&state->leak_list);
  75                 kmem_cache_free(extent_state_cache, state);
  76         }
  77
  78         while (!list_empty(&buffers)) {
  79                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  80                 pr_err("BTRFS: buffer leak start %llu len %lu refs %d bflags %lu\n",
  81                        eb->start, eb->len, atomic_read(&eb->refs), eb->bflags);
  82                 list_del(&eb->leak_list);
  83                 kmem_cache_free(extent_buffer_cache, eb);
  84         }
  85 }
  86
  87 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  88         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  89 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  90                 struct extent_io_tree *tree, u64 start, u64 end)
  91 {
  92         struct inode *inode = tree->private_data;
  93         u64 isize;
  94
  95         if (!inode || !is_data_inode(inode))
  96                 return;
  97
  98         isize = i_size_read(inode);
  99         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
 100                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
 101                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
 102                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
 103         }
 104 }
 105 #else
 106 #define btrfs_leak_debug_add(new, head) do {} while (0)
 107 #define btrfs_leak_debug_del(entry)     do {} while (0)
 108 #define btrfs_leak_debug_check()        do {} while (0)
 109 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 110 #endif
 111
 112 #define BUFFER_LRU_MAX 64
 113
 114 struct tree_entry {
 115         u64 start;
 116         u64 end;
 117         struct rb_node rb_node;
 118 };
 119
 120 struct extent_page_data {
 121         struct bio *bio;
 122         struct extent_io_tree *tree;
 123         /* tells writepage not to lock the state bits for this range
 124          * it still does the unlocking
 125          */
 126         unsigned int extent_locked:1;
 127
 128         /* tells the submit_bio code to use REQ_SYNC */
 129         unsigned int sync_io:1;
 130 };
 131
 132 static int add_extent_changeset(struct extent_state *state, unsigned bits,
 133                                  struct extent_changeset *changeset,
 134                                  int set)
 135 {
 136         int ret;
 137
 138         if (!changeset)
 139                 return 0;
 140         if (set && (state->state & bits) == bits)
 141                 return 0;
 142         if (!set && (state->state & bits) == 0)
 143                 return 0;
 144         changeset->bytes_changed += state->end - state->start + 1;
 145         ret = ulist_add(&changeset->range_changed, state->start, state->end,
 146                         GFP_ATOMIC);
 147         return ret;
 148 }
 149
 150 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
 151                                        unsigned long bio_flags)
 152 {
 153         blk_status_t ret = 0;
 154         struct bio_vec *bvec = bio_last_bvec_all(bio);
 155         struct bio_vec bv;
 156         struct extent_io_tree *tree = bio->bi_private;
 157         u64 start;
 158
 159         mp_bvec_last_segment(bvec, &bv);
 160         start = page_offset(bv.bv_page) + bv.bv_offset;
 161
 162         bio->bi_private = NULL;
 163
 164         if (tree->ops)
 165                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
 166                                            mirror_num, bio_flags, start);
 167         else
 168                 btrfsic_submit_bio(bio);
 169
 170         return blk_status_to_errno(ret);
 171 }
 172
 173 static void flush_write_bio(struct extent_page_data *epd)
 174 {
 175         if (epd->bio) {
 176                 int ret;
 177
 178                 ret = submit_one_bio(epd->bio, 0, 0);
 179                 BUG_ON(ret < 0); /* -ENOMEM */
 180                 epd->bio = NULL;
 181         }
 182 }
 183
 184 int __init extent_io_init(void)
 185 {
 186         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 187                         sizeof(struct extent_state), 0,
 188                         SLAB_MEM_SPREAD, NULL);
 189         if (!extent_state_cache)
 190                 return -ENOMEM;
 191
 192         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 193                         sizeof(struct extent_buffer), 0,
 194                         SLAB_MEM_SPREAD, NULL);
 195         if (!extent_buffer_cache)
 196                 goto free_state_cache;
 197
 198         if (bioset_init(&btrfs_bioset, BIO_POOL_SIZE,
 199                         offsetof(struct btrfs_io_bio, bio),
 200                         BIOSET_NEED_BVECS))
 201                 goto free_buffer_cache;
 202
 203         if (bioset_integrity_create(&btrfs_bioset, BIO_POOL_SIZE))
 204                 goto free_bioset;
 205
 206         return 0;
 207
 208 free_bioset:
 209         bioset_exit(&btrfs_bioset);
 210
 211 free_buffer_cache:
 212         kmem_cache_destroy(extent_buffer_cache);
 213         extent_buffer_cache = NULL;
 214
 215 free_state_cache:
 216         kmem_cache_destroy(extent_state_cache);
 217         extent_state_cache = NULL;
 218         return -ENOMEM;
 219 }
 220
 221 void __cold extent_io_exit(void)
 222 {
 223         btrfs_leak_debug_check();
 224
 225         /*
 226          * Make sure all delayed rcu free are flushed before we
 227          * destroy caches.
 228          */
 229         rcu_barrier();
 230         kmem_cache_destroy(extent_state_cache);
 231         kmem_cache_destroy(extent_buffer_cache);
 232         bioset_exit(&btrfs_bioset);
 233 }
 234
 235 void extent_io_tree_init(struct extent_io_tree *tree,
 236                          void *private_data)
 237 {
 238         tree->state = RB_ROOT;
 239         tree->ops = NULL;
 240         tree->dirty_bytes = 0;
 241         spin_lock_init(&tree->lock);
 242         tree->private_data = private_data;
 243 }
 244
 245 static struct extent_state *alloc_extent_state(gfp_t mask)
 246 {
 247         struct extent_state *state;
 248
 249         /*
 250          * The given mask might be not appropriate for the slab allocator,
 251          * drop the unsupported bits
 252          */
 253         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
 254         state = kmem_cache_alloc(extent_state_cache, mask);
 255         if (!state)
 256                 return state;
 257         state->state = 0;
 258         state->failrec = NULL;
 259         RB_CLEAR_NODE(&state->rb_node);
 260         btrfs_leak_debug_add(&state->leak_list, &states);
 261         refcount_set(&state->refs, 1);
 262         init_waitqueue_head(&state->wq);
 263         trace_alloc_extent_state(state, mask, _RET_IP_);
 264         return state;
 265 }
 266
 267 void free_extent_state(struct extent_state *state)
 268 {
 269         if (!state)
 270                 return;
 271         if (refcount_dec_and_test(&state->refs)) {
 272                 WARN_ON(extent_state_in_tree(state));
 273                 btrfs_leak_debug_del(&state->leak_list);
 274                 trace_free_extent_state(state, _RET_IP_);
 275                 kmem_cache_free(extent_state_cache, state);
 276         }
 277 }
 278
 279 static struct rb_node *tree_insert(struct rb_root *root,
 280                                    struct rb_node *search_start,
 281                                    u64 offset,
 282                                    struct rb_node *node,
 283                                    struct rb_node ***p_in,
 284                                    struct rb_node **parent_in)
 285 {
 286         struct rb_node **p;
 287         struct rb_node *parent = NULL;
 288         struct tree_entry *entry;
 289
 290         if (p_in && parent_in) {
 291                 p = *p_in;
 292                 parent = *parent_in;
 293                 goto do_insert;
 294         }
 295
 296         p = search_start ? &search_start : &root->rb_node;
 297         while (*p) {
 298                 parent = *p;
 299                 entry = rb_entry(parent, struct tree_entry, rb_node);
 300
 301                 if (offset < entry->start)
 302                         p = &(*p)->rb_left;
 303                 else if (offset > entry->end)
 304                         p = &(*p)->rb_right;
 305                 else
 306                         return parent;
 307         }
 308
 309 do_insert:
 310         rb_link_node(node, parent, p);
 311         rb_insert_color(node, root);
 312         return NULL;
 313 }
 314
 315 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 316                                       struct rb_node **next_ret,
 317                                       struct rb_node **prev_ret,
 318                                       struct rb_node ***p_ret,
 319                                       struct rb_node **parent_ret)
 320 {
 321         struct rb_root *root = &tree->state;
 322         struct rb_node **n = &root->rb_node;
 323         struct rb_node *prev = NULL;
 324         struct rb_node *orig_prev = NULL;
 325         struct tree_entry *entry;
 326         struct tree_entry *prev_entry = NULL;
 327
 328         while (*n) {
 329                 prev = *n;
 330                 entry = rb_entry(prev, struct tree_entry, rb_node);
 331                 prev_entry = entry;
 332
 333                 if (offset < entry->start)
 334                         n = &(*n)->rb_left;
 335                 else if (offset > entry->end)
 336                         n = &(*n)->rb_right;
 337                 else
 338                         return *n;
 339         }
 340
 341         if (p_ret)
 342                 *p_ret = n;
 343         if (parent_ret)
 344                 *parent_ret = prev;
 345
 346         if (next_ret) {
 347                 orig_prev = prev;
 348                 while (prev && offset > prev_entry->end) {
 349                         prev = rb_next(prev);
 350                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 351                 }
 352                 *next_ret = prev;
 353                 prev = orig_prev;
 354         }
 355
 356         if (prev_ret) {
 357                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 358                 while (prev && offset < prev_entry->start) {
 359                         prev = rb_prev(prev);
 360                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 361                 }
 362                 *prev_ret = prev;
 363         }
 364         return NULL;
 365 }
 366
 367 static inline struct rb_node *
 368 tree_search_for_insert(struct extent_io_tree *tree,
 369                        u64 offset,
 370                        struct rb_node ***p_ret,
 371                        struct rb_node **parent_ret)
 372 {
 373         struct rb_node *next= NULL;
 374         struct rb_node *ret;
 375
 376         ret = __etree_search(tree, offset, &next, NULL, p_ret, parent_ret);
 377         if (!ret)
 378                 return next;
 379         return ret;
 380 }
 381
 382 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 383                                           u64 offset)
 384 {
 385         return tree_search_for_insert(tree, offset, NULL, NULL);
 386 }
 387
 388 /*
 389  * utility function to look for merge candidates inside a given range.
 390  * Any extents with matching state are merged together into a single
 391  * extent in the tree.  Extents with EXTENT_IO in their state field
 392  * are not merged because the end_io handlers need to be able to do
 393  * operations on them without sleeping (or doing allocations/splits).
 394  *
 395  * This should be called with the tree lock held.
 396  */
 397 static void merge_state(struct extent_io_tree *tree,
 398                         struct extent_state *state)
 399 {
 400         struct extent_state *other;
 401         struct rb_node *other_node;
 402
 403         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 404                 return;
 405
 406         other_node = rb_prev(&state->rb_node);
 407         if (other_node) {
 408                 other = rb_entry(other_node, struct extent_state, rb_node);
 409                 if (other->end == state->start - 1 &&
 410                     other->state == state->state) {
 411                         if (tree->private_data &&
 412                             is_data_inode(tree->private_data))
 413                                 btrfs_merge_delalloc_extent(tree->private_data,
 414                                                             state, other);
 415                         state->start = other->start;
 416                         rb_erase(&other->rb_node, &tree->state);
 417                         RB_CLEAR_NODE(&other->rb_node);
 418                         free_extent_state(other);
 419                 }
 420         }
 421         other_node = rb_next(&state->rb_node);
 422         if (other_node) {
 423                 other = rb_entry(other_node, struct extent_state, rb_node);
 424                 if (other->start == state->end + 1 &&
 425                     other->state == state->state) {
 426                         if (tree->private_data &&
 427                             is_data_inode(tree->private_data))
 428                                 btrfs_merge_delalloc_extent(tree->private_data,
 429                                                             state, other);
 430                         state->end = other->end;
 431                         rb_erase(&other->rb_node, &tree->state);
 432                         RB_CLEAR_NODE(&other->rb_node);
 433                         free_extent_state(other);
 434                 }
 435         }
 436 }
 437
 438 static void set_state_bits(struct extent_io_tree *tree,
 439                            struct extent_state *state, unsigned *bits,
 440                            struct extent_changeset *changeset);
 441
 442 /*
 443  * insert an extent_state struct into the tree.  'bits' are set on the
 444  * struct before it is inserted.
 445  *
 446  * This may return -EEXIST if the extent is already there, in which case the
 447  * state struct is freed.
 448  *
 449  * The tree lock is not taken internally.  This is a utility function and
 450  * probably isn't what you want to call (see set/clear_extent_bit).
 451  */
 452 static int insert_state(struct extent_io_tree *tree,
 453                         struct extent_state *state, u64 start, u64 end,
 454                         struct rb_node ***p,
 455                         struct rb_node **parent,
 456                         unsigned *bits, struct extent_changeset *changeset)
 457 {
 458         struct rb_node *node;
 459
 460         if (end < start)
 461                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 462                        end, start);
 463         state->start = start;
 464         state->end = end;
 465
 466         set_state_bits(tree, state, bits, changeset);
 467
 468         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 469         if (node) {
 470                 struct extent_state *found;
 471                 found = rb_entry(node, struct extent_state, rb_node);
 472                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
 473                        found->start, found->end, start, end);
 474                 return -EEXIST;
 475         }
 476         merge_state(tree, state);
 477         return 0;
 478 }
 479
 480 /*
 481  * split a given extent state struct in two, inserting the preallocated
 482  * struct 'prealloc' as the newly created second half.  'split' indicates an
 483  * offset inside 'orig' where it should be split.
 484  *
 485  * Before calling,
 486  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 487  * are two extent state structs in the tree:
 488  * prealloc: [orig->start, split - 1]
 489  * orig: [ split, orig->end ]
 490  *
 491  * The tree locks are not taken by this function. They need to be held
 492  * by the caller.
 493  */
 494 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 495                        struct extent_state *prealloc, u64 split)
 496 {
 497         struct rb_node *node;
 498
 499         if (tree->private_data && is_data_inode(tree->private_data))
 500                 btrfs_split_delalloc_extent(tree->private_data, orig, split);
 501
 502         prealloc->start = orig->start;
 503         prealloc->end = split - 1;
 504         prealloc->state = orig->state;
 505         orig->start = split;
 506
 507         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 508                            &prealloc->rb_node, NULL, NULL);
 509         if (node) {
 510                 free_extent_state(prealloc);
 511                 return -EEXIST;
 512         }
 513         return 0;
 514 }
 515
 516 static struct extent_state *next_state(struct extent_state *state)
 517 {
 518         struct rb_node *next = rb_next(&state->rb_node);
 519         if (next)
 520                 return rb_entry(next, struct extent_state, rb_node);
 521         else
 522                 return NULL;
 523 }
 524
 525 /*
 526  * utility function to clear some bits in an extent state struct.
 527  * it will optionally wake up anyone waiting on this state (wake == 1).
 528  *
 529  * If no bits are set on the state struct after clearing things, the
 530  * struct is freed and removed from the tree
 531  */
 532 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 533                                             struct extent_state *state,
 534                                             unsigned *bits, int wake,
 535                                             struct extent_changeset *changeset)
 536 {
 537         struct extent_state *next;
 538         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 539         int ret;
 540
 541         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 542                 u64 range = state->end - state->start + 1;
 543                 WARN_ON(range > tree->dirty_bytes);
 544                 tree->dirty_bytes -= range;
 545         }
 546
 547         if (tree->private_data && is_data_inode(tree->private_data))
 548                 btrfs_clear_delalloc_extent(tree->private_data, state, bits);
 549
 550         ret = add_extent_changeset(state, bits_to_clear, changeset, 0);
 551         BUG_ON(ret < 0);
 552         state->state &= ~bits_to_clear;
 553         if (wake)
 554                 wake_up(&state->wq);
 555         if (state->state == 0) {
 556                 next = next_state(state);
 557                 if (extent_state_in_tree(state)) {
 558                         rb_erase(&state->rb_node, &tree->state);
 559                         RB_CLEAR_NODE(&state->rb_node);
 560                         free_extent_state(state);
 561                 } else {
 562                         WARN_ON(1);
 563                 }
 564         } else {
 565                 merge_state(tree, state);
 566                 next = next_state(state);
 567         }
 568         return next;
 569 }
 570
 571 static struct extent_state *
 572 alloc_extent_state_atomic(struct extent_state *prealloc)
 573 {
 574         if (!prealloc)
 575                 prealloc = alloc_extent_state(GFP_ATOMIC);
 576
 577         return prealloc;
 578 }
 579
 580 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 581 {
 582         struct inode *inode = tree->private_data;
 583
 584         btrfs_panic(btrfs_sb(inode->i_sb), err,
 585         "locking error: extent tree was modified by another thread while locked");
 586 }
 587
 588 /*
 589  * clear some bits on a range in the tree.  This may require splitting
 590  * or inserting elements in the tree, so the gfp mask is used to
 591  * indicate which allocations or sleeping are allowed.
 592  *
 593  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 594  * the given range from the tree regardless of state (ie for truncate).
 595  *
 596  * the range [start, end] is inclusive.
 597  *
 598  * This takes the tree lock, and returns 0 on success and < 0 on error.
 599  */
 600 int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 601                               unsigned bits, int wake, int delete,
 602                               struct extent_state **cached_state,
 603                               gfp_t mask, struct extent_changeset *changeset)
 604 {
 605         struct extent_state *state;
 606         struct extent_state *cached;
 607         struct extent_state *prealloc = NULL;
 608         struct rb_node *node;
 609         u64 last_end;
 610         int err;
 611         int clear = 0;
 612
 613         btrfs_debug_check_extent_io_range(tree, start, end);
 614
 615         if (bits & EXTENT_DELALLOC)
 616                 bits |= EXTENT_NORESERVE;
 617
 618         if (delete)
 619                 bits |= ~EXTENT_CTLBITS;
 620
 621         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 622                 clear = 1;
 623 again:
 624         if (!prealloc && gfpflags_allow_blocking(mask)) {
 625                 /*
 626                  * Don't care for allocation failure here because we might end
 627                  * up not needing the pre-allocated extent state at all, which
 628                  * is the case if we only have in the tree extent states that
 629                  * cover our input range and don't cover too any other range.
 630                  * If we end up needing a new extent state we allocate it later.
 631                  */
 632                 prealloc = alloc_extent_state(mask);
 633         }
 634
 635         spin_lock(&tree->lock);
 636         if (cached_state) {
 637                 cached = *cached_state;
 638
 639                 if (clear) {
 640                         *cached_state = NULL;
 641                         cached_state = NULL;
 642                 }
 643
 644                 if (cached && extent_state_in_tree(cached) &&
 645                     cached->start <= start && cached->end > start) {
 646                         if (clear)
 647                                 refcount_dec(&cached->refs);
 648                         state = cached;
 649                         goto hit_next;
 650                 }
 651                 if (clear)
 652                         free_extent_state(cached);
 653         }
 654         /*
 655          * this search will find the extents that end after
 656          * our range starts
 657          */
 658         node = tree_search(tree, start);
 659         if (!node)
 660                 goto out;
 661         state = rb_entry(node, struct extent_state, rb_node);
 662 hit_next:
 663         if (state->start > end)
 664                 goto out;
 665         WARN_ON(state->end < start);
 666         last_end = state->end;
 667
 668         /* the state doesn't have the wanted bits, go ahead */
 669         if (!(state->state & bits)) {
 670                 state = next_state(state);
 671                 goto next;
 672         }
 673
 674         /*
 675          *     | ---- desired range ---- |
 676          *  | state | or
 677          *  | ------------- state -------------- |
 678          *
 679          * We need to split the extent we found, and may flip
 680          * bits on second half.
 681          *
 682          * If the extent we found extends past our range, we
 683          * just split and search again.  It'll get split again
 684          * the next time though.
 685          *
 686          * If the extent we found is inside our range, we clear
 687          * the desired bit on it.
 688          */
 689
 690         if (state->start < start) {
 691                 prealloc = alloc_extent_state_atomic(prealloc);
 692                 BUG_ON(!prealloc);
 693                 err = split_state(tree, state, prealloc, start);
 694                 if (err)
 695                         extent_io_tree_panic(tree, err);
 696
 697                 prealloc = NULL;
 698                 if (err)
 699                         goto out;
 700                 if (state->end <= end) {
 701                         state = clear_state_bit(tree, state, &bits, wake,
 702                                                 changeset);
 703                         goto next;
 704                 }
 705                 goto search_again;
 706         }
 707         /*
 708          * | ---- desired range ---- |
 709          *                        | state |
 710          * We need to split the extent, and clear the bit
 711          * on the first half
 712          */
 713         if (state->start <= end && state->end > end) {
 714                 prealloc = alloc_extent_state_atomic(prealloc);
 715                 BUG_ON(!prealloc);
 716                 err = split_state(tree, state, prealloc, end + 1);
 717                 if (err)
 718                         extent_io_tree_panic(tree, err);
 719
 720                 if (wake)
 721                         wake_up(&state->wq);
 722
 723                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 724
 725                 prealloc = NULL;
 726                 goto out;
 727         }
 728
 729         state = clear_state_bit(tree, state, &bits, wake, changeset);
 730 next:
 731         if (last_end == (u64)-1)
 732                 goto out;
 733         start = last_end + 1;
 734         if (start <= end && state && !need_resched())
 735                 goto hit_next;
 736
 737 search_again:
 738         if (start > end)
 739                 goto out;
 740         spin_unlock(&tree->lock);
 741         if (gfpflags_allow_blocking(mask))
 742                 cond_resched();
 743         goto again;
 744
 745 out:
 746         spin_unlock(&tree->lock);
 747         if (prealloc)
 748                 free_extent_state(prealloc);
 749
 750         return 0;
 751
 752 }
 753
 754 static void wait_on_state(struct extent_io_tree *tree,
 755                           struct extent_state *state)
 756                 __releases(tree->lock)
 757                 __acquires(tree->lock)
 758 {
 759         DEFINE_WAIT(wait);
 760         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 761         spin_unlock(&tree->lock);
 762         schedule();
 763         spin_lock(&tree->lock);
 764         finish_wait(&state->wq, &wait);
 765 }
 766
 767 /*
 768  * waits for one or more bits to clear on a range in the state tree.
 769  * The range [start, end] is inclusive.
 770  * The tree lock is taken by this function
 771  */
 772 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 773                             unsigned long bits)
 774 {
 775         struct extent_state *state;
 776         struct rb_node *node;
 777
 778         btrfs_debug_check_extent_io_range(tree, start, end);
 779
 780         spin_lock(&tree->lock);
 781 again:
 782         while (1) {
 783                 /*
 784                  * this search will find all the extents that end after
 785                  * our range starts
 786                  */
 787                 node = tree_search(tree, start);
 788 process_node:
 789                 if (!node)
 790                         break;
 791
 792                 state = rb_entry(node, struct extent_state, rb_node);
 793
 794                 if (state->start > end)
 795                         goto out;
 796
 797                 if (state->state & bits) {
 798                         start = state->start;
 799                         refcount_inc(&state->refs);
 800                         wait_on_state(tree, state);
 801                         free_extent_state(state);
 802                         goto again;
 803                 }
 804                 start = state->end + 1;
 805
 806                 if (start > end)
 807                         break;
 808
 809                 if (!cond_resched_lock(&tree->lock)) {
 810                         node = rb_next(node);
 811                         goto process_node;
 812                 }
 813         }
 814 out:
 815         spin_unlock(&tree->lock);
 816 }
 817
 818 static void set_state_bits(struct extent_io_tree *tree,
 819                            struct extent_state *state,
 820                            unsigned *bits, struct extent_changeset *changeset)
 821 {
 822         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 823         int ret;
 824
 825         if (tree->private_data && is_data_inode(tree->private_data))
 826                 btrfs_set_delalloc_extent(tree->private_data, state, bits);
 827
 828         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 829                 u64 range = state->end - state->start + 1;
 830                 tree->dirty_bytes += range;
 831         }
 832         ret = add_extent_changeset(state, bits_to_set, changeset, 1);
 833         BUG_ON(ret < 0);
 834         state->state |= bits_to_set;
 835 }
 836
 837 static void cache_state_if_flags(struct extent_state *state,
 838                                  struct extent_state **cached_ptr,
 839                                  unsigned flags)
 840 {
 841         if (cached_ptr && !(*cached_ptr)) {
 842                 if (!flags || (state->state & flags)) {
 843                         *cached_ptr = state;
 844                         refcount_inc(&state->refs);
 845                 }
 846         }
 847 }
 848
 849 static void cache_state(struct extent_state *state,
 850                         struct extent_state **cached_ptr)
 851 {
 852         return cache_state_if_flags(state, cached_ptr,
 853                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 854 }
 855
 856 /*
 857  * set some bits on a range in the tree.  This may require allocations or
 858  * sleeping, so the gfp mask is used to indicate what is allowed.
 859  *
 860  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 861  * part of the range already has the desired bits set.  The start of the
 862  * existing range is returned in failed_start in this case.
 863  *
 864  * [start, end] is inclusive This takes the tree lock.
 865  */
 866
 867 static int __must_check
 868 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 869                  unsigned bits, unsigned exclusive_bits,
 870                  u64 *failed_start, struct extent_state **cached_state,
 871                  gfp_t mask, struct extent_changeset *changeset)
 872 {
 873         struct extent_state *state;
 874         struct extent_state *prealloc = NULL;
 875         struct rb_node *node;
 876         struct rb_node **p;
 877         struct rb_node *parent;
 878         int err = 0;
 879         u64 last_start;
 880         u64 last_end;
 881
 882         btrfs_debug_check_extent_io_range(tree, start, end);
 883
 884 again:
 885         if (!prealloc && gfpflags_allow_blocking(mask)) {
 886                 /*
 887                  * Don't care for allocation failure here because we might end
 888                  * up not needing the pre-allocated extent state at all, which
 889                  * is the case if we only have in the tree extent states that
 890                  * cover our input range and don't cover too any other range.
 891                  * If we end up needing a new extent state we allocate it later.
 892                  */
 893                 prealloc = alloc_extent_state(mask);
 894         }
 895
 896         spin_lock(&tree->lock);
 897         if (cached_state && *cached_state) {
 898                 state = *cached_state;
 899                 if (state->start <= start && state->end > start &&
 900                     extent_state_in_tree(state)) {
 901                         node = &state->rb_node;
 902                         goto hit_next;
 903                 }
 904         }
 905         /*
 906          * this search will find all the extents that end after
 907          * our range starts.
 908          */
 909         node = tree_search_for_insert(tree, start, &p, &parent);
 910         if (!node) {
 911                 prealloc = alloc_extent_state_atomic(prealloc);
 912                 BUG_ON(!prealloc);
 913                 err = insert_state(tree, prealloc, start, end,
 914                                    &p, &parent, &bits, changeset);
 915                 if (err)
 916                         extent_io_tree_panic(tree, err);
 917
 918                 cache_state(prealloc, cached_state);
 919                 prealloc = NULL;
 920                 goto out;
 921         }
 922         state = rb_entry(node, struct extent_state, rb_node);
 923 hit_next:
 924         last_start = state->start;
 925         last_end = state->end;
 926
 927         /*
 928          * | ---- desired range ---- |
 929          * | state |
 930          *
 931          * Just lock what we found and keep going
 932          */
 933         if (state->start == start && state->end <= end) {
 934                 if (state->state & exclusive_bits) {
 935                         *failed_start = state->start;
 936                         err = -EEXIST;
 937                         goto out;
 938                 }
 939
 940                 set_state_bits(tree, state, &bits, changeset);
 941                 cache_state(state, cached_state);
 942                 merge_state(tree, state);
 943                 if (last_end == (u64)-1)
 944                         goto out;
 945                 start = last_end + 1;
 946                 state = next_state(state);
 947                 if (start < end && state && state->start == start &&
 948                     !need_resched())
 949                         goto hit_next;
 950                 goto search_again;
 951         }
 952
 953         /*
 954          *     | ---- desired range ---- |
 955          * | state |
 956          *   or
 957          * | ------------- state -------------- |
 958          *
 959          * We need to split the extent we found, and may flip bits on
 960          * second half.
 961          *
 962          * If the extent we found extends past our
 963          * range, we just split and search again.  It'll get split
 964          * again the next time though.
 965          *
 966          * If the extent we found is inside our range, we set the
 967          * desired bit on it.
 968          */
 969         if (state->start < start) {
 970                 if (state->state & exclusive_bits) {
 971                         *failed_start = start;
 972                         err = -EEXIST;
 973                         goto out;
 974                 }
 975
 976                 prealloc = alloc_extent_state_atomic(prealloc);
 977                 BUG_ON(!prealloc);
 978                 err = split_state(tree, state, prealloc, start);
 979                 if (err)
 980                         extent_io_tree_panic(tree, err);
 981
 982                 prealloc = NULL;
 983                 if (err)
 984                         goto out;
 985                 if (state->end <= end) {
 986                         set_state_bits(tree, state, &bits, changeset);
 987                         cache_state(state, cached_state);
 988                         merge_state(tree, state);
 989                         if (last_end == (u64)-1)
 990                                 goto out;
 991                         start = last_end + 1;
 992                         state = next_state(state);
 993                         if (start < end && state && state->start == start &&
 994                             !need_resched())
 995                                 goto hit_next;
 996                 }
 997                 goto search_again;
 998         }
 999         /*
1000          * | ---- desired range ---- |
1001          *     | state | or               | state |
1002          *
1003          * There's a hole, we need to insert something in it and
1004          * ignore the extent we found.
1005          */
1006         if (state->start > start) {
1007                 u64 this_end;
1008                 if (end < last_start)
1009                         this_end = end;
1010                 else
1011                         this_end = last_start - 1;
1012
1013                 prealloc = alloc_extent_state_atomic(prealloc);
1014                 BUG_ON(!prealloc);
1015
1016                 /*
1017                  * Avoid to free 'prealloc' if it can be merged with
1018                  * the later extent.
1019                  */
1020                 err = insert_state(tree, prealloc, start, this_end,
1021                                    NULL, NULL, &bits, changeset);
1022                 if (err)
1023                         extent_io_tree_panic(tree, err);
1024
1025                 cache_state(prealloc, cached_state);
1026                 prealloc = NULL;
1027                 start = this_end + 1;
1028                 goto search_again;
1029         }
1030         /*
1031          * | ---- desired range ---- |
1032          *                        | state |
1033          * We need to split the extent, and set the bit
1034          * on the first half
1035          */
1036         if (state->start <= end && state->end > end) {
1037                 if (state->state & exclusive_bits) {
1038                         *failed_start = start;
1039                         err = -EEXIST;
1040                         goto out;
1041                 }
1042
1043                 prealloc = alloc_extent_state_atomic(prealloc);
1044                 BUG_ON(!prealloc);
1045                 err = split_state(tree, state, prealloc, end + 1);
1046                 if (err)
1047                         extent_io_tree_panic(tree, err);
1048
1049                 set_state_bits(tree, prealloc, &bits, changeset);
1050                 cache_state(prealloc, cached_state);
1051                 merge_state(tree, prealloc);
1052                 prealloc = NULL;
1053                 goto out;
1054         }
1055
1056 search_again:
1057         if (start > end)
1058                 goto out;
1059         spin_unlock(&tree->lock);
1060         if (gfpflags_allow_blocking(mask))
1061                 cond_resched();
1062         goto again;
1063
1064 out:
1065         spin_unlock(&tree->lock);
1066         if (prealloc)
1067                 free_extent_state(prealloc);
1068
1069         return err;
1070
1071 }
1072
1073 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1074                    unsigned bits, u64 * failed_start,
1075                    struct extent_state **cached_state, gfp_t mask)
1076 {
1077         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1078                                 cached_state, mask, NULL);
1079 }
1080
1081
1082 /**
1083  * convert_extent_bit - convert all bits in a given range from one bit to
1084  *                      another
1085  * @tree:       the io tree to search
1086  * @start:      the start offset in bytes
1087  * @end:        the end offset in bytes (inclusive)
1088  * @bits:       the bits to set in this range
1089  * @clear_bits: the bits to clear in this range
1090  * @cached_state:       state that we're going to cache
1091  *
1092  * This will go through and set bits for the given range.  If any states exist
1093  * already in this range they are set with the given bit and cleared of the
1094  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1095  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1096  * boundary bits like LOCK.
1097  *
1098  * All allocations are done with GFP_NOFS.
1099  */
1100 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1101                        unsigned bits, unsigned clear_bits,
1102                        struct extent_state **cached_state)
1103 {
1104         struct extent_state *state;
1105         struct extent_state *prealloc = NULL;
1106         struct rb_node *node;
1107         struct rb_node **p;
1108         struct rb_node *parent;
1109         int err = 0;
1110         u64 last_start;
1111         u64 last_end;
1112         bool first_iteration = true;
1113
1114         btrfs_debug_check_extent_io_range(tree, start, end);
1115
1116 again:
1117         if (!prealloc) {
1118                 /*
1119                  * Best effort, don't worry if extent state allocation fails
1120                  * here for the first iteration. We might have a cached state
1121                  * that matches exactly the target range, in which case no
1122                  * extent state allocations are needed. We'll only know this
1123                  * after locking the tree.
1124                  */
1125                 prealloc = alloc_extent_state(GFP_NOFS);
1126                 if (!prealloc && !first_iteration)
1127                         return -ENOMEM;
1128         }
1129
1130         spin_lock(&tree->lock);
1131         if (cached_state && *cached_state) {
1132                 state = *cached_state;
1133                 if (state->start <= start && state->end > start &&
1134                     extent_state_in_tree(state)) {
1135                         node = &state->rb_node;
1136                         goto hit_next;
1137                 }
1138         }
1139
1140         /*
1141          * this search will find all the extents that end after
1142          * our range starts.
1143          */
1144         node = tree_search_for_insert(tree, start, &p, &parent);
1145         if (!node) {
1146                 prealloc = alloc_extent_state_atomic(prealloc);
1147                 if (!prealloc) {
1148                         err = -ENOMEM;
1149                         goto out;
1150                 }
1151                 err = insert_state(tree, prealloc, start, end,
1152                                    &p, &parent, &bits, NULL);
1153                 if (err)
1154                         extent_io_tree_panic(tree, err);
1155                 cache_state(prealloc, cached_state);
1156                 prealloc = NULL;
1157                 goto out;
1158         }
1159         state = rb_entry(node, struct extent_state, rb_node);
1160 hit_next:
1161         last_start = state->start;
1162         last_end = state->end;
1163
1164         /*
1165          * | ---- desired range ---- |
1166          * | state |
1167          *
1168          * Just lock what we found and keep going
1169          */
1170         if (state->start == start && state->end <= end) {
1171                 set_state_bits(tree, state, &bits, NULL);
1172                 cache_state(state, cached_state);
1173                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1174                 if (last_end == (u64)-1)
1175                         goto out;
1176                 start = last_end + 1;
1177                 if (start < end && state && state->start == start &&
1178                     !need_resched())
1179                         goto hit_next;
1180                 goto search_again;
1181         }
1182
1183         /*
1184          *     | ---- desired range ---- |
1185          * | state |
1186          *   or
1187          * | ------------- state -------------- |
1188          *
1189          * We need to split the extent we found, and may flip bits on
1190          * second half.
1191          *
1192          * If the extent we found extends past our
1193          * range, we just split and search again.  It'll get split
1194          * again the next time though.
1195          *
1196          * If the extent we found is inside our range, we set the
1197          * desired bit on it.
1198          */
1199         if (state->start < start) {
1200                 prealloc = alloc_extent_state_atomic(prealloc);
1201                 if (!prealloc) {
1202                         err = -ENOMEM;
1203                         goto out;
1204                 }
1205                 err = split_state(tree, state, prealloc, start);
1206                 if (err)
1207                         extent_io_tree_panic(tree, err);
1208                 prealloc = NULL;
1209                 if (err)
1210                         goto out;
1211                 if (state->end <= end) {
1212                         set_state_bits(tree, state, &bits, NULL);
1213                         cache_state(state, cached_state);
1214                         state = clear_state_bit(tree, state, &clear_bits, 0,
1215                                                 NULL);
1216                         if (last_end == (u64)-1)
1217                                 goto out;
1218                         start = last_end + 1;
1219                         if (start < end && state && state->start == start &&
1220                             !need_resched())
1221                                 goto hit_next;
1222                 }
1223                 goto search_again;
1224         }
1225         /*
1226          * | ---- desired range ---- |
1227          *     | state | or               | state |
1228          *
1229          * There's a hole, we need to insert something in it and
1230          * ignore the extent we found.
1231          */
1232         if (state->start > start) {
1233                 u64 this_end;
1234                 if (end < last_start)
1235                         this_end = end;
1236                 else
1237                         this_end = last_start - 1;
1238
1239                 prealloc = alloc_extent_state_atomic(prealloc);
1240                 if (!prealloc) {
1241                         err = -ENOMEM;
1242                         goto out;
1243                 }
1244
1245                 /*
1246                  * Avoid to free 'prealloc' if it can be merged with
1247                  * the later extent.
1248                  */
1249                 err = insert_state(tree, prealloc, start, this_end,
1250                                    NULL, NULL, &bits, NULL);
1251                 if (err)
1252                         extent_io_tree_panic(tree, err);
1253                 cache_state(prealloc, cached_state);
1254                 prealloc = NULL;
1255                 start = this_end + 1;
1256                 goto search_again;
1257         }
1258         /*
1259          * | ---- desired range ---- |
1260          *                        | state |
1261          * We need to split the extent, and set the bit
1262          * on the first half
1263          */
1264         if (state->start <= end && state->end > end) {
1265                 prealloc = alloc_extent_state_atomic(prealloc);
1266                 if (!prealloc) {
1267                         err = -ENOMEM;
1268                         goto out;
1269                 }
1270
1271                 err = split_state(tree, state, prealloc, end + 1);
1272                 if (err)
1273                         extent_io_tree_panic(tree, err);
1274
1275                 set_state_bits(tree, prealloc, &bits, NULL);
1276                 cache_state(prealloc, cached_state);
1277                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1278                 prealloc = NULL;
1279                 goto out;
1280         }
1281
1282 search_again:
1283         if (start > end)
1284                 goto out;
1285         spin_unlock(&tree->lock);
1286         cond_resched();
1287         first_iteration = false;
1288         goto again;
1289
1290 out:
1291         spin_unlock(&tree->lock);
1292         if (prealloc)
1293                 free_extent_state(prealloc);
1294
1295         return err;
1296 }
1297
1298 /* wrappers around set/clear extent bit */
1299 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1300                            unsigned bits, struct extent_changeset *changeset)
1301 {
1302         /*
1303          * We don't support EXTENT_LOCKED yet, as current changeset will
1304          * record any bits changed, so for EXTENT_LOCKED case, it will
1305          * either fail with -EEXIST or changeset will record the whole
1306          * range.
1307          */
1308         BUG_ON(bits & EXTENT_LOCKED);
1309
1310         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1311                                 changeset);
1312 }
1313
1314 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1315                      unsigned bits, int wake, int delete,
1316                      struct extent_state **cached)
1317 {
1318         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1319                                   cached, GFP_NOFS, NULL);
1320 }
1321
1322 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1323                 unsigned bits, struct extent_changeset *changeset)
1324 {
1325         /*
1326          * Don't support EXTENT_LOCKED case, same reason as
1327          * set_record_extent_bits().
1328          */
1329         BUG_ON(bits & EXTENT_LOCKED);
1330
1331         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1332                                   changeset);
1333 }
1334
1335 /*
1336  * either insert or lock state struct between start and end use mask to tell
1337  * us if waiting is desired.
1338  */
1339 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1340                      struct extent_state **cached_state)
1341 {
1342         int err;
1343         u64 failed_start;
1344
1345         while (1) {
1346                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1347                                        EXTENT_LOCKED, &failed_start,
1348                                        cached_state, GFP_NOFS, NULL);
1349                 if (err == -EEXIST) {
1350                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1351                         start = failed_start;
1352                 } else
1353                         break;
1354                 WARN_ON(start > end);
1355         }
1356         return err;
1357 }
1358
1359 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1360 {
1361         int err;
1362         u64 failed_start;
1363
1364         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1365                                &failed_start, NULL, GFP_NOFS, NULL);
1366         if (err == -EEXIST) {
1367                 if (failed_start > start)
1368                         clear_extent_bit(tree, start, failed_start - 1,
1369                                          EXTENT_LOCKED, 1, 0, NULL);
1370                 return 0;
1371         }
1372         return 1;
1373 }
1374
1375 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1376 {
1377         unsigned long index = start >> PAGE_SHIFT;
1378         unsigned long end_index = end >> PAGE_SHIFT;
1379         struct page *page;
1380
1381         while (index <= end_index) {
1382                 page = find_get_page(inode->i_mapping, index);
1383                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1384                 clear_page_dirty_for_io(page);
1385                 put_page(page);
1386                 index++;
1387         }
1388 }
1389
1390 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1391 {
1392         unsigned long index = start >> PAGE_SHIFT;
1393         unsigned long end_index = end >> PAGE_SHIFT;
1394         struct page *page;
1395
1396         while (index <= end_index) {
1397                 page = find_get_page(inode->i_mapping, index);
1398                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1399                 __set_page_dirty_nobuffers(page);
1400                 account_page_redirty(page);
1401                 put_page(page);
1402                 index++;
1403         }
1404 }
1405
1406 /* find the first state struct with 'bits' set after 'start', and
1407  * return it.  tree->lock must be held.  NULL will returned if
1408  * nothing was found after 'start'
1409  */
1410 static struct extent_state *
1411 find_first_extent_bit_state(struct extent_io_tree *tree,
1412                             u64 start, unsigned bits)
1413 {
1414         struct rb_node *node;
1415         struct extent_state *state;
1416
1417         /*
1418          * this search will find all the extents that end after
1419          * our range starts.
1420          */
1421         node = tree_search(tree, start);
1422         if (!node)
1423                 goto out;
1424
1425         while (1) {
1426                 state = rb_entry(node, struct extent_state, rb_node);
1427                 if (state->end >= start && (state->state & bits))
1428                         return state;
1429
1430                 node = rb_next(node);
1431                 if (!node)
1432                         break;
1433         }
1434 out:
1435         return NULL;
1436 }
1437
1438 /*
1439  * find the first offset in the io tree with 'bits' set. zero is
1440  * returned if we find something, and *start_ret and *end_ret are
1441  * set to reflect the state struct that was found.
1442  *
1443  * If nothing was found, 1 is returned. If found something, return 0.
1444  */
1445 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1446                           u64 *start_ret, u64 *end_ret, unsigned bits,
1447                           struct extent_state **cached_state)
1448 {
1449         struct extent_state *state;
1450         int ret = 1;
1451
1452         spin_lock(&tree->lock);
1453         if (cached_state && *cached_state) {
1454                 state = *cached_state;
1455                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1456                         while ((state = next_state(state)) != NULL) {
1457                                 if (state->state & bits)
1458                                         goto got_it;
1459                         }
1460                         free_extent_state(*cached_state);
1461                         *cached_state = NULL;
1462                         goto out;
1463                 }
1464                 free_extent_state(*cached_state);
1465                 *cached_state = NULL;
1466         }
1467
1468         state = find_first_extent_bit_state(tree, start, bits);
1469 got_it:
1470         if (state) {
1471                 cache_state_if_flags(state, cached_state, 0);
1472                 *start_ret = state->start;
1473                 *end_ret = state->end;
1474                 ret = 0;
1475         }
1476 out:
1477         spin_unlock(&tree->lock);
1478         return ret;
1479 }
1480
1481 /*
1482  * find a contiguous range of bytes in the file marked as delalloc, not
1483  * more than 'max_bytes'.  start and end are used to return the range,
1484  *
1485  * true is returned if we find something, false if nothing was in the tree
1486  */
1487 static noinline bool find_delalloc_range(struct extent_io_tree *tree,
1488                                         u64 *start, u64 *end, u64 max_bytes,
1489                                         struct extent_state **cached_state)
1490 {
1491         struct rb_node *node;
1492         struct extent_state *state;
1493         u64 cur_start = *start;
1494         bool found = false;
1495         u64 total_bytes = 0;
1496
1497         spin_lock(&tree->lock);
1498
1499         /*
1500          * this search will find all the extents that end after
1501          * our range starts.
1502          */
1503         node = tree_search(tree, cur_start);
1504         if (!node) {
1505                 *end = (u64)-1;
1506                 goto out;
1507         }
1508
1509         while (1) {
1510                 state = rb_entry(node, struct extent_state, rb_node);
1511                 if (found && (state->start != cur_start ||
1512                               (state->state & EXTENT_BOUNDARY))) {
1513                         goto out;
1514                 }
1515                 if (!(state->state & EXTENT_DELALLOC)) {
1516                         if (!found)
1517                                 *end = state->end;
1518                         goto out;
1519                 }
1520                 if (!found) {
1521                         *start = state->start;
1522                         *cached_state = state;
1523                         refcount_inc(&state->refs);
1524                 }
1525                 found = true;
1526                 *end = state->end;
1527                 cur_start = state->end + 1;
1528                 node = rb_next(node);
1529                 total_bytes += state->end - state->start + 1;
1530                 if (total_bytes >= max_bytes)
1531                         break;
1532                 if (!node)
1533                         break;
1534         }
1535 out:
1536         spin_unlock(&tree->lock);
1537         return found;
1538 }
1539
1540 static int __process_pages_contig(struct address_space *mapping,
1541                                   struct page *locked_page,
1542                                   pgoff_t start_index, pgoff_t end_index,
1543                                   unsigned long page_ops, pgoff_t *index_ret);
1544
1545 static noinline void __unlock_for_delalloc(struct inode *inode,
1546                                            struct page *locked_page,
1547                                            u64 start, u64 end)
1548 {
1549         unsigned long index = start >> PAGE_SHIFT;
1550         unsigned long end_index = end >> PAGE_SHIFT;
1551
1552         ASSERT(locked_page);
1553         if (index == locked_page->index && end_index == index)
1554                 return;
1555
1556         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1557                                PAGE_UNLOCK, NULL);
1558 }
1559
1560 static noinline int lock_delalloc_pages(struct inode *inode,
1561                                         struct page *locked_page,
1562                                         u64 delalloc_start,
1563                                         u64 delalloc_end)
1564 {
1565         unsigned long index = delalloc_start >> PAGE_SHIFT;
1566         unsigned long index_ret = index;
1567         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1568         int ret;
1569
1570         ASSERT(locked_page);
1571         if (index == locked_page->index && index == end_index)
1572                 return 0;
1573
1574         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1575                                      end_index, PAGE_LOCK, &index_ret);
1576         if (ret == -EAGAIN)
1577                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1578                                       (u64)index_ret << PAGE_SHIFT);
1579         return ret;
1580 }
1581
1582 /*
1583  * Find and lock a contiguous range of bytes in the file marked as delalloc, no
1584  * more than @max_bytes.  @Start and @end are used to return the range,
1585  *
1586  * Return: true if we find something
1587  *         false if nothing was in the tree
1588  */
1589 EXPORT_FOR_TESTS
1590 noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
1591                                     struct extent_io_tree *tree,
1592                                     struct page *locked_page, u64 *start,
1593                                     u64 *end)
1594 {
1595         u64 max_bytes = BTRFS_MAX_EXTENT_SIZE;
1596         u64 delalloc_start;
1597         u64 delalloc_end;
1598         bool found;
1599         struct extent_state *cached_state = NULL;
1600         int ret;
1601         int loops = 0;
1602
1603 again:
1604         /* step one, find a bunch of delalloc bytes starting at start */
1605         delalloc_start = *start;
1606         delalloc_end = 0;
1607         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1608                                     max_bytes, &cached_state);
1609         if (!found || delalloc_end <= *start) {
1610                 *start = delalloc_start;
1611                 *end = delalloc_end;
1612                 free_extent_state(cached_state);
1613                 return false;
1614         }
1615
1616         /*
1617          * start comes from the offset of locked_page.  We have to lock
1618          * pages in order, so we can't process delalloc bytes before
1619          * locked_page
1620          */
1621         if (delalloc_start < *start)
1622                 delalloc_start = *start;
1623
1624         /*
1625          * make sure to limit the number of pages we try to lock down
1626          */
1627         if (delalloc_end + 1 - delalloc_start > max_bytes)
1628                 delalloc_end = delalloc_start + max_bytes - 1;
1629
1630         /* step two, lock all the pages after the page that has start */
1631         ret = lock_delalloc_pages(inode, locked_page,
1632                                   delalloc_start, delalloc_end);
1633         ASSERT(!ret || ret == -EAGAIN);
1634         if (ret == -EAGAIN) {
1635                 /* some of the pages are gone, lets avoid looping by
1636                  * shortening the size of the delalloc range we're searching
1637                  */
1638                 free_extent_state(cached_state);
1639                 cached_state = NULL;
1640                 if (!loops) {
1641                         max_bytes = PAGE_SIZE;
1642                         loops = 1;
1643                         goto again;
1644                 } else {
1645                         found = false;
1646                         goto out_failed;
1647                 }
1648         }
1649
1650         /* step three, lock the state bits for the whole range */
1651         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1652
1653         /* then test to make sure it is all still delalloc */
1654         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1655                              EXTENT_DELALLOC, 1, cached_state);
1656         if (!ret) {
1657                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1658                                      &cached_state);
1659                 __unlock_for_delalloc(inode, locked_page,
1660                               delalloc_start, delalloc_end);
1661                 cond_resched();
1662                 goto again;
1663         }
1664         free_extent_state(cached_state);
1665         *start = delalloc_start;
1666         *end = delalloc_end;
1667 out_failed:
1668         return found;
1669 }
1670
1671 static int __process_pages_contig(struct address_space *mapping,
1672                                   struct page *locked_page,
1673                                   pgoff_t start_index, pgoff_t end_index,
1674                                   unsigned long page_ops, pgoff_t *index_ret)
1675 {
1676         unsigned long nr_pages = end_index - start_index + 1;
1677         unsigned long pages_locked = 0;
1678         pgoff_t index = start_index;
1679         struct page *pages[16];
1680         unsigned ret;
1681         int err = 0;
1682         int i;
1683
1684         if (page_ops & PAGE_LOCK) {
1685                 ASSERT(page_ops == PAGE_LOCK);
1686                 ASSERT(index_ret && *index_ret == start_index);
1687         }
1688
1689         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1690                 mapping_set_error(mapping, -EIO);
1691
1692         while (nr_pages > 0) {
1693                 ret = find_get_pages_contig(mapping, index,
1694                                      min_t(unsigned long,
1695                                      nr_pages, ARRAY_SIZE(pages)), pages);
1696                 if (ret == 0) {
1697                         /*
1698                          * Only if we're going to lock these pages,
1699                          * can we find nothing at @index.
1700                          */
1701                         ASSERT(page_ops & PAGE_LOCK);
1702                         err = -EAGAIN;
1703                         goto out;
1704                 }
1705
1706                 for (i = 0; i < ret; i++) {
1707                         if (page_ops & PAGE_SET_PRIVATE2)
1708                                 SetPagePrivate2(pages[i]);
1709
1710                         if (pages[i] == locked_page) {
1711                                 put_page(pages[i]);
1712                                 pages_locked++;
1713                                 continue;
1714                         }
1715                         if (page_ops & PAGE_CLEAR_DIRTY)
1716                                 clear_page_dirty_for_io(pages[i]);
1717                         if (page_ops & PAGE_SET_WRITEBACK)
1718                                 set_page_writeback(pages[i]);
1719                         if (page_ops & PAGE_SET_ERROR)
1720                                 SetPageError(pages[i]);
1721                         if (page_ops & PAGE_END_WRITEBACK)
1722                                 end_page_writeback(pages[i]);
1723                         if (page_ops & PAGE_UNLOCK)
1724                                 unlock_page(pages[i]);
1725                         if (page_ops & PAGE_LOCK) {
1726                                 lock_page(pages[i]);
1727                                 if (!PageDirty(pages[i]) ||
1728                                     pages[i]->mapping != mapping) {
1729                                         unlock_page(pages[i]);
1730                                         put_page(pages[i]);
1731                                         err = -EAGAIN;
1732                                         goto out;
1733                                 }
1734                         }
1735                         put_page(pages[i]);
1736                         pages_locked++;
1737                 }
1738                 nr_pages -= ret;
1739                 index += ret;
1740                 cond_resched();
1741         }
1742 out:
1743         if (err && index_ret)
1744                 *index_ret = start_index + pages_locked - 1;
1745         return err;
1746 }
1747
1748 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1749                                  u64 delalloc_end, struct page *locked_page,
1750                                  unsigned clear_bits,
1751                                  unsigned long page_ops)
1752 {
1753         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1754                          NULL);
1755
1756         __process_pages_contig(inode->i_mapping, locked_page,
1757                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1758                                page_ops, NULL);
1759 }
1760
1761 /*
1762  * count the number of bytes in the tree that have a given bit(s)
1763  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1764  * cached.  The total number found is returned.
1765  */
1766 u64 count_range_bits(struct extent_io_tree *tree,
1767                      u64 *start, u64 search_end, u64 max_bytes,
1768                      unsigned bits, int contig)
1769 {
1770         struct rb_node *node;
1771         struct extent_state *state;
1772         u64 cur_start = *start;
1773         u64 total_bytes = 0;
1774         u64 last = 0;
1775         int found = 0;
1776
1777         if (WARN_ON(search_end <= cur_start))
1778                 return 0;
1779
1780         spin_lock(&tree->lock);
1781         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1782                 total_bytes = tree->dirty_bytes;
1783                 goto out;
1784         }
1785         /*
1786          * this search will find all the extents that end after
1787          * our range starts.
1788          */
1789         node = tree_search(tree, cur_start);
1790         if (!node)
1791                 goto out;
1792
1793         while (1) {
1794                 state = rb_entry(node, struct extent_state, rb_node);
1795                 if (state->start > search_end)
1796                         break;
1797                 if (contig && found && state->start > last + 1)
1798                         break;
1799                 if (state->end >= cur_start && (state->state & bits) == bits) {
1800                         total_bytes += min(search_end, state->end) + 1 -
1801                                        max(cur_start, state->start);
1802                         if (total_bytes >= max_bytes)
1803                                 break;
1804                         if (!found) {
1805                                 *start = max(cur_start, state->start);
1806                                 found = 1;
1807                         }
1808                         last = state->end;
1809                 } else if (contig && found) {
1810                         break;
1811                 }
1812                 node = rb_next(node);
1813                 if (!node)
1814                         break;
1815         }
1816 out:
1817         spin_unlock(&tree->lock);
1818         return total_bytes;
1819 }
1820
1821 /*
1822  * set the private field for a given byte offset in the tree.  If there isn't
1823  * an extent_state there already, this does nothing.
1824  */
1825 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1826                 struct io_failure_record *failrec)
1827 {
1828         struct rb_node *node;
1829         struct extent_state *state;
1830         int ret = 0;
1831
1832         spin_lock(&tree->lock);
1833         /*
1834          * this search will find all the extents that end after
1835          * our range starts.
1836          */
1837         node = tree_search(tree, start);
1838         if (!node) {
1839                 ret = -ENOENT;
1840                 goto out;
1841         }
1842         state = rb_entry(node, struct extent_state, rb_node);
1843         if (state->start != start) {
1844                 ret = -ENOENT;
1845                 goto out;
1846         }
1847         state->failrec = failrec;
1848 out:
1849         spin_unlock(&tree->lock);
1850         return ret;
1851 }
1852
1853 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1854                 struct io_failure_record **failrec)
1855 {
1856         struct rb_node *node;
1857         struct extent_state *state;
1858         int ret = 0;
1859
1860         spin_lock(&tree->lock);
1861         /*
1862          * this search will find all the extents that end after
1863          * our range starts.
1864          */
1865         node = tree_search(tree, start);
1866         if (!node) {
1867                 ret = -ENOENT;
1868                 goto out;
1869         }
1870         state = rb_entry(node, struct extent_state, rb_node);
1871         if (state->start != start) {
1872                 ret = -ENOENT;
1873                 goto out;
1874         }
1875         *failrec = state->failrec;
1876 out:
1877         spin_unlock(&tree->lock);
1878         return ret;
1879 }
1880
1881 /*
1882  * searches a range in the state tree for a given mask.
1883  * If 'filled' == 1, this returns 1 only if every extent in the tree
1884  * has the bits set.  Otherwise, 1 is returned if any bit in the
1885  * range is found set.
1886  */
1887 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1888                    unsigned bits, int filled, struct extent_state *cached)
1889 {
1890         struct extent_state *state = NULL;
1891         struct rb_node *node;
1892         int bitset = 0;
1893
1894         spin_lock(&tree->lock);
1895         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1896             cached->end > start)
1897                 node = &cached->rb_node;
1898         else
1899                 node = tree_search(tree, start);
1900         while (node && start <= end) {
1901                 state = rb_entry(node, struct extent_state, rb_node);
1902
1903                 if (filled && state->start > start) {
1904                         bitset = 0;
1905                         break;
1906                 }
1907
1908                 if (state->start > end)
1909                         break;
1910
1911                 if (state->state & bits) {
1912                         bitset = 1;
1913                         if (!filled)
1914                                 break;
1915                 } else if (filled) {
1916                         bitset = 0;
1917                         break;
1918                 }
1919
1920                 if (state->end == (u64)-1)
1921                         break;
1922
1923                 start = state->end + 1;
1924                 if (start > end)
1925                         break;
1926                 node = rb_next(node);
1927                 if (!node) {
1928                         if (filled)
1929                                 bitset = 0;
1930                         break;
1931                 }
1932         }
1933         spin_unlock(&tree->lock);
1934         return bitset;
1935 }
1936
1937 /*
1938  * helper function to set a given page up to date if all the
1939  * extents in the tree for that page are up to date
1940  */
1941 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1942 {
1943         u64 start = page_offset(page);
1944         u64 end = start + PAGE_SIZE - 1;
1945         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1946                 SetPageUptodate(page);
1947 }
1948
1949 int free_io_failure(struct extent_io_tree *failure_tree,
1950                     struct extent_io_tree *io_tree,
1951                     struct io_failure_record *rec)
1952 {
1953         int ret;
1954         int err = 0;
1955
1956         set_state_failrec(failure_tree, rec->start, NULL);
1957         ret = clear_extent_bits(failure_tree, rec->start,
1958                                 rec->start + rec->len - 1,
1959                                 EXTENT_LOCKED | EXTENT_DIRTY);
1960         if (ret)
1961                 err = ret;
1962
1963         ret = clear_extent_bits(io_tree, rec->start,
1964                                 rec->start + rec->len - 1,
1965                                 EXTENT_DAMAGED);
1966         if (ret && !err)
1967                 err = ret;
1968
1969         kfree(rec);
1970         return err;
1971 }
1972
1973 /*
1974  * this bypasses the standard btrfs submit functions deliberately, as
1975  * the standard behavior is to write all copies in a raid setup. here we only
1976  * want to write the one bad copy. so we do the mapping for ourselves and issue
1977  * submit_bio directly.
1978  * to avoid any synchronization issues, wait for the data after writing, which
1979  * actually prevents the read that triggered the error from finishing.
1980  * currently, there can be no more than two copies of every data bit. thus,
1981  * exactly one rewrite is required.
1982  */
1983 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
1984                       u64 length, u64 logical, struct page *page,
1985                       unsigned int pg_offset, int mirror_num)
1986 {
1987         struct bio *bio;
1988         struct btrfs_device *dev;
1989         u64 map_length = 0;
1990         u64 sector;
1991         struct btrfs_bio *bbio = NULL;
1992         int ret;
1993
1994         ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
1995         BUG_ON(!mirror_num);
1996
1997         bio = btrfs_io_bio_alloc(1);
1998         bio->bi_iter.bi_size = 0;
1999         map_length = length;
2000
2001         /*
2002          * Avoid races with device replace and make sure our bbio has devices
2003          * associated to its stripes that don't go away while we are doing the
2004          * read repair operation.
2005          */
2006         btrfs_bio_counter_inc_blocked(fs_info);
2007         if (btrfs_is_parity_mirror(fs_info, logical, length)) {
2008                 /*
2009                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2010                  * to update all raid stripes, but here we just want to correct
2011                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2012                  * stripe's dev and sector.
2013                  */
2014                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2015                                       &map_length, &bbio, 0);
2016                 if (ret) {
2017                         btrfs_bio_counter_dec(fs_info);
2018                         bio_put(bio);
2019                         return -EIO;
2020                 }
2021                 ASSERT(bbio->mirror_num == 1);
2022         } else {
2023                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2024                                       &map_length, &bbio, mirror_num);
2025                 if (ret) {
2026                         btrfs_bio_counter_dec(fs_info);
2027                         bio_put(bio);
2028                         return -EIO;
2029                 }
2030                 BUG_ON(mirror_num != bbio->mirror_num);
2031         }
2032
2033         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2034         bio->bi_iter.bi_sector = sector;
2035         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2036         btrfs_put_bbio(bbio);
2037         if (!dev || !dev->bdev ||
2038             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2039                 btrfs_bio_counter_dec(fs_info);
2040                 bio_put(bio);
2041                 return -EIO;
2042         }
2043         bio_set_dev(bio, dev->bdev);
2044         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2045         bio_add_page(bio, page, length, pg_offset);
2046
2047         if (btrfsic_submit_bio_wait(bio)) {
2048                 /* try to remap that extent elsewhere? */
2049                 btrfs_bio_counter_dec(fs_info);
2050                 bio_put(bio);
2051                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2052                 return -EIO;
2053         }
2054
2055         btrfs_info_rl_in_rcu(fs_info,
2056                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2057                                   ino, start,
2058                                   rcu_str_deref(dev->name), sector);
2059         btrfs_bio_counter_dec(fs_info);
2060         bio_put(bio);
2061         return 0;
2062 }
2063
2064 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2065                          struct extent_buffer *eb, int mirror_num)
2066 {
2067         u64 start = eb->start;
2068         int i, num_pages = num_extent_pages(eb);
2069         int ret = 0;
2070
2071         if (sb_rdonly(fs_info->sb))
2072                 return -EROFS;
2073
2074         for (i = 0; i < num_pages; i++) {
2075                 struct page *p = eb->pages[i];
2076
2077                 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2078                                         start - page_offset(p), mirror_num);
2079                 if (ret)
2080                         break;
2081                 start += PAGE_SIZE;
2082         }
2083
2084         return ret;
2085 }
2086
2087 /*
2088  * each time an IO finishes, we do a fast check in the IO failure tree
2089  * to see if we need to process or clean up an io_failure_record
2090  */
2091 int clean_io_failure(struct btrfs_fs_info *fs_info,
2092                      struct extent_io_tree *failure_tree,
2093                      struct extent_io_tree *io_tree, u64 start,
2094                      struct page *page, u64 ino, unsigned int pg_offset)
2095 {
2096         u64 private;
2097         struct io_failure_record *failrec;
2098         struct extent_state *state;
2099         int num_copies;
2100         int ret;
2101
2102         private = 0;
2103         ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2104                                EXTENT_DIRTY, 0);
2105         if (!ret)
2106                 return 0;
2107
2108         ret = get_state_failrec(failure_tree, start, &failrec);
2109         if (ret)
2110                 return 0;
2111
2112         BUG_ON(!failrec->this_mirror);
2113
2114         if (failrec->in_validation) {
2115                 /* there was no real error, just free the record */
2116                 btrfs_debug(fs_info,
2117                         "clean_io_failure: freeing dummy error at %llu",
2118                         failrec->start);
2119                 goto out;
2120         }
2121         if (sb_rdonly(fs_info->sb))
2122                 goto out;
2123
2124         spin_lock(&io_tree->lock);
2125         state = find_first_extent_bit_state(io_tree,
2126                                             failrec->start,
2127                                             EXTENT_LOCKED);
2128         spin_unlock(&io_tree->lock);
2129
2130         if (state && state->start <= failrec->start &&
2131             state->end >= failrec->start + failrec->len - 1) {
2132                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2133                                               failrec->len);
2134                 if (num_copies > 1)  {
2135                         repair_io_failure(fs_info, ino, start, failrec->len,
2136                                           failrec->logical, page, pg_offset,
2137                                           failrec->failed_mirror);
2138                 }
2139         }
2140
2141 out:
2142         free_io_failure(failure_tree, io_tree, failrec);
2143
2144         return 0;
2145 }
2146
2147 /*
2148  * Can be called when
2149  * - hold extent lock
2150  * - under ordered extent
2151  * - the inode is freeing
2152  */
2153 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2154 {
2155         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2156         struct io_failure_record *failrec;
2157         struct extent_state *state, *next;
2158
2159         if (RB_EMPTY_ROOT(&failure_tree->state))
2160                 return;
2161
2162         spin_lock(&failure_tree->lock);
2163         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2164         while (state) {
2165                 if (state->start > end)
2166                         break;
2167
2168                 ASSERT(state->end <= end);
2169
2170                 next = next_state(state);
2171
2172                 failrec = state->failrec;
2173                 free_extent_state(state);
2174                 kfree(failrec);
2175
2176                 state = next;
2177         }
2178         spin_unlock(&failure_tree->lock);
2179 }
2180
2181 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2182                 struct io_failure_record **failrec_ret)
2183 {
2184         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2185         struct io_failure_record *failrec;
2186         struct extent_map *em;
2187         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2188         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2189         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2190         int ret;
2191         u64 logical;
2192
2193         ret = get_state_failrec(failure_tree, start, &failrec);
2194         if (ret) {
2195                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2196                 if (!failrec)
2197                         return -ENOMEM;
2198
2199                 failrec->start = start;
2200                 failrec->len = end - start + 1;
2201                 failrec->this_mirror = 0;
2202                 failrec->bio_flags = 0;
2203                 failrec->in_validation = 0;
2204
2205                 read_lock(&em_tree->lock);
2206                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2207                 if (!em) {
2208                         read_unlock(&em_tree->lock);
2209                         kfree(failrec);
2210                         return -EIO;
2211                 }
2212
2213                 if (em->start > start || em->start + em->len <= start) {
2214                         free_extent_map(em);
2215                         em = NULL;
2216                 }
2217                 read_unlock(&em_tree->lock);
2218                 if (!em) {
2219                         kfree(failrec);
2220                         return -EIO;
2221                 }
2222
2223                 logical = start - em->start;
2224                 logical = em->block_start + logical;
2225                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2226                         logical = em->block_start;
2227                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2228                         extent_set_compress_type(&failrec->bio_flags,
2229                                                  em->compress_type);
2230                 }
2231
2232                 btrfs_debug(fs_info,
2233                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2234                         logical, start, failrec->len);
2235
2236                 failrec->logical = logical;
2237                 free_extent_map(em);
2238
2239                 /* set the bits in the private failure tree */
2240                 ret = set_extent_bits(failure_tree, start, end,
2241                                         EXTENT_LOCKED | EXTENT_DIRTY);
2242                 if (ret >= 0)
2243                         ret = set_state_failrec(failure_tree, start, failrec);
2244                 /* set the bits in the inode's tree */
2245                 if (ret >= 0)
2246                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2247                 if (ret < 0) {
2248                         kfree(failrec);
2249                         return ret;
2250                 }
2251         } else {
2252                 btrfs_debug(fs_info,
2253                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2254                         failrec->logical, failrec->start, failrec->len,
2255                         failrec->in_validation);
2256                 /*
2257                  * when data can be on disk more than twice, add to failrec here
2258                  * (e.g. with a list for failed_mirror) to make
2259                  * clean_io_failure() clean all those errors at once.
2260                  */
2261         }
2262
2263         *failrec_ret = failrec;
2264
2265         return 0;
2266 }
2267
2268 bool btrfs_check_repairable(struct inode *inode, unsigned failed_bio_pages,
2269                            struct io_failure_record *failrec, int failed_mirror)
2270 {
2271         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2272         int num_copies;
2273
2274         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2275         if (num_copies == 1) {
2276                 /*
2277                  * we only have a single copy of the data, so don't bother with
2278                  * all the retry and error correction code that follows. no
2279                  * matter what the error is, it is very likely to persist.
2280                  */
2281                 btrfs_debug(fs_info,
2282                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2283                         num_copies, failrec->this_mirror, failed_mirror);
2284                 return false;
2285         }
2286
2287         /*
2288          * there are two premises:
2289          *      a) deliver good data to the caller
2290          *      b) correct the bad sectors on disk
2291          */
2292         if (failed_bio_pages > 1) {
2293                 /*
2294                  * to fulfill b), we need to know the exact failing sectors, as
2295                  * we don't want to rewrite any more than the failed ones. thus,
2296                  * we need separate read requests for the failed bio
2297                  *
2298                  * if the following BUG_ON triggers, our validation request got
2299                  * merged. we need separate requests for our algorithm to work.
2300                  */
2301                 BUG_ON(failrec->in_validation);
2302                 failrec->in_validation = 1;
2303                 failrec->this_mirror = failed_mirror;
2304         } else {
2305                 /*
2306                  * we're ready to fulfill a) and b) alongside. get a good copy
2307                  * of the failed sector and if we succeed, we have setup
2308                  * everything for repair_io_failure to do the rest for us.
2309                  */
2310                 if (failrec->in_validation) {
2311                         BUG_ON(failrec->this_mirror != failed_mirror);
2312                         failrec->in_validation = 0;
2313                         failrec->this_mirror = 0;
2314                 }
2315                 failrec->failed_mirror = failed_mirror;
2316                 failrec->this_mirror++;
2317                 if (failrec->this_mirror == failed_mirror)
2318                         failrec->this_mirror++;
2319         }
2320
2321         if (failrec->this_mirror > num_copies) {
2322                 btrfs_debug(fs_info,
2323                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2324                         num_copies, failrec->this_mirror, failed_mirror);
2325                 return false;
2326         }
2327
2328         return true;
2329 }
2330
2331
2332 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2333                                     struct io_failure_record *failrec,
2334                                     struct page *page, int pg_offset, int icsum,
2335                                     bio_end_io_t *endio_func, void *data)
2336 {
2337         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2338         struct bio *bio;
2339         struct btrfs_io_bio *btrfs_failed_bio;
2340         struct btrfs_io_bio *btrfs_bio;
2341
2342         bio = btrfs_io_bio_alloc(1);
2343         bio->bi_end_io = endio_func;
2344         bio->bi_iter.bi_sector = failrec->logical >> 9;
2345         bio_set_dev(bio, fs_info->fs_devices->latest_bdev);
2346         bio->bi_iter.bi_size = 0;
2347         bio->bi_private = data;
2348
2349         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2350         if (btrfs_failed_bio->csum) {
2351                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2352
2353                 btrfs_bio = btrfs_io_bio(bio);
2354                 btrfs_bio->csum = btrfs_bio->csum_inline;
2355                 icsum *= csum_size;
2356                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2357                        csum_size);
2358         }
2359
2360         bio_add_page(bio, page, failrec->len, pg_offset);
2361
2362         return bio;
2363 }
2364
2365 /*
2366  * This is a generic handler for readpage errors. If other copies exist, read
2367  * those and write back good data to the failed position. Does not investigate
2368  * in remapping the failed extent elsewhere, hoping the device will be smart
2369  * enough to do this as needed
2370  */
2371 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2372                               struct page *page, u64 start, u64 end,
2373                               int failed_mirror)
2374 {
2375         struct io_failure_record *failrec;
2376         struct inode *inode = page->mapping->host;
2377         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2378         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2379         struct bio *bio;
2380         int read_mode = 0;
2381         blk_status_t status;
2382         int ret;
2383         unsigned failed_bio_pages = failed_bio->bi_iter.bi_size >> PAGE_SHIFT;
2384
2385         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2386
2387         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2388         if (ret)
2389                 return ret;
2390
2391         if (!btrfs_check_repairable(inode, failed_bio_pages, failrec,
2392                                     failed_mirror)) {
2393                 free_io_failure(failure_tree, tree, failrec);
2394                 return -EIO;
2395         }
2396
2397         if (failed_bio_pages > 1)
2398                 read_mode |= REQ_FAILFAST_DEV;
2399
2400         phy_offset >>= inode->i_sb->s_blocksize_bits;
2401         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2402                                       start - page_offset(page),
2403                                       (int)phy_offset, failed_bio->bi_end_io,
2404                                       NULL);
2405         bio->bi_opf = REQ_OP_READ | read_mode;
2406
2407         btrfs_debug(btrfs_sb(inode->i_sb),
2408                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2409                 read_mode, failrec->this_mirror, failrec->in_validation);
2410
2411         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2412                                          failrec->bio_flags, 0);
2413         if (status) {
2414                 free_io_failure(failure_tree, tree, failrec);
2415                 bio_put(bio);
2416                 ret = blk_status_to_errno(status);
2417         }
2418
2419         return ret;
2420 }
2421
2422 /* lots and lots of room for performance fixes in the end_bio funcs */
2423
2424 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2425 {
2426         int uptodate = (err == 0);
2427         int ret = 0;
2428
2429         btrfs_writepage_endio_finish_ordered(page, start, end, uptodate);
2430
2431         if (!uptodate) {
2432                 ClearPageUptodate(page);
2433                 SetPageError(page);
2434                 ret = err < 0 ? err : -EIO;
2435                 mapping_set_error(page->mapping, ret);
2436         }
2437 }
2438
2439 /*
2440  * after a writepage IO is done, we need to:
2441  * clear the uptodate bits on error
2442  * clear the writeback bits in the extent tree for this IO
2443  * end_page_writeback if the page has no more pending IO
2444  *
2445  * Scheduling is not allowed, so the extent state tree is expected
2446  * to have one and only one object corresponding to this IO.
2447  */
2448 static void end_bio_extent_writepage(struct bio *bio)
2449 {
2450         int error = blk_status_to_errno(bio->bi_status);
2451         struct bio_vec *bvec;
2452         u64 start;
2453         u64 end;
2454         int i;
2455         struct bvec_iter_all iter_all;
2456
2457         ASSERT(!bio_flagged(bio, BIO_CLONED));
2458         bio_for_each_segment_all(bvec, bio, i, iter_all) {
2459                 struct page *page = bvec->bv_page;
2460                 struct inode *inode = page->mapping->host;
2461                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2462
2463                 /* We always issue full-page reads, but if some block
2464                  * in a page fails to read, blk_update_request() will
2465                  * advance bv_offset and adjust bv_len to compensate.
2466                  * Print a warning for nonzero offsets, and an error
2467                  * if they don't add up to a full page.  */
2468                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2469                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2470                                 btrfs_err(fs_info,
2471                                    "partial page write in btrfs with offset %u and length %u",
2472                                         bvec->bv_offset, bvec->bv_len);
2473                         else
2474                                 btrfs_info(fs_info,
2475                                    "incomplete page write in btrfs with offset %u and length %u",
2476                                         bvec->bv_offset, bvec->bv_len);
2477                 }
2478
2479                 start = page_offset(page);
2480                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2481
2482                 end_extent_writepage(page, error, start, end);
2483                 end_page_writeback(page);
2484         }
2485
2486         bio_put(bio);
2487 }
2488
2489 static void
2490 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2491                               int uptodate)
2492 {
2493         struct extent_state *cached = NULL;
2494         u64 end = start + len - 1;
2495
2496         if (uptodate && tree->track_uptodate)
2497                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2498         unlock_extent_cached_atomic(tree, start, end, &cached);
2499 }
2500
2501 /*
2502  * after a readpage IO is done, we need to:
2503  * clear the uptodate bits on error
2504  * set the uptodate bits if things worked
2505  * set the page up to date if all extents in the tree are uptodate
2506  * clear the lock bit in the extent tree
2507  * unlock the page if there are no other extents locked for it
2508  *
2509  * Scheduling is not allowed, so the extent state tree is expected
2510  * to have one and only one object corresponding to this IO.
2511  */
2512 static void end_bio_extent_readpage(struct bio *bio)
2513 {
2514         struct bio_vec *bvec;
2515         int uptodate = !bio->bi_status;
2516         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2517         struct extent_io_tree *tree, *failure_tree;
2518         u64 offset = 0;
2519         u64 start;
2520         u64 end;
2521         u64 len;
2522         u64 extent_start = 0;
2523         u64 extent_len = 0;
2524         int mirror;
2525         int ret;
2526         int i;
2527         struct bvec_iter_all iter_all;
2528
2529         ASSERT(!bio_flagged(bio, BIO_CLONED));
2530         bio_for_each_segment_all(bvec, bio, i, iter_all) {
2531                 struct page *page = bvec->bv_page;
2532                 struct inode *inode = page->mapping->host;
2533                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2534                 bool data_inode = btrfs_ino(BTRFS_I(inode))
2535                         != BTRFS_BTREE_INODE_OBJECTID;
2536
2537                 btrfs_debug(fs_info,
2538                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2539                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2540                         io_bio->mirror_num);
2541                 tree = &BTRFS_I(inode)->io_tree;
2542                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2543
2544                 /* We always issue full-page reads, but if some block
2545                  * in a page fails to read, blk_update_request() will
2546                  * advance bv_offset and adjust bv_len to compensate.
2547                  * Print a warning for nonzero offsets, and an error
2548                  * if they don't add up to a full page.  */
2549                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2550                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2551                                 btrfs_err(fs_info,
2552                                         "partial page read in btrfs with offset %u and length %u",
2553                                         bvec->bv_offset, bvec->bv_len);
2554                         else
2555                                 btrfs_info(fs_info,
2556                                         "incomplete page read in btrfs with offset %u and length %u",
2557                                         bvec->bv_offset, bvec->bv_len);
2558                 }
2559
2560                 start = page_offset(page);
2561                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2562                 len = bvec->bv_len;
2563
2564                 mirror = io_bio->mirror_num;
2565                 if (likely(uptodate)) {
2566                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2567                                                               page, start, end,
2568                                                               mirror);
2569                         if (ret)
2570                                 uptodate = 0;
2571                         else
2572                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2573                                                  failure_tree, tree, start,
2574                                                  page,
2575                                                  btrfs_ino(BTRFS_I(inode)), 0);
2576                 }
2577
2578                 if (likely(uptodate))
2579                         goto readpage_ok;
2580
2581                 if (data_inode) {
2582
2583                         /*
2584                          * The generic bio_readpage_error handles errors the
2585                          * following way: If possible, new read requests are
2586                          * created and submitted and will end up in
2587                          * end_bio_extent_readpage as well (if we're lucky,
2588                          * not in the !uptodate case). In that case it returns
2589                          * 0 and we just go on with the next page in our bio.
2590                          * If it can't handle the error it will return -EIO and
2591                          * we remain responsible for that page.
2592                          */
2593                         ret = bio_readpage_error(bio, offset, page, start, end,
2594                                                  mirror);
2595                         if (ret == 0) {
2596                                 uptodate = !bio->bi_status;
2597                                 offset += len;
2598                                 continue;
2599                         }
2600                 } else {
2601                         struct extent_buffer *eb;
2602
2603                         eb = (struct extent_buffer *)page->private;
2604                         set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
2605                         eb->read_mirror = mirror;
2606                         atomic_dec(&eb->io_pages);
2607                         if (test_and_clear_bit(EXTENT_BUFFER_READAHEAD,
2608                                                &eb->bflags))
2609                                 btree_readahead_hook(eb, -EIO);
2610
2611                         ret = -EIO;
2612                 }
2613 readpage_ok:
2614                 if (likely(uptodate)) {
2615                         loff_t i_size = i_size_read(inode);
2616                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2617                         unsigned off;
2618
2619                         /* Zero out the end if this page straddles i_size */
2620                         off = offset_in_page(i_size);
2621                         if (page->index == end_index && off)
2622                                 zero_user_segment(page, off, PAGE_SIZE);
2623                         SetPageUptodate(page);
2624                 } else {
2625                         ClearPageUptodate(page);
2626                         SetPageError(page);
2627                 }
2628                 unlock_page(page);
2629                 offset += len;
2630
2631                 if (unlikely(!uptodate)) {
2632                         if (extent_len) {
2633                                 endio_readpage_release_extent(tree,
2634                                                               extent_start,
2635                                                               extent_len, 1);
2636                                 extent_start = 0;
2637                                 extent_len = 0;
2638                         }
2639                         endio_readpage_release_extent(tree, start,
2640                                                       end - start + 1, 0);
2641                 } else if (!extent_len) {
2642                         extent_start = start;
2643                         extent_len = end + 1 - start;
2644                 } else if (extent_start + extent_len == start) {
2645                         extent_len += end + 1 - start;
2646                 } else {
2647                         endio_readpage_release_extent(tree, extent_start,
2648                                                       extent_len, uptodate);
2649                         extent_start = start;
2650                         extent_len = end + 1 - start;
2651                 }
2652         }
2653
2654         if (extent_len)
2655                 endio_readpage_release_extent(tree, extent_start, extent_len,
2656                                               uptodate);
2657         btrfs_io_bio_free_csum(io_bio);
2658         bio_put(bio);
2659 }
2660
2661 /*
2662  * Initialize the members up to but not including 'bio'. Use after allocating a
2663  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2664  * 'bio' because use of __GFP_ZERO is not supported.
2665  */
2666 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2667 {
2668         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2669 }
2670
2671 /*
2672  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2673  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2674  * for the appropriate container_of magic
2675  */
2676 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte)
2677 {
2678         struct bio *bio;
2679
2680         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, &btrfs_bioset);
2681         bio_set_dev(bio, bdev);
2682         bio->bi_iter.bi_sector = first_byte >> 9;
2683         btrfs_io_bio_init(btrfs_io_bio(bio));
2684         return bio;
2685 }
2686
2687 struct bio *btrfs_bio_clone(struct bio *bio)
2688 {
2689         struct btrfs_io_bio *btrfs_bio;
2690         struct bio *new;
2691
2692         /* Bio allocation backed by a bioset does not fail */
2693         new = bio_clone_fast(bio, GFP_NOFS, &btrfs_bioset);
2694         btrfs_bio = btrfs_io_bio(new);
2695         btrfs_io_bio_init(btrfs_bio);
2696         btrfs_bio->iter = bio->bi_iter;
2697         return new;
2698 }
2699
2700 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2701 {
2702         struct bio *bio;
2703
2704         /* Bio allocation backed by a bioset does not fail */
2705         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, &btrfs_bioset);
2706         btrfs_io_bio_init(btrfs_io_bio(bio));
2707         return bio;
2708 }
2709
2710 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2711 {
2712         struct bio *bio;
2713         struct btrfs_io_bio *btrfs_bio;
2714
2715         /* this will never fail when it's backed by a bioset */
2716         bio = bio_clone_fast(orig, GFP_NOFS, &btrfs_bioset);
2717         ASSERT(bio);
2718
2719         btrfs_bio = btrfs_io_bio(bio);
2720         btrfs_io_bio_init(btrfs_bio);
2721
2722         bio_trim(bio, offset >> 9, size >> 9);
2723         btrfs_bio->iter = bio->bi_iter;
2724         return bio;
2725 }
2726
2727 /*
2728  * @opf:        bio REQ_OP_* and REQ_* flags as one value
2729  * @tree:       tree so we can call our merge_bio hook
2730  * @wbc:        optional writeback control for io accounting
2731  * @page:       page to add to the bio
2732  * @pg_offset:  offset of the new bio or to check whether we are adding
2733  *              a contiguous page to the previous one
2734  * @size:       portion of page that we want to write
2735  * @offset:     starting offset in the page
2736  * @bdev:       attach newly created bios to this bdev
2737  * @bio_ret:    must be valid pointer, newly allocated bio will be stored there
2738  * @end_io_func:     end_io callback for new bio
2739  * @mirror_num:      desired mirror to read/write
2740  * @prev_bio_flags:  flags of previous bio to see if we can merge the current one
2741  * @bio_flags:  flags of the current bio to see if we can merge them
2742  */
2743 static int submit_extent_page(unsigned int opf, struct extent_io_tree *tree,
2744                               struct writeback_control *wbc,
2745                               struct page *page, u64 offset,
2746                               size_t size, unsigned long pg_offset,
2747                               struct block_device *bdev,
2748                               struct bio **bio_ret,
2749                               bio_end_io_t end_io_func,
2750                               int mirror_num,
2751                               unsigned long prev_bio_flags,
2752                               unsigned long bio_flags,
2753                               bool force_bio_submit)
2754 {
2755         int ret = 0;
2756         struct bio *bio;
2757         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2758         sector_t sector = offset >> 9;
2759
2760         ASSERT(bio_ret);
2761
2762         if (*bio_ret) {
2763                 bool contig;
2764                 bool can_merge = true;
2765
2766                 bio = *bio_ret;
2767                 if (prev_bio_flags & EXTENT_BIO_COMPRESSED)
2768                         contig = bio->bi_iter.bi_sector == sector;
2769                 else
2770                         contig = bio_end_sector(bio) == sector;
2771
2772                 ASSERT(tree->ops);
2773                 if (btrfs_bio_fits_in_stripe(page, page_size, bio, bio_flags))
2774                         can_merge = false;
2775
2776                 if (prev_bio_flags != bio_flags || !contig || !can_merge ||
2777                     force_bio_submit ||
2778                     bio_add_page(bio, page, page_size, pg_offset) < page_size) {
2779                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2780                         if (ret < 0) {
2781                                 *bio_ret = NULL;
2782                                 return ret;
2783                         }
2784                         bio = NULL;
2785                 } else {
2786                         if (wbc)
2787                                 wbc_account_io(wbc, page, page_size);
2788                         return 0;
2789                 }
2790         }
2791
2792         bio = btrfs_bio_alloc(bdev, offset);
2793         bio_add_page(bio, page, page_size, pg_offset);
2794         bio->bi_end_io = end_io_func;
2795         bio->bi_private = tree;
2796         bio->bi_write_hint = page->mapping->host->i_write_hint;
2797         bio->bi_opf = opf;
2798         if (wbc) {
2799                 wbc_init_bio(wbc, bio);
2800                 wbc_account_io(wbc, page, page_size);
2801         }
2802
2803         *bio_ret = bio;
2804
2805         return ret;
2806 }
2807
2808 static void attach_extent_buffer_page(struct extent_buffer *eb,
2809                                       struct page *page)
2810 {
2811         if (!PagePrivate(page)) {
2812                 SetPagePrivate(page);
2813                 get_page(page);
2814                 set_page_private(page, (unsigned long)eb);
2815         } else {
2816                 WARN_ON(page->private != (unsigned long)eb);
2817         }
2818 }
2819
2820 void set_page_extent_mapped(struct page *page)
2821 {
2822         if (!PagePrivate(page)) {
2823                 SetPagePrivate(page);
2824                 get_page(page);
2825                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2826         }
2827 }
2828
2829 static struct extent_map *
2830 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2831                  u64 start, u64 len, get_extent_t *get_extent,
2832                  struct extent_map **em_cached)
2833 {
2834         struct extent_map *em;
2835
2836         if (em_cached && *em_cached) {
2837                 em = *em_cached;
2838                 if (extent_map_in_tree(em) && start >= em->start &&
2839                     start < extent_map_end(em)) {
2840                         refcount_inc(&em->refs);
2841                         return em;
2842                 }
2843
2844                 free_extent_map(em);
2845                 *em_cached = NULL;
2846         }
2847
2848         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2849         if (em_cached && !IS_ERR_OR_NULL(em)) {
2850                 BUG_ON(*em_cached);
2851                 refcount_inc(&em->refs);
2852                 *em_cached = em;
2853         }
2854         return em;
2855 }
2856 /*
2857  * basic readpage implementation.  Locked extent state structs are inserted
2858  * into the tree that are removed when the IO is done (by the end_io
2859  * handlers)
2860  * XXX JDM: This needs looking at to ensure proper page locking
2861  * return 0 on success, otherwise return error
2862  */
2863 static int __do_readpage(struct extent_io_tree *tree,
2864                          struct page *page,
2865                          get_extent_t *get_extent,
2866                          struct extent_map **em_cached,
2867                          struct bio **bio, int mirror_num,
2868                          unsigned long *bio_flags, unsigned int read_flags,
2869                          u64 *prev_em_start)
2870 {
2871         struct inode *inode = page->mapping->host;
2872         u64 start = page_offset(page);
2873         const u64 end = start + PAGE_SIZE - 1;
2874         u64 cur = start;
2875         u64 extent_offset;
2876         u64 last_byte = i_size_read(inode);
2877         u64 block_start;
2878         u64 cur_end;
2879         struct extent_map *em;
2880         struct block_device *bdev;
2881         int ret = 0;
2882         int nr = 0;
2883         size_t pg_offset = 0;
2884         size_t iosize;
2885         size_t disk_io_size;
2886         size_t blocksize = inode->i_sb->s_blocksize;
2887         unsigned long this_bio_flag = 0;
2888
2889         set_page_extent_mapped(page);
2890
2891         if (!PageUptodate(page)) {
2892                 if (cleancache_get_page(page) == 0) {
2893                         BUG_ON(blocksize != PAGE_SIZE);
2894                         unlock_extent(tree, start, end);
2895                         goto out;
2896                 }
2897         }
2898
2899         if (page->index == last_byte >> PAGE_SHIFT) {
2900                 char *userpage;
2901                 size_t zero_offset = offset_in_page(last_byte);
2902
2903                 if (zero_offset) {
2904                         iosize = PAGE_SIZE - zero_offset;
2905                         userpage = kmap_atomic(page);
2906                         memset(userpage + zero_offset, 0, iosize);
2907                         flush_dcache_page(page);
2908                         kunmap_atomic(userpage);
2909                 }
2910         }
2911         while (cur <= end) {
2912                 bool force_bio_submit = false;
2913                 u64 offset;
2914
2915                 if (cur >= last_byte) {
2916                         char *userpage;
2917                         struct extent_state *cached = NULL;
2918
2919                         iosize = PAGE_SIZE - pg_offset;
2920                         userpage = kmap_atomic(page);
2921                         memset(userpage + pg_offset, 0, iosize);
2922                         flush_dcache_page(page);
2923                         kunmap_atomic(userpage);
2924                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2925                                             &cached, GFP_NOFS);
2926                         unlock_extent_cached(tree, cur,
2927                                              cur + iosize - 1, &cached);
2928                         break;
2929                 }
2930                 em = __get_extent_map(inode, page, pg_offset, cur,
2931                                       end - cur + 1, get_extent, em_cached);
2932                 if (IS_ERR_OR_NULL(em)) {
2933                         SetPageError(page);
2934                         unlock_extent(tree, cur, end);
2935                         break;
2936                 }
2937                 extent_offset = cur - em->start;
2938                 BUG_ON(extent_map_end(em) <= cur);
2939                 BUG_ON(end < cur);
2940
2941                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2942                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2943                         extent_set_compress_type(&this_bio_flag,
2944                                                  em->compress_type);
2945                 }
2946
2947                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2948                 cur_end = min(extent_map_end(em) - 1, end);
2949                 iosize = ALIGN(iosize, blocksize);
2950                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2951                         disk_io_size = em->block_len;
2952                         offset = em->block_start;
2953                 } else {
2954                         offset = em->block_start + extent_offset;
2955                         disk_io_size = iosize;
2956                 }
2957                 bdev = em->bdev;
2958                 block_start = em->block_start;
2959                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2960                         block_start = EXTENT_MAP_HOLE;
2961
2962                 /*
2963                  * If we have a file range that points to a compressed extent
2964                  * and it's followed by a consecutive file range that points to
2965                  * to the same compressed extent (possibly with a different
2966                  * offset and/or length, so it either points to the whole extent
2967                  * or only part of it), we must make sure we do not submit a
2968                  * single bio to populate the pages for the 2 ranges because
2969                  * this makes the compressed extent read zero out the pages
2970                  * belonging to the 2nd range. Imagine the following scenario:
2971                  *
2972                  *  File layout
2973                  *  [0 - 8K]                     [8K - 24K]
2974                  *    |                               |
2975                  *    |                               |
2976                  * points to extent X,         points to extent X,
2977                  * offset 4K, length of 8K     offset 0, length 16K
2978                  *
2979                  * [extent X, compressed length = 4K uncompressed length = 16K]
2980                  *
2981                  * If the bio to read the compressed extent covers both ranges,
2982                  * it will decompress extent X into the pages belonging to the
2983                  * first range and then it will stop, zeroing out the remaining
2984                  * pages that belong to the other range that points to extent X.
2985                  * So here we make sure we submit 2 bios, one for the first
2986                  * range and another one for the third range. Both will target
2987                  * the same physical extent from disk, but we can't currently
2988                  * make the compressed bio endio callback populate the pages
2989                  * for both ranges because each compressed bio is tightly
2990                  * coupled with a single extent map, and each range can have
2991                  * an extent map with a different offset value relative to the
2992                  * uncompressed data of our extent and different lengths. This
2993                  * is a corner case so we prioritize correctness over
2994                  * non-optimal behavior (submitting 2 bios for the same extent).
2995                  */
2996                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
2997                     prev_em_start && *prev_em_start != (u64)-1 &&
2998                     *prev_em_start != em->start)
2999                         force_bio_submit = true;
3000
3001                 if (prev_em_start)
3002                         *prev_em_start = em->start;
3003
3004                 free_extent_map(em);
3005                 em = NULL;
3006
3007                 /* we've found a hole, just zero and go on */
3008                 if (block_start == EXTENT_MAP_HOLE) {
3009                         char *userpage;
3010                         struct extent_state *cached = NULL;
3011
3012                         userpage = kmap_atomic(page);
3013                         memset(userpage + pg_offset, 0, iosize);
3014                         flush_dcache_page(page);
3015                         kunmap_atomic(userpage);
3016
3017                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3018                                             &cached, GFP_NOFS);
3019                         unlock_extent_cached(tree, cur,
3020                                              cur + iosize - 1, &cached);
3021                         cur = cur + iosize;
3022                         pg_offset += iosize;
3023                         continue;
3024                 }
3025                 /* the get_extent function already copied into the page */
3026                 if (test_range_bit(tree, cur, cur_end,
3027                                    EXTENT_UPTODATE, 1, NULL)) {
3028                         check_page_uptodate(tree, page);
3029                         unlock_extent(tree, cur, cur + iosize - 1);
3030                         cur = cur + iosize;
3031                         pg_offset += iosize;
3032                         continue;
3033                 }
3034                 /* we have an inline extent but it didn't get marked up
3035                  * to date.  Error out
3036                  */
3037                 if (block_start == EXTENT_MAP_INLINE) {
3038                         SetPageError(page);
3039                         unlock_extent(tree, cur, cur + iosize - 1);
3040                         cur = cur + iosize;
3041                         pg_offset += iosize;
3042                         continue;
3043                 }
3044
3045                 ret = submit_extent_page(REQ_OP_READ | read_flags, tree, NULL,
3046                                          page, offset, disk_io_size,
3047                                          pg_offset, bdev, bio,
3048                                          end_bio_extent_readpage, mirror_num,
3049                                          *bio_flags,
3050                                          this_bio_flag,
3051                                          force_bio_submit);
3052                 if (!ret) {
3053                         nr++;
3054                         *bio_flags = this_bio_flag;
3055                 } else {
3056                         SetPageError(page);
3057                         unlock_extent(tree, cur, cur + iosize - 1);
3058                         goto out;
3059                 }
3060                 cur = cur + iosize;
3061                 pg_offset += iosize;
3062         }
3063 out:
3064         if (!nr) {
3065                 if (!PageError(page))
3066                         SetPageUptodate(page);
3067                 unlock_page(page);
3068         }
3069         return ret;
3070 }
3071
3072 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3073                                              struct page *pages[], int nr_pages,
3074                                              u64 start, u64 end,
3075                                              struct extent_map **em_cached,
3076                                              struct bio **bio,
3077                                              unsigned long *bio_flags,
3078                                              u64 *prev_em_start)
3079 {
3080         struct inode *inode;
3081         struct btrfs_ordered_extent *ordered;
3082         int index;
3083
3084         inode = pages[0]->mapping->host;
3085         while (1) {
3086                 lock_extent(tree, start, end);
3087                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3088                                                      end - start + 1);
3089                 if (!ordered)
3090                         break;
3091                 unlock_extent(tree, start, end);
3092                 btrfs_start_ordered_extent(inode, ordered, 1);
3093                 btrfs_put_ordered_extent(ordered);
3094         }
3095
3096         for (index = 0; index < nr_pages; index++) {
3097                 __do_readpage(tree, pages[index], btrfs_get_extent, em_cached,
3098                                 bio, 0, bio_flags, REQ_RAHEAD, prev_em_start);
3099                 put_page(pages[index]);
3100         }
3101 }
3102
3103 static void __extent_readpages(struct extent_io_tree *tree,
3104                                struct page *pages[],
3105                                int nr_pages,
3106                                struct extent_map **em_cached,
3107                                struct bio **bio, unsigned long *bio_flags,
3108                                u64 *prev_em_start)
3109 {
3110         u64 start = 0;
3111         u64 end = 0;
3112         u64 page_start;
3113         int index;
3114         int first_index = 0;
3115
3116         for (index = 0; index < nr_pages; index++) {
3117                 page_start = page_offset(pages[index]);
3118                 if (!end) {
3119                         start = page_start;
3120                         end = start + PAGE_SIZE - 1;
3121                         first_index = index;
3122                 } else if (end + 1 == page_start) {
3123                         end += PAGE_SIZE;
3124                 } else {
3125                         __do_contiguous_readpages(tree, &pages[first_index],
3126                                                   index - first_index, start,
3127                                                   end, em_cached,
3128                                                   bio, bio_flags,
3129                                                   prev_em_start);
3130                         start = page_start;
3131                         end = start + PAGE_SIZE - 1;
3132                         first_index = index;
3133                 }
3134         }
3135
3136         if (end)
3137                 __do_contiguous_readpages(tree, &pages[first_index],
3138                                           index - first_index, start,
3139                                           end, em_cached, bio,
3140                                           bio_flags, prev_em_start);
3141 }
3142
3143 static int __extent_read_full_page(struct extent_io_tree *tree,
3144                                    struct page *page,
3145                                    get_extent_t *get_extent,
3146                                    struct bio **bio, int mirror_num,
3147                                    unsigned long *bio_flags,
3148                                    unsigned int read_flags)
3149 {
3150         struct inode *inode = page->mapping->host;
3151         struct btrfs_ordered_extent *ordered;
3152         u64 start = page_offset(page);
3153         u64 end = start + PAGE_SIZE - 1;
3154         int ret;
3155
3156         while (1) {
3157                 lock_extent(tree, start, end);
3158                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3159                                                 PAGE_SIZE);
3160                 if (!ordered)
3161                         break;
3162                 unlock_extent(tree, start, end);
3163                 btrfs_start_ordered_extent(inode, ordered, 1);
3164                 btrfs_put_ordered_extent(ordered);
3165         }
3166
3167         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3168                             bio_flags, read_flags, NULL);
3169         return ret;
3170 }
3171
3172 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3173                             get_extent_t *get_extent, int mirror_num)
3174 {
3175         struct bio *bio = NULL;
3176         unsigned long bio_flags = 0;
3177         int ret;
3178
3179         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3180                                       &bio_flags, 0);
3181         if (bio)
3182                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3183         return ret;
3184 }
3185
3186 static void update_nr_written(struct writeback_control *wbc,
3187                               unsigned long nr_written)
3188 {
3189         wbc->nr_to_write -= nr_written;
3190 }
3191
3192 /*
3193  * helper for __extent_writepage, doing all of the delayed allocation setup.
3194  *
3195  * This returns 1 if btrfs_run_delalloc_range function did all the work required
3196  * to write the page (copy into inline extent).  In this case the IO has
3197  * been started and the page is already unlocked.
3198  *
3199  * This returns 0 if all went well (page still locked)
3200  * This returns < 0 if there were errors (page still locked)
3201  */
3202 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3203                 struct page *page, struct writeback_control *wbc,
3204                 u64 delalloc_start, unsigned long *nr_written)
3205 {
3206         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
3207         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3208         bool found;
3209         u64 delalloc_to_write = 0;
3210         u64 delalloc_end = 0;
3211         int ret;
3212         int page_started = 0;
3213
3214
3215         while (delalloc_end < page_end) {
3216                 found = find_lock_delalloc_range(inode, tree,
3217                                                page,
3218                                                &delalloc_start,
3219                                                &delalloc_end);
3220                 if (!found) {
3221                         delalloc_start = delalloc_end + 1;
3222                         continue;
3223                 }
3224                 ret = btrfs_run_delalloc_range(inode, page, delalloc_start,
3225                                 delalloc_end, &page_started, nr_written, wbc);
3226                 /* File system has been set read-only */
3227                 if (ret) {
3228                         SetPageError(page);
3229                         /*
3230                          * btrfs_run_delalloc_range should return < 0 for error
3231                          * but just in case, we use > 0 here meaning the IO is
3232                          * started, so we don't want to return > 0 unless
3233                          * things are going well.
3234                          */
3235                         ret = ret < 0 ? ret : -EIO;
3236                         goto done;
3237                 }
3238                 /*
3239                  * delalloc_end is already one less than the total length, so
3240                  * we don't subtract one from PAGE_SIZE
3241                  */
3242                 delalloc_to_write += (delalloc_end - delalloc_start +
3243                                       PAGE_SIZE) >> PAGE_SHIFT;
3244                 delalloc_start = delalloc_end + 1;
3245         }
3246         if (wbc->nr_to_write < delalloc_to_write) {
3247                 int thresh = 8192;
3248
3249                 if (delalloc_to_write < thresh * 2)
3250                         thresh = delalloc_to_write;
3251                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3252                                          thresh);
3253         }
3254
3255         /* did the fill delalloc function already unlock and start
3256          * the IO?
3257          */
3258         if (page_started) {
3259                 /*
3260                  * we've unlocked the page, so we can't update
3261                  * the mapping's writeback index, just update
3262                  * nr_to_write.
3263                  */
3264                 wbc->nr_to_write -= *nr_written;
3265                 return 1;
3266         }
3267
3268         ret = 0;
3269
3270 done:
3271         return ret;
3272 }
3273
3274 /*
3275  * helper for __extent_writepage.  This calls the writepage start hooks,
3276  * and does the loop to map the page into extents and bios.
3277  *
3278  * We return 1 if the IO is started and the page is unlocked,
3279  * 0 if all went well (page still locked)
3280  * < 0 if there were errors (page still locked)
3281  */
3282 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3283                                  struct page *page,
3284                                  struct writeback_control *wbc,
3285                                  struct extent_page_data *epd,
3286                                  loff_t i_size,
3287                                  unsigned long nr_written,
3288                                  unsigned int write_flags, int *nr_ret)
3289 {
3290         struct extent_io_tree *tree = epd->tree;
3291         u64 start = page_offset(page);
3292         u64 page_end = start + PAGE_SIZE - 1;
3293         u64 end;
3294         u64 cur = start;
3295         u64 extent_offset;
3296         u64 block_start;
3297         u64 iosize;
3298         struct extent_map *em;
3299         struct block_device *bdev;
3300         size_t pg_offset = 0;
3301         size_t blocksize;
3302         int ret = 0;
3303         int nr = 0;
3304         bool compressed;
3305
3306         ret = btrfs_writepage_cow_fixup(page, start, page_end);
3307         if (ret) {
3308                 /* Fixup worker will requeue */
3309                 if (ret == -EBUSY)
3310                         wbc->pages_skipped++;
3311                 else
3312                         redirty_page_for_writepage(wbc, page);
3313
3314                 update_nr_written(wbc, nr_written);
3315                 unlock_page(page);
3316                 return 1;
3317         }
3318
3319         /*
3320          * we don't want to touch the inode after unlocking the page,
3321          * so we update the mapping writeback index now
3322          */
3323         update_nr_written(wbc, nr_written + 1);
3324
3325         end = page_end;
3326         if (i_size <= start) {
3327                 btrfs_writepage_endio_finish_ordered(page, start, page_end, 1);
3328                 goto done;
3329         }
3330
3331         blocksize = inode->i_sb->s_blocksize;
3332
3333         while (cur <= end) {
3334                 u64 em_end;
3335                 u64 offset;
3336
3337                 if (cur >= i_size) {
3338                         btrfs_writepage_endio_finish_ordered(page, cur,
3339                                                              page_end, 1);
3340                         break;
3341                 }
3342                 em = btrfs_get_extent(BTRFS_I(inode), page, pg_offset, cur,
3343                                      end - cur + 1, 1);
3344                 if (IS_ERR_OR_NULL(em)) {
3345                         SetPageError(page);
3346                         ret = PTR_ERR_OR_ZERO(em);
3347                         break;
3348                 }
3349
3350                 extent_offset = cur - em->start;
3351                 em_end = extent_map_end(em);
3352                 BUG_ON(em_end <= cur);
3353                 BUG_ON(end < cur);
3354                 iosize = min(em_end - cur, end - cur + 1);
3355                 iosize = ALIGN(iosize, blocksize);
3356                 offset = em->block_start + extent_offset;
3357                 bdev = em->bdev;
3358                 block_start = em->block_start;
3359                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3360                 free_extent_map(em);
3361                 em = NULL;
3362
3363                 /*
3364                  * compressed and inline extents are written through other
3365                  * paths in the FS
3366                  */
3367                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3368                     block_start == EXTENT_MAP_INLINE) {
3369                         /*
3370                          * end_io notification does not happen here for
3371                          * compressed extents
3372                          */
3373                         if (!compressed)
3374                                 btrfs_writepage_endio_finish_ordered(page, cur,
3375                                                             cur + iosize - 1,
3376                                                             1);
3377                         else if (compressed) {
3378                                 /* we don't want to end_page_writeback on
3379                                  * a compressed extent.  this happens
3380                                  * elsewhere
3381                                  */
3382                                 nr++;
3383                         }
3384
3385                         cur += iosize;
3386                         pg_offset += iosize;
3387                         continue;
3388                 }
3389
3390                 btrfs_set_range_writeback(tree, cur, cur + iosize - 1);
3391                 if (!PageWriteback(page)) {
3392                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3393                                    "page %lu not writeback, cur %llu end %llu",
3394                                page->index, cur, end);
3395                 }
3396
3397                 ret = submit_extent_page(REQ_OP_WRITE | write_flags, tree, wbc,
3398                                          page, offset, iosize, pg_offset,
3399                                          bdev, &epd->bio,
3400                                          end_bio_extent_writepage,
3401                                          0, 0, 0, false);
3402                 if (ret) {
3403                         SetPageError(page);
3404                         if (PageWriteback(page))
3405                                 end_page_writeback(page);
3406                 }
3407
3408                 cur = cur + iosize;
3409                 pg_offset += iosize;
3410                 nr++;
3411         }
3412 done:
3413         *nr_ret = nr;
3414         return ret;
3415 }
3416
3417 /*
3418  * the writepage semantics are similar to regular writepage.  extent
3419  * records are inserted to lock ranges in the tree, and as dirty areas
3420  * are found, they are marked writeback.  Then the lock bits are removed
3421  * and the end_io handler clears the writeback ranges
3422  */
3423 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3424                               struct extent_page_data *epd)
3425 {
3426         struct inode *inode = page->mapping->host;
3427         u64 start = page_offset(page);
3428         u64 page_end = start + PAGE_SIZE - 1;
3429         int ret;
3430         int nr = 0;
3431         size_t pg_offset = 0;
3432         loff_t i_size = i_size_read(inode);
3433         unsigned long end_index = i_size >> PAGE_SHIFT;
3434         unsigned int write_flags = 0;
3435         unsigned long nr_written = 0;
3436
3437         write_flags = wbc_to_write_flags(wbc);
3438
3439         trace___extent_writepage(page, inode, wbc);
3440
3441         WARN_ON(!PageLocked(page));
3442
3443         ClearPageError(page);
3444
3445         pg_offset = offset_in_page(i_size);
3446         if (page->index > end_index ||
3447            (page->index == end_index && !pg_offset)) {
3448                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3449                 unlock_page(page);
3450                 return 0;
3451         }
3452
3453         if (page->index == end_index) {
3454                 char *userpage;
3455
3456                 userpage = kmap_atomic(page);
3457                 memset(userpage + pg_offset, 0,
3458                        PAGE_SIZE - pg_offset);
3459                 kunmap_atomic(userpage);
3460                 flush_dcache_page(page);
3461         }
3462
3463         pg_offset = 0;
3464
3465         set_page_extent_mapped(page);
3466
3467         if (!epd->extent_locked) {
3468                 ret = writepage_delalloc(inode, page, wbc, start, &nr_written);
3469                 if (ret == 1)
3470                         goto done_unlocked;
3471                 if (ret)
3472                         goto done;
3473         }
3474
3475         ret = __extent_writepage_io(inode, page, wbc, epd,
3476                                     i_size, nr_written, write_flags, &nr);
3477         if (ret == 1)
3478                 goto done_unlocked;
3479
3480 done:
3481         if (nr == 0) {
3482                 /* make sure the mapping tag for page dirty gets cleared */
3483                 set_page_writeback(page);
3484                 end_page_writeback(page);
3485         }
3486         if (PageError(page)) {
3487                 ret = ret < 0 ? ret : -EIO;
3488                 end_extent_writepage(page, ret, start, page_end);
3489         }
3490         unlock_page(page);
3491         return ret;
3492
3493 done_unlocked:
3494         return 0;
3495 }
3496
3497 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3498 {
3499         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3500                        TASK_UNINTERRUPTIBLE);
3501 }
3502
3503 static noinline_for_stack int
3504 lock_extent_buffer_for_io(struct extent_buffer *eb,
3505                           struct btrfs_fs_info *fs_info,
3506                           struct extent_page_data *epd)
3507 {
3508         int i, num_pages;
3509         int flush = 0;
3510         int ret = 0;
3511
3512         if (!btrfs_try_tree_write_lock(eb)) {
3513                 flush = 1;
3514                 flush_write_bio(epd);
3515                 btrfs_tree_lock(eb);
3516         }
3517
3518         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3519                 btrfs_tree_unlock(eb);
3520                 if (!epd->sync_io)
3521                         return 0;
3522                 if (!flush) {
3523                         flush_write_bio(epd);
3524                         flush = 1;
3525                 }
3526                 while (1) {
3527                         wait_on_extent_buffer_writeback(eb);
3528                         btrfs_tree_lock(eb);
3529                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3530                                 break;
3531                         btrfs_tree_unlock(eb);
3532                 }
3533         }
3534
3535         /*
3536          * We need to do this to prevent races in people who check if the eb is
3537          * under IO since we can end up having no IO bits set for a short period
3538          * of time.
3539          */
3540         spin_lock(&eb->refs_lock);
3541         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3542                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3543                 spin_unlock(&eb->refs_lock);
3544                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3545                 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
3546                                          -eb->len,
3547                                          fs_info->dirty_metadata_batch);
3548                 ret = 1;
3549         } else {
3550                 spin_unlock(&eb->refs_lock);
3551         }
3552
3553         btrfs_tree_unlock(eb);
3554
3555         if (!ret)
3556                 return ret;
3557
3558         num_pages = num_extent_pages(eb);
3559         for (i = 0; i < num_pages; i++) {
3560                 struct page *p = eb->pages[i];
3561
3562                 if (!trylock_page(p)) {
3563                         if (!flush) {
3564                                 flush_write_bio(epd);
3565                                 flush = 1;
3566                         }
3567                         lock_page(p);
3568                 }
3569         }
3570
3571         return ret;
3572 }
3573
3574 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3575 {
3576         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3577         smp_mb__after_atomic();
3578         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3579 }
3580
3581 static void set_btree_ioerr(struct page *page)
3582 {
3583         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3584
3585         SetPageError(page);
3586         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3587                 return;
3588
3589         /*
3590          * If writeback for a btree extent that doesn't belong to a log tree
3591          * failed, increment the counter transaction->eb_write_errors.
3592          * We do this because while the transaction is running and before it's
3593          * committing (when we call filemap_fdata[write|wait]_range against
3594          * the btree inode), we might have
3595          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3596          * returns an error or an error happens during writeback, when we're
3597          * committing the transaction we wouldn't know about it, since the pages
3598          * can be no longer dirty nor marked anymore for writeback (if a
3599          * subsequent modification to the extent buffer didn't happen before the
3600          * transaction commit), which makes filemap_fdata[write|wait]_range not
3601          * able to find the pages tagged with SetPageError at transaction
3602          * commit time. So if this happens we must abort the transaction,
3603          * otherwise we commit a super block with btree roots that point to
3604          * btree nodes/leafs whose content on disk is invalid - either garbage
3605          * or the content of some node/leaf from a past generation that got
3606          * cowed or deleted and is no longer valid.
3607          *
3608          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3609          * not be enough - we need to distinguish between log tree extents vs
3610          * non-log tree extents, and the next filemap_fdatawait_range() call
3611          * will catch and clear such errors in the mapping - and that call might
3612          * be from a log sync and not from a transaction commit. Also, checking
3613          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3614          * not done and would not be reliable - the eb might have been released
3615          * from memory and reading it back again means that flag would not be
3616          * set (since it's a runtime flag, not persisted on disk).
3617          *
3618          * Using the flags below in the btree inode also makes us achieve the
3619          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3620          * writeback for all dirty pages and before filemap_fdatawait_range()
3621          * is called, the writeback for all dirty pages had already finished
3622          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3623          * filemap_fdatawait_range() would return success, as it could not know
3624          * that writeback errors happened (the pages were no longer tagged for
3625          * writeback).
3626          */
3627         switch (eb->log_index) {
3628         case -1:
3629                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3630                 break;
3631         case 0:
3632                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3633                 break;
3634         case 1:
3635                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3636                 break;
3637         default:
3638                 BUG(); /* unexpected, logic error */
3639         }
3640 }
3641
3642 static void end_bio_extent_buffer_writepage(struct bio *bio)
3643 {
3644         struct bio_vec *bvec;
3645         struct extent_buffer *eb;
3646         int i, done;
3647         struct bvec_iter_all iter_all;
3648
3649         ASSERT(!bio_flagged(bio, BIO_CLONED));
3650         bio_for_each_segment_all(bvec, bio, i, iter_all) {
3651                 struct page *page = bvec->bv_page;
3652
3653                 eb = (struct extent_buffer *)page->private;
3654                 BUG_ON(!eb);
3655                 done = atomic_dec_and_test(&eb->io_pages);
3656
3657                 if (bio->bi_status ||
3658                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3659                         ClearPageUptodate(page);
3660                         set_btree_ioerr(page);
3661                 }
3662
3663                 end_page_writeback(page);
3664
3665                 if (!done)
3666                         continue;
3667
3668                 end_extent_buffer_writeback(eb);
3669         }
3670
3671         bio_put(bio);
3672 }
3673
3674 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3675                         struct btrfs_fs_info *fs_info,
3676                         struct writeback_control *wbc,
3677                         struct extent_page_data *epd)
3678 {
3679         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3680         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3681         u64 offset = eb->start;
3682         u32 nritems;
3683         int i, num_pages;
3684         unsigned long start, end;
3685         unsigned int write_flags = wbc_to_write_flags(wbc) | REQ_META;
3686         int ret = 0;
3687
3688         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3689         num_pages = num_extent_pages(eb);
3690         atomic_set(&eb->io_pages, num_pages);
3691
3692         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3693         nritems = btrfs_header_nritems(eb);
3694         if (btrfs_header_level(eb) > 0) {
3695                 end = btrfs_node_key_ptr_offset(nritems);
3696
3697                 memzero_extent_buffer(eb, end, eb->len - end);
3698         } else {
3699                 /*
3700                  * leaf:
3701                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3702                  */
3703                 start = btrfs_item_nr_offset(nritems);
3704                 end = BTRFS_LEAF_DATA_OFFSET + leaf_data_end(fs_info, eb);
3705                 memzero_extent_buffer(eb, start, end - start);
3706         }
3707
3708         for (i = 0; i < num_pages; i++) {
3709                 struct page *p = eb->pages[i];
3710
3711                 clear_page_dirty_for_io(p);
3712                 set_page_writeback(p);
3713                 ret = submit_extent_page(REQ_OP_WRITE | write_flags, tree, wbc,
3714                                          p, offset, PAGE_SIZE, 0, bdev,
3715                                          &epd->bio,
3716                                          end_bio_extent_buffer_writepage,
3717                                          0, 0, 0, false);
3718                 if (ret) {
3719                         set_btree_ioerr(p);
3720                         if (PageWriteback(p))
3721                                 end_page_writeback(p);
3722                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3723                                 end_extent_buffer_writeback(eb);
3724                         ret = -EIO;
3725                         break;
3726                 }
3727                 offset += PAGE_SIZE;
3728                 update_nr_written(wbc, 1);
3729                 unlock_page(p);
3730         }
3731
3732         if (unlikely(ret)) {
3733                 for (; i < num_pages; i++) {
3734                         struct page *p = eb->pages[i];
3735                         clear_page_dirty_for_io(p);
3736                         unlock_page(p);
3737                 }
3738         }
3739
3740         return ret;
3741 }
3742
3743 int btree_write_cache_pages(struct address_space *mapping,
3744                                    struct writeback_control *wbc)
3745 {
3746         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3747         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3748         struct extent_buffer *eb, *prev_eb = NULL;
3749         struct extent_page_data epd = {
3750                 .bio = NULL,
3751                 .tree = tree,
3752                 .extent_locked = 0,
3753                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3754         };
3755         int ret = 0;
3756         int done = 0;
3757         int nr_to_write_done = 0;
3758         struct pagevec pvec;
3759         int nr_pages;
3760         pgoff_t index;
3761         pgoff_t end;            /* Inclusive */
3762         int scanned = 0;
3763         xa_mark_t tag;
3764
3765         pagevec_init(&pvec);
3766         if (wbc->range_cyclic) {
3767                 index = mapping->writeback_index; /* Start from prev offset */
3768                 end = -1;
3769         } else {
3770                 index = wbc->range_start >> PAGE_SHIFT;
3771                 end = wbc->range_end >> PAGE_SHIFT;
3772                 scanned = 1;
3773         }
3774         if (wbc->sync_mode == WB_SYNC_ALL)
3775                 tag = PAGECACHE_TAG_TOWRITE;
3776         else
3777                 tag = PAGECACHE_TAG_DIRTY;
3778 retry:
3779         if (wbc->sync_mode == WB_SYNC_ALL)
3780                 tag_pages_for_writeback(mapping, index, end);
3781         while (!done && !nr_to_write_done && (index <= end) &&
3782                (nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
3783                         tag))) {
3784                 unsigned i;
3785
3786                 scanned = 1;
3787                 for (i = 0; i < nr_pages; i++) {
3788                         struct page *page = pvec.pages[i];
3789
3790                         if (!PagePrivate(page))
3791                                 continue;
3792
3793                         spin_lock(&mapping->private_lock);
3794                         if (!PagePrivate(page)) {
3795                                 spin_unlock(&mapping->private_lock);
3796                                 continue;
3797                         }
3798
3799                         eb = (struct extent_buffer *)page->private;
3800
3801                         /*
3802                          * Shouldn't happen and normally this would be a BUG_ON
3803                          * but no sense in crashing the users box for something
3804                          * we can survive anyway.
3805                          */
3806                         if (WARN_ON(!eb)) {
3807                                 spin_unlock(&mapping->private_lock);
3808                                 continue;
3809                         }
3810
3811                         if (eb == prev_eb) {
3812                                 spin_unlock(&mapping->private_lock);
3813                                 continue;
3814                         }
3815
3816                         ret = atomic_inc_not_zero(&eb->refs);
3817                         spin_unlock(&mapping->private_lock);
3818                         if (!ret)
3819                                 continue;
3820
3821                         prev_eb = eb;
3822                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3823                         if (!ret) {
3824                                 free_extent_buffer(eb);
3825                                 continue;
3826                         }
3827
3828                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3829                         if (ret) {
3830                                 done = 1;
3831                                 free_extent_buffer(eb);
3832                                 break;
3833                         }
3834                         free_extent_buffer(eb);
3835
3836                         /*
3837                          * the filesystem may choose to bump up nr_to_write.
3838                          * We have to make sure to honor the new nr_to_write
3839                          * at any time
3840                          */
3841                         nr_to_write_done = wbc->nr_to_write <= 0;
3842                 }
3843                 pagevec_release(&pvec);
3844                 cond_resched();
3845         }
3846         if (!scanned && !done) {
3847                 /*
3848                  * We hit the last page and there is more work to be done: wrap
3849                  * back to the start of the file
3850                  */
3851                 scanned = 1;
3852                 index = 0;
3853                 goto retry;
3854         }
3855         flush_write_bio(&epd);
3856         return ret;
3857 }
3858
3859 /**
3860  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3861  * @mapping: address space structure to write
3862  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3863  * @data: data passed to __extent_writepage function
3864  *
3865  * If a page is already under I/O, write_cache_pages() skips it, even
3866  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3867  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3868  * and msync() need to guarantee that all the data which was dirty at the time
3869  * the call was made get new I/O started against them.  If wbc->sync_mode is
3870  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3871  * existing IO to complete.
3872  */
3873 static int extent_write_cache_pages(struct address_space *mapping,
3874                              struct writeback_control *wbc,
3875                              struct extent_page_data *epd)
3876 {
3877         struct inode *inode = mapping->host;
3878         int ret = 0;
3879         int done = 0;
3880         int nr_to_write_done = 0;
3881         struct pagevec pvec;
3882         int nr_pages;
3883         pgoff_t index;
3884         pgoff_t end;            /* Inclusive */
3885         pgoff_t done_index;
3886         int range_whole = 0;
3887         int scanned = 0;
3888         xa_mark_t tag;
3889
3890         /*
3891          * We have to hold onto the inode so that ordered extents can do their
3892          * work when the IO finishes.  The alternative to this is failing to add
3893          * an ordered extent if the igrab() fails there and that is a huge pain
3894          * to deal with, so instead just hold onto the inode throughout the
3895          * writepages operation.  If it fails here we are freeing up the inode
3896          * anyway and we'd rather not waste our time writing out stuff that is
3897          * going to be truncated anyway.
3898          */
3899         if (!igrab(inode))
3900                 return 0;
3901
3902         pagevec_init(&pvec);
3903         if (wbc->range_cyclic) {
3904                 index = mapping->writeback_index; /* Start from prev offset */
3905                 end = -1;
3906         } else {
3907                 index = wbc->range_start >> PAGE_SHIFT;
3908                 end = wbc->range_end >> PAGE_SHIFT;
3909                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3910                         range_whole = 1;
3911                 scanned = 1;
3912         }
3913
3914         /*
3915          * We do the tagged writepage as long as the snapshot flush bit is set
3916          * and we are the first one who do the filemap_flush() on this inode.
3917          *
3918          * The nr_to_write == LONG_MAX is needed to make sure other flushers do
3919          * not race in and drop the bit.
3920          */
3921         if (range_whole && wbc->nr_to_write == LONG_MAX &&
3922             test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
3923                                &BTRFS_I(inode)->runtime_flags))
3924                 wbc->tagged_writepages = 1;
3925
3926         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
3927                 tag = PAGECACHE_TAG_TOWRITE;
3928         else
3929                 tag = PAGECACHE_TAG_DIRTY;
3930 retry:
3931         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
3932                 tag_pages_for_writeback(mapping, index, end);
3933         done_index = index;
3934         while (!done && !nr_to_write_done && (index <= end) &&
3935                         (nr_pages = pagevec_lookup_range_tag(&pvec, mapping,
3936                                                 &index, end, tag))) {
3937                 unsigned i;
3938
3939                 scanned = 1;
3940                 for (i = 0; i < nr_pages; i++) {
3941                         struct page *page = pvec.pages[i];
3942
3943                         done_index = page->index;
3944                         /*
3945                          * At this point we hold neither the i_pages lock nor
3946                          * the page lock: the page may be truncated or
3947                          * invalidated (changing page->mapping to NULL),
3948                          * or even swizzled back from swapper_space to
3949                          * tmpfs file mapping
3950                          */
3951                         if (!trylock_page(page)) {
3952                                 flush_write_bio(epd);
3953                                 lock_page(page);
3954                         }
3955
3956                         if (unlikely(page->mapping != mapping)) {
3957                                 unlock_page(page);
3958                                 continue;
3959                         }
3960
3961                         if (wbc->sync_mode != WB_SYNC_NONE) {
3962                                 if (PageWriteback(page))
3963                                         flush_write_bio(epd);
3964                                 wait_on_page_writeback(page);
3965                         }
3966
3967                         if (PageWriteback(page) ||
3968                             !clear_page_dirty_for_io(page)) {
3969                                 unlock_page(page);
3970                                 continue;
3971                         }
3972
3973                         ret = __extent_writepage(page, wbc, epd);
3974
3975                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3976                                 unlock_page(page);
3977                                 ret = 0;
3978                         }
3979                         if (ret < 0) {
3980                                 /*
3981                                  * done_index is set past this page,
3982                                  * so media errors will not choke
3983                                  * background writeout for the entire
3984                                  * file. This has consequences for
3985                                  * range_cyclic semantics (ie. it may
3986                                  * not be suitable for data integrity
3987                                  * writeout).
3988                                  */
3989                                 done_index = page->index + 1;
3990                                 done = 1;
3991                                 break;
3992                         }
3993
3994                         /*
3995                          * the filesystem may choose to bump up nr_to_write.
3996                          * We have to make sure to honor the new nr_to_write
3997                          * at any time
3998                          */
3999                         nr_to_write_done = wbc->nr_to_write <= 0;
4000                 }
4001                 pagevec_release(&pvec);
4002                 cond_resched();
4003         }
4004         if (!scanned && !done) {
4005                 /*
4006                  * We hit the last page and there is more work to be done: wrap
4007                  * back to the start of the file
4008                  */
4009                 scanned = 1;
4010                 index = 0;
4011                 goto retry;
4012         }
4013
4014         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4015                 mapping->writeback_index = done_index;
4016
4017         btrfs_add_delayed_iput(inode);
4018         return ret;
4019 }
4020
4021 int extent_write_full_page(struct page *page, struct writeback_control *wbc)
4022 {
4023         int ret;
4024         struct extent_page_data epd = {
4025                 .bio = NULL,
4026                 .tree = &BTRFS_I(page->mapping->host)->io_tree,
4027                 .extent_locked = 0,
4028                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4029         };
4030
4031         ret = __extent_writepage(page, wbc, &epd);
4032
4033         flush_write_bio(&epd);
4034         return ret;
4035 }
4036
4037 int extent_write_locked_range(struct inode *inode, u64 start, u64 end,
4038                               int mode)
4039 {
4040         int ret = 0;
4041         struct address_space *mapping = inode->i_mapping;
4042         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
4043         struct page *page;
4044         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4045                 PAGE_SHIFT;
4046
4047         struct extent_page_data epd = {
4048                 .bio = NULL,
4049                 .tree = tree,
4050                 .extent_locked = 1,
4051                 .sync_io = mode == WB_SYNC_ALL,
4052         };
4053         struct writeback_control wbc_writepages = {
4054                 .sync_mode      = mode,
4055                 .nr_to_write    = nr_pages * 2,
4056                 .range_start    = start,
4057                 .range_end      = end + 1,
4058         };
4059
4060         while (start <= end) {
4061                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4062                 if (clear_page_dirty_for_io(page))
4063                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4064                 else {
4065                         btrfs_writepage_endio_finish_ordered(page, start,
4066                                                     start + PAGE_SIZE - 1, 1);
4067                         unlock_page(page);
4068                 }
4069                 put_page(page);
4070                 start += PAGE_SIZE;
4071         }
4072
4073         flush_write_bio(&epd);
4074         return ret;
4075 }
4076
4077 int extent_writepages(struct address_space *mapping,
4078                       struct writeback_control *wbc)
4079 {
4080         int ret = 0;
4081         struct extent_page_data epd = {
4082                 .bio = NULL,
4083                 .tree = &BTRFS_I(mapping->host)->io_tree,
4084                 .extent_locked = 0,
4085                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4086         };
4087
4088         ret = extent_write_cache_pages(mapping, wbc, &epd);
4089         flush_write_bio(&epd);
4090         return ret;
4091 }
4092
4093 int extent_readpages(struct address_space *mapping, struct list_head *pages,
4094                      unsigned nr_pages)
4095 {
4096         struct bio *bio = NULL;
4097         unsigned long bio_flags = 0;
4098         struct page *pagepool[16];
4099         struct extent_map *em_cached = NULL;
4100         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
4101         int nr = 0;
4102         u64 prev_em_start = (u64)-1;
4103
4104         while (!list_empty(pages)) {
4105                 for (nr = 0; nr < ARRAY_SIZE(pagepool) && !list_empty(pages);) {
4106                         struct page *page = lru_to_page(pages);
4107
4108                         prefetchw(&page->flags);
4109                         list_del(&page->lru);
4110                         if (add_to_page_cache_lru(page, mapping, page->index,
4111                                                 readahead_gfp_mask(mapping))) {
4112                                 put_page(page);
4113                                 continue;
4114                         }
4115
4116                         pagepool[nr++] = page;
4117                 }
4118
4119                 __extent_readpages(tree, pagepool, nr, &em_cached, &bio,
4120                                    &bio_flags, &prev_em_start);
4121         }
4122
4123         if (em_cached)
4124                 free_extent_map(em_cached);
4125
4126         if (bio)
4127                 return submit_one_bio(bio, 0, bio_flags);
4128         return 0;
4129 }
4130
4131 /*
4132  * basic invalidatepage code, this waits on any locked or writeback
4133  * ranges corresponding to the page, and then deletes any extent state
4134  * records from the tree
4135  */
4136 int extent_invalidatepage(struct extent_io_tree *tree,
4137                           struct page *page, unsigned long offset)
4138 {
4139         struct extent_state *cached_state = NULL;
4140         u64 start = page_offset(page);
4141         u64 end = start + PAGE_SIZE - 1;
4142         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4143
4144         start += ALIGN(offset, blocksize);
4145         if (start > end)
4146                 return 0;
4147
4148         lock_extent_bits(tree, start, end, &cached_state);
4149         wait_on_page_writeback(page);
4150         clear_extent_bit(tree, start, end,
4151                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4152                          EXTENT_DO_ACCOUNTING,
4153                          1, 1, &cached_state);
4154         return 0;
4155 }
4156
4157 /*
4158  * a helper for releasepage, this tests for areas of the page that
4159  * are locked or under IO and drops the related state bits if it is safe
4160  * to drop the page.
4161  */
4162 static int try_release_extent_state(struct extent_io_tree *tree,
4163                                     struct page *page, gfp_t mask)
4164 {
4165         u64 start = page_offset(page);
4166         u64 end = start + PAGE_SIZE - 1;
4167         int ret = 1;
4168
4169         if (test_range_bit(tree, start, end,
4170                            EXTENT_IOBITS, 0, NULL))
4171                 ret = 0;
4172         else {
4173                 /*
4174                  * at this point we can safely clear everything except the
4175                  * locked bit and the nodatasum bit
4176                  */
4177                 ret = __clear_extent_bit(tree, start, end,
4178                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4179                                  0, 0, NULL, mask, NULL);
4180
4181                 /* if clear_extent_bit failed for enomem reasons,
4182                  * we can't allow the release to continue.
4183                  */
4184                 if (ret < 0)
4185                         ret = 0;
4186                 else
4187                         ret = 1;
4188         }
4189         return ret;
4190 }
4191
4192 /*
4193  * a helper for releasepage.  As long as there are no locked extents
4194  * in the range corresponding to the page, both state records and extent
4195  * map records are removed
4196  */
4197 int try_release_extent_mapping(struct page *page, gfp_t mask)
4198 {
4199         struct extent_map *em;
4200         u64 start = page_offset(page);
4201         u64 end = start + PAGE_SIZE - 1;
4202         struct btrfs_inode *btrfs_inode = BTRFS_I(page->mapping->host);
4203         struct extent_io_tree *tree = &btrfs_inode->io_tree;
4204         struct extent_map_tree *map = &btrfs_inode->extent_tree;
4205
4206         if (gfpflags_allow_blocking(mask) &&
4207             page->mapping->host->i_size > SZ_16M) {
4208                 u64 len;
4209                 while (start <= end) {
4210                         len = end - start + 1;
4211                         write_lock(&map->lock);
4212                         em = lookup_extent_mapping(map, start, len);
4213                         if (!em) {
4214                                 write_unlock(&map->lock);
4215                                 break;
4216                         }
4217                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4218                             em->start != start) {
4219                                 write_unlock(&map->lock);
4220                                 free_extent_map(em);
4221                                 break;
4222                         }
4223                         if (!test_range_bit(tree, em->start,
4224                                             extent_map_end(em) - 1,
4225                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4226                                             0, NULL)) {
4227                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4228                                         &btrfs_inode->runtime_flags);
4229                                 remove_extent_mapping(map, em);
4230                                 /* once for the rb tree */
4231                                 free_extent_map(em);
4232                         }
4233                         start = extent_map_end(em);
4234                         write_unlock(&map->lock);
4235
4236                         /* once for us */
4237                         free_extent_map(em);
4238                 }
4239         }
4240         return try_release_extent_state(tree, page, mask);
4241 }
4242
4243 /*
4244  * helper function for fiemap, which doesn't want to see any holes.
4245  * This maps until we find something past 'last'
4246  */
4247 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4248                                                 u64 offset, u64 last)
4249 {
4250         u64 sectorsize = btrfs_inode_sectorsize(inode);
4251         struct extent_map *em;
4252         u64 len;
4253
4254         if (offset >= last)
4255                 return NULL;
4256
4257         while (1) {
4258                 len = last - offset;
4259                 if (len == 0)
4260                         break;
4261                 len = ALIGN(len, sectorsize);
4262                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), offset, len);
4263                 if (IS_ERR_OR_NULL(em))
4264                         return em;
4265
4266                 /* if this isn't a hole return it */
4267                 if (em->block_start != EXTENT_MAP_HOLE)
4268                         return em;
4269
4270                 /* this is a hole, advance to the next extent */
4271                 offset = extent_map_end(em);
4272                 free_extent_map(em);
4273                 if (offset >= last)
4274                         break;
4275         }
4276         return NULL;
4277 }
4278
4279 /*
4280  * To cache previous fiemap extent
4281  *
4282  * Will be used for merging fiemap extent
4283  */
4284 struct fiemap_cache {
4285         u64 offset;
4286         u64 phys;
4287         u64 len;
4288         u32 flags;
4289         bool cached;
4290 };
4291
4292 /*
4293  * Helper to submit fiemap extent.
4294  *
4295  * Will try to merge current fiemap extent specified by @offset, @phys,
4296  * @len and @flags with cached one.
4297  * And only when we fails to merge, cached one will be submitted as
4298  * fiemap extent.
4299  *
4300  * Return value is the same as fiemap_fill_next_extent().
4301  */
4302 static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
4303                                 struct fiemap_cache *cache,
4304                                 u64 offset, u64 phys, u64 len, u32 flags)
4305 {
4306         int ret = 0;
4307
4308         if (!cache->cached)
4309                 goto assign;
4310
4311         /*
4312          * Sanity check, extent_fiemap() should have ensured that new
4313          * fiemap extent won't overlap with cached one.
4314          * Not recoverable.
4315          *
4316          * NOTE: Physical address can overlap, due to compression
4317          */
4318         if (cache->offset + cache->len > offset) {
4319                 WARN_ON(1);
4320                 return -EINVAL;
4321         }
4322
4323         /*
4324          * Only merges fiemap extents if
4325          * 1) Their logical addresses are continuous
4326          *
4327          * 2) Their physical addresses are continuous
4328          *    So truly compressed (physical size smaller than logical size)
4329          *    extents won't get merged with each other
4330          *
4331          * 3) Share same flags except FIEMAP_EXTENT_LAST
4332          *    So regular extent won't get merged with prealloc extent
4333          */
4334         if (cache->offset + cache->len  == offset &&
4335             cache->phys + cache->len == phys  &&
4336             (cache->flags & ~FIEMAP_EXTENT_LAST) ==
4337                         (flags & ~FIEMAP_EXTENT_LAST)) {
4338                 cache->len += len;
4339                 cache->flags |= flags;
4340                 goto try_submit_last;
4341         }
4342
4343         /* Not mergeable, need to submit cached one */
4344         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4345                                       cache->len, cache->flags);
4346         cache->cached = false;
4347         if (ret)
4348                 return ret;
4349 assign:
4350         cache->cached = true;
4351         cache->offset = offset;
4352         cache->phys = phys;
4353         cache->len = len;
4354         cache->flags = flags;
4355 try_submit_last:
4356         if (cache->flags & FIEMAP_EXTENT_LAST) {
4357                 ret = fiemap_fill_next_extent(fieinfo, cache->offset,
4358                                 cache->phys, cache->len, cache->flags);
4359                 cache->cached = false;
4360         }
4361         return ret;
4362 }
4363
4364 /*
4365  * Emit last fiemap cache
4366  *
4367  * The last fiemap cache may still be cached in the following case:
4368  * 0                  4k                    8k
4369  * |<- Fiemap range ->|
4370  * |<------------  First extent ----------->|
4371  *
4372  * In this case, the first extent range will be cached but not emitted.
4373  * So we must emit it before ending extent_fiemap().
4374  */
4375 static int emit_last_fiemap_cache(struct btrfs_fs_info *fs_info,
4376                                   struct fiemap_extent_info *fieinfo,
4377                                   struct fiemap_cache *cache)
4378 {
4379         int ret;
4380
4381         if (!cache->cached)
4382                 return 0;
4383
4384         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4385                                       cache->len, cache->flags);
4386         cache->cached = false;
4387         if (ret > 0)
4388                 ret = 0;
4389         return ret;
4390 }
4391
4392 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4393                 __u64 start, __u64 len)
4394 {
4395         int ret = 0;
4396         u64 off = start;
4397         u64 max = start + len;
4398         u32 flags = 0;
4399         u32 found_type;
4400         u64 last;
4401         u64 last_for_get_extent = 0;
4402         u64 disko = 0;
4403         u64 isize = i_size_read(inode);
4404         struct btrfs_key found_key;
4405         struct extent_map *em = NULL;
4406         struct extent_state *cached_state = NULL;
4407         struct btrfs_path *path;
4408         struct btrfs_root *root = BTRFS_I(inode)->root;
4409         struct fiemap_cache cache = { 0 };
4410         int end = 0;
4411         u64 em_start = 0;
4412         u64 em_len = 0;
4413         u64 em_end = 0;
4414
4415         if (len == 0)
4416                 return -EINVAL;
4417
4418         path = btrfs_alloc_path();
4419         if (!path)
4420                 return -ENOMEM;
4421         path->leave_spinning = 1;
4422
4423         start = round_down(start, btrfs_inode_sectorsize(inode));
4424         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4425
4426         /*
4427          * lookup the last file extent.  We're not using i_size here
4428          * because there might be preallocation past i_size
4429          */
4430         ret = btrfs_lookup_file_extent(NULL, root, path,
4431                         btrfs_ino(BTRFS_I(inode)), -1, 0);
4432         if (ret < 0) {
4433                 btrfs_free_path(path);
4434                 return ret;
4435         } else {
4436                 WARN_ON(!ret);
4437                 if (ret == 1)
4438                         ret = 0;
4439         }
4440
4441         path->slots[0]--;
4442         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4443         found_type = found_key.type;
4444
4445         /* No extents, but there might be delalloc bits */
4446         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4447             found_type != BTRFS_EXTENT_DATA_KEY) {
4448                 /* have to trust i_size as the end */
4449                 last = (u64)-1;
4450                 last_for_get_extent = isize;
4451         } else {
4452                 /*
4453                  * remember the start of the last extent.  There are a
4454                  * bunch of different factors that go into the length of the
4455                  * extent, so its much less complex to remember where it started
4456                  */
4457                 last = found_key.offset;
4458                 last_for_get_extent = last + 1;
4459         }
4460         btrfs_release_path(path);
4461
4462         /*
4463          * we might have some extents allocated but more delalloc past those
4464          * extents.  so, we trust isize unless the start of the last extent is
4465          * beyond isize
4466          */
4467         if (last < isize) {
4468                 last = (u64)-1;
4469                 last_for_get_extent = isize;
4470         }
4471
4472         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4473                          &cached_state);
4474
4475         em = get_extent_skip_holes(inode, start, last_for_get_extent);
4476         if (!em)
4477                 goto out;
4478         if (IS_ERR(em)) {
4479                 ret = PTR_ERR(em);
4480                 goto out;
4481         }
4482
4483         while (!end) {
4484                 u64 offset_in_extent = 0;
4485
4486                 /* break if the extent we found is outside the range */
4487                 if (em->start >= max || extent_map_end(em) < off)
4488                         break;
4489
4490                 /*
4491                  * get_extent may return an extent that starts before our
4492                  * requested range.  We have to make sure the ranges
4493                  * we return to fiemap always move forward and don't
4494                  * overlap, so adjust the offsets here
4495                  */
4496                 em_start = max(em->start, off);
4497
4498                 /*
4499                  * record the offset from the start of the extent
4500                  * for adjusting the disk offset below.  Only do this if the
4501                  * extent isn't compressed since our in ram offset may be past
4502                  * what we have actually allocated on disk.
4503                  */
4504                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4505                         offset_in_extent = em_start - em->start;
4506                 em_end = extent_map_end(em);
4507                 em_len = em_end - em_start;
4508                 flags = 0;
4509                 if (em->block_start < EXTENT_MAP_LAST_BYTE)
4510                         disko = em->block_start + offset_in_extent;
4511                 else
4512                         disko = 0;
4513
4514                 /*
4515                  * bump off for our next call to get_extent
4516                  */
4517                 off = extent_map_end(em);
4518                 if (off >= max)
4519                         end = 1;
4520
4521                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4522                         end = 1;
4523                         flags |= FIEMAP_EXTENT_LAST;
4524                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4525                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4526                                   FIEMAP_EXTENT_NOT_ALIGNED);
4527                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4528                         flags |= (FIEMAP_EXTENT_DELALLOC |
4529                                   FIEMAP_EXTENT_UNKNOWN);
4530                 } else if (fieinfo->fi_extents_max) {
4531                         u64 bytenr = em->block_start -
4532                                 (em->start - em->orig_start);
4533
4534                         /*
4535                          * As btrfs supports shared space, this information
4536                          * can be exported to userspace tools via
4537                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4538                          * then we're just getting a count and we can skip the
4539                          * lookup stuff.
4540                          */
4541                         ret = btrfs_check_shared(root,
4542                                                  btrfs_ino(BTRFS_I(inode)),
4543                                                  bytenr);
4544                         if (ret < 0)
4545                                 goto out_free;
4546                         if (ret)
4547                                 flags |= FIEMAP_EXTENT_SHARED;
4548                         ret = 0;
4549                 }
4550                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4551                         flags |= FIEMAP_EXTENT_ENCODED;
4552                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4553                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4554
4555                 free_extent_map(em);
4556                 em = NULL;
4557                 if ((em_start >= last) || em_len == (u64)-1 ||
4558                    (last == (u64)-1 && isize <= em_end)) {
4559                         flags |= FIEMAP_EXTENT_LAST;
4560                         end = 1;
4561                 }
4562
4563                 /* now scan forward to see if this is really the last extent. */
4564                 em = get_extent_skip_holes(inode, off, last_for_get_extent);
4565                 if (IS_ERR(em)) {
4566                         ret = PTR_ERR(em);
4567                         goto out;
4568                 }
4569                 if (!em) {
4570                         flags |= FIEMAP_EXTENT_LAST;
4571                         end = 1;
4572                 }
4573                 ret = emit_fiemap_extent(fieinfo, &cache, em_start, disko,
4574                                            em_len, flags);
4575                 if (ret) {
4576                         if (ret == 1)
4577                                 ret = 0;
4578                         goto out_free;
4579                 }
4580         }
4581 out_free:
4582         if (!ret)
4583                 ret = emit_last_fiemap_cache(root->fs_info, fieinfo, &cache);
4584         free_extent_map(em);
4585 out:
4586         btrfs_free_path(path);
4587         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4588                              &cached_state);
4589         return ret;
4590 }
4591
4592 static void __free_extent_buffer(struct extent_buffer *eb)
4593 {
4594         btrfs_leak_debug_del(&eb->leak_list);
4595         kmem_cache_free(extent_buffer_cache, eb);
4596 }
4597
4598 int extent_buffer_under_io(struct extent_buffer *eb)
4599 {
4600         return (atomic_read(&eb->io_pages) ||
4601                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4602                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4603 }
4604
4605 /*
4606  * Release all pages attached to the extent buffer.
4607  */
4608 static void btrfs_release_extent_buffer_pages(struct extent_buffer *eb)
4609 {
4610         int i;
4611         int num_pages;
4612         int mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
4613
4614         BUG_ON(extent_buffer_under_io(eb));
4615
4616         num_pages = num_extent_pages(eb);
4617         for (i = 0; i < num_pages; i++) {
4618                 struct page *page = eb->pages[i];
4619
4620                 if (!page)
4621                         continue;
4622                 if (mapped)
4623                         spin_lock(&page->mapping->private_lock);
4624                 /*
4625                  * We do this since we'll remove the pages after we've
4626                  * removed the eb from the radix tree, so we could race
4627                  * and have this page now attached to the new eb.  So
4628                  * only clear page_private if it's still connected to
4629                  * this eb.
4630                  */
4631                 if (PagePrivate(page) &&
4632                     page->private == (unsigned long)eb) {
4633                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4634                         BUG_ON(PageDirty(page));
4635                         BUG_ON(PageWriteback(page));
4636                         /*
4637                          * We need to make sure we haven't be attached
4638                          * to a new eb.
4639                          */
4640                         ClearPagePrivate(page);
4641                         set_page_private(page, 0);
4642                         /* One for the page private */
4643                         put_page(page);
4644                 }
4645
4646                 if (mapped)
4647                         spin_unlock(&page->mapping->private_lock);
4648
4649                 /* One for when we allocated the page */
4650                 put_page(page);
4651         }
4652 }
4653
4654 /*
4655  * Helper for releasing the extent buffer.
4656  */
4657 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4658 {
4659         btrfs_release_extent_buffer_pages(eb);
4660         __free_extent_buffer(eb);
4661 }
4662
4663 static struct extent_buffer *
4664 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4665                       unsigned long len)
4666 {
4667         struct extent_buffer *eb = NULL;
4668
4669         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4670         eb->start = start;
4671         eb->len = len;
4672         eb->fs_info = fs_info;
4673         eb->bflags = 0;
4674         rwlock_init(&eb->lock);
4675         atomic_set(&eb->write_locks, 0);
4676         atomic_set(&eb->read_locks, 0);
4677         atomic_set(&eb->blocking_readers, 0);
4678         atomic_set(&eb->blocking_writers, 0);
4679         atomic_set(&eb->spinning_readers, 0);
4680         atomic_set(&eb->spinning_writers, 0);
4681         eb->lock_nested = 0;
4682         init_waitqueue_head(&eb->write_lock_wq);
4683         init_waitqueue_head(&eb->read_lock_wq);
4684
4685         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4686
4687         spin_lock_init(&eb->refs_lock);
4688         atomic_set(&eb->refs, 1);
4689         atomic_set(&eb->io_pages, 0);
4690
4691         /*
4692          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4693          */
4694         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4695                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4696         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4697
4698         return eb;
4699 }
4700
4701 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4702 {
4703         int i;
4704         struct page *p;
4705         struct extent_buffer *new;
4706         int num_pages = num_extent_pages(src);
4707
4708         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4709         if (new == NULL)
4710                 return NULL;
4711
4712         for (i = 0; i < num_pages; i++) {
4713                 p = alloc_page(GFP_NOFS);
4714                 if (!p) {
4715                         btrfs_release_extent_buffer(new);
4716                         return NULL;
4717                 }
4718                 attach_extent_buffer_page(new, p);
4719                 WARN_ON(PageDirty(p));
4720                 SetPageUptodate(p);
4721                 new->pages[i] = p;
4722                 copy_page(page_address(p), page_address(src->pages[i]));
4723         }
4724
4725         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4726         set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
4727
4728         return new;
4729 }
4730
4731 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4732                                                   u64 start, unsigned long len)
4733 {
4734         struct extent_buffer *eb;
4735         int num_pages;
4736         int i;
4737
4738         eb = __alloc_extent_buffer(fs_info, start, len);
4739         if (!eb)
4740                 return NULL;
4741
4742         num_pages = num_extent_pages(eb);
4743         for (i = 0; i < num_pages; i++) {
4744                 eb->pages[i] = alloc_page(GFP_NOFS);
4745                 if (!eb->pages[i])
4746                         goto err;
4747         }
4748         set_extent_buffer_uptodate(eb);
4749         btrfs_set_header_nritems(eb, 0);
4750         set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
4751
4752         return eb;
4753 err:
4754         for (; i > 0; i--)
4755                 __free_page(eb->pages[i - 1]);
4756         __free_extent_buffer(eb);
4757         return NULL;
4758 }
4759
4760 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4761                                                 u64 start)
4762 {
4763         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4764 }
4765
4766 static void check_buffer_tree_ref(struct extent_buffer *eb)
4767 {
4768         int refs;
4769         /* the ref bit is tricky.  We have to make sure it is set
4770          * if we have the buffer dirty.   Otherwise the
4771          * code to free a buffer can end up dropping a dirty
4772          * page
4773          *
4774          * Once the ref bit is set, it won't go away while the
4775          * buffer is dirty or in writeback, and it also won't
4776          * go away while we have the reference count on the
4777          * eb bumped.
4778          *
4779          * We can't just set the ref bit without bumping the
4780          * ref on the eb because free_extent_buffer might
4781          * see the ref bit and try to clear it.  If this happens
4782          * free_extent_buffer might end up dropping our original
4783          * ref by mistake and freeing the page before we are able
4784          * to add one more ref.
4785          *
4786          * So bump the ref count first, then set the bit.  If someone
4787          * beat us to it, drop the ref we added.
4788          */
4789         refs = atomic_read(&eb->refs);
4790         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4791                 return;
4792
4793         spin_lock(&eb->refs_lock);
4794         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4795                 atomic_inc(&eb->refs);
4796         spin_unlock(&eb->refs_lock);
4797 }
4798
4799 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4800                 struct page *accessed)
4801 {
4802         int num_pages, i;
4803
4804         check_buffer_tree_ref(eb);
4805
4806         num_pages = num_extent_pages(eb);
4807         for (i = 0; i < num_pages; i++) {
4808                 struct page *p = eb->pages[i];
4809
4810                 if (p != accessed)
4811                         mark_page_accessed(p);
4812         }
4813 }
4814
4815 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4816                                          u64 start)
4817 {
4818         struct extent_buffer *eb;
4819
4820         rcu_read_lock();
4821         eb = radix_tree_lookup(&fs_info->buffer_radix,
4822                                start >> PAGE_SHIFT);
4823         if (eb && atomic_inc_not_zero(&eb->refs)) {
4824                 rcu_read_unlock();
4825                 /*
4826                  * Lock our eb's refs_lock to avoid races with
4827                  * free_extent_buffer. When we get our eb it might be flagged
4828                  * with EXTENT_BUFFER_STALE and another task running
4829                  * free_extent_buffer might have seen that flag set,
4830                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4831                  * writeback flags not set) and it's still in the tree (flag
4832                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4833                  * of decrementing the extent buffer's reference count twice.
4834                  * So here we could race and increment the eb's reference count,
4835                  * clear its stale flag, mark it as dirty and drop our reference
4836                  * before the other task finishes executing free_extent_buffer,
4837                  * which would later result in an attempt to free an extent
4838                  * buffer that is dirty.
4839                  */
4840                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4841                         spin_lock(&eb->refs_lock);
4842                         spin_unlock(&eb->refs_lock);
4843                 }
4844                 mark_extent_buffer_accessed(eb, NULL);
4845                 return eb;
4846         }
4847         rcu_read_unlock();
4848
4849         return NULL;
4850 }
4851
4852 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4853 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4854                                         u64 start)
4855 {
4856         struct extent_buffer *eb, *exists = NULL;
4857         int ret;
4858
4859         eb = find_extent_buffer(fs_info, start);
4860         if (eb)
4861                 return eb;
4862         eb = alloc_dummy_extent_buffer(fs_info, start);
4863         if (!eb)
4864                 return NULL;
4865         eb->fs_info = fs_info;
4866 again:
4867         ret = radix_tree_preload(GFP_NOFS);
4868         if (ret)
4869                 goto free_eb;
4870         spin_lock(&fs_info->buffer_lock);
4871         ret = radix_tree_insert(&fs_info->buffer_radix,
4872                                 start >> PAGE_SHIFT, eb);
4873         spin_unlock(&fs_info->buffer_lock);
4874         radix_tree_preload_end();
4875         if (ret == -EEXIST) {
4876                 exists = find_extent_buffer(fs_info, start);
4877                 if (exists)
4878                         goto free_eb;
4879                 else
4880                         goto again;
4881         }
4882         check_buffer_tree_ref(eb);
4883         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4884
4885         return eb;
4886 free_eb:
4887         btrfs_release_extent_buffer(eb);
4888         return exists;
4889 }
4890 #endif
4891
4892 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4893                                           u64 start)
4894 {
4895         unsigned long len = fs_info->nodesize;
4896         int num_pages;
4897         int i;
4898         unsigned long index = start >> PAGE_SHIFT;
4899         struct extent_buffer *eb;
4900         struct extent_buffer *exists = NULL;
4901         struct page *p;
4902         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4903         int uptodate = 1;
4904         int ret;
4905
4906         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
4907                 btrfs_err(fs_info, "bad tree block start %llu", start);
4908                 return ERR_PTR(-EINVAL);
4909         }
4910
4911         eb = find_extent_buffer(fs_info, start);
4912         if (eb)
4913                 return eb;
4914
4915         eb = __alloc_extent_buffer(fs_info, start, len);
4916         if (!eb)
4917                 return ERR_PTR(-ENOMEM);
4918
4919         num_pages = num_extent_pages(eb);
4920         for (i = 0; i < num_pages; i++, index++) {
4921                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4922                 if (!p) {
4923                         exists = ERR_PTR(-ENOMEM);
4924                         goto free_eb;
4925                 }
4926
4927                 spin_lock(&mapping->private_lock);
4928                 if (PagePrivate(p)) {
4929                         /*
4930                          * We could have already allocated an eb for this page
4931                          * and attached one so lets see if we can get a ref on
4932                          * the existing eb, and if we can we know it's good and
4933                          * we can just return that one, else we know we can just
4934                          * overwrite page->private.
4935                          */
4936                         exists = (struct extent_buffer *)p->private;
4937                         if (atomic_inc_not_zero(&exists->refs)) {
4938                                 spin_unlock(&mapping->private_lock);
4939                                 unlock_page(p);
4940                                 put_page(p);
4941                                 mark_extent_buffer_accessed(exists, p);
4942                                 goto free_eb;
4943                         }
4944                         exists = NULL;
4945
4946                         /*
4947                          * Do this so attach doesn't complain and we need to
4948                          * drop the ref the old guy had.
4949                          */
4950                         ClearPagePrivate(p);
4951                         WARN_ON(PageDirty(p));
4952                         put_page(p);
4953                 }
4954                 attach_extent_buffer_page(eb, p);
4955                 spin_unlock(&mapping->private_lock);
4956                 WARN_ON(PageDirty(p));
4957                 eb->pages[i] = p;
4958                 if (!PageUptodate(p))
4959                         uptodate = 0;
4960
4961                 /*
4962                  * We can't unlock the pages just yet since the extent buffer
4963                  * hasn't been properly inserted in the radix tree, this
4964                  * opens a race with btree_releasepage which can free a page
4965                  * while we are still filling in all pages for the buffer and
4966                  * we could crash.
4967                  */
4968         }
4969         if (uptodate)
4970                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4971 again:
4972         ret = radix_tree_preload(GFP_NOFS);
4973         if (ret) {
4974                 exists = ERR_PTR(ret);
4975                 goto free_eb;
4976         }
4977
4978         spin_lock(&fs_info->buffer_lock);
4979         ret = radix_tree_insert(&fs_info->buffer_radix,
4980                                 start >> PAGE_SHIFT, eb);
4981         spin_unlock(&fs_info->buffer_lock);
4982         radix_tree_preload_end();
4983         if (ret == -EEXIST) {
4984                 exists = find_extent_buffer(fs_info, start);
4985                 if (exists)
4986                         goto free_eb;
4987                 else
4988                         goto again;
4989         }
4990         /* add one reference for the tree */
4991         check_buffer_tree_ref(eb);
4992         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4993
4994         /*
4995          * Now it's safe to unlock the pages because any calls to
4996          * btree_releasepage will correctly detect that a page belongs to a
4997          * live buffer and won't free them prematurely.
4998          */
4999         for (i = 0; i < num_pages; i++)
5000                 unlock_page(eb->pages[i]);
5001         return eb;
5002
5003 free_eb:
5004         WARN_ON(!atomic_dec_and_test(&eb->refs));
5005         for (i = 0; i < num_pages; i++) {
5006                 if (eb->pages[i])
5007                         unlock_page(eb->pages[i]);
5008         }
5009
5010         btrfs_release_extent_buffer(eb);
5011         return exists;
5012 }
5013
5014 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5015 {
5016         struct extent_buffer *eb =
5017                         container_of(head, struct extent_buffer, rcu_head);
5018
5019         __free_extent_buffer(eb);
5020 }
5021
5022 static int release_extent_buffer(struct extent_buffer *eb)
5023 {
5024         lockdep_assert_held(&eb->refs_lock);
5025
5026         WARN_ON(atomic_read(&eb->refs) == 0);
5027         if (atomic_dec_and_test(&eb->refs)) {
5028                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5029                         struct btrfs_fs_info *fs_info = eb->fs_info;
5030
5031                         spin_unlock(&eb->refs_lock);
5032
5033                         spin_lock(&fs_info->buffer_lock);
5034                         radix_tree_delete(&fs_info->buffer_radix,
5035                                           eb->start >> PAGE_SHIFT);
5036                         spin_unlock(&fs_info->buffer_lock);
5037                 } else {
5038                         spin_unlock(&eb->refs_lock);
5039                 }
5040
5041                 /* Should be safe to release our pages at this point */
5042                 btrfs_release_extent_buffer_pages(eb);
5043 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5044                 if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
5045                         __free_extent_buffer(eb);
5046                         return 1;
5047                 }
5048 #endif
5049                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5050                 return 1;
5051         }
5052         spin_unlock(&eb->refs_lock);
5053
5054         return 0;
5055 }
5056
5057 void free_extent_buffer(struct extent_buffer *eb)
5058 {
5059         int refs;
5060         int old;
5061         if (!eb)
5062                 return;
5063
5064         while (1) {
5065                 refs = atomic_read(&eb->refs);
5066                 if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
5067                     || (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
5068                         refs == 1))
5069                         break;
5070                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5071                 if (old == refs)
5072                         return;
5073         }
5074
5075         spin_lock(&eb->refs_lock);
5076         if (atomic_read(&eb->refs) == 2 &&
5077             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5078             !extent_buffer_under_io(eb) &&
5079             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5080                 atomic_dec(&eb->refs);
5081
5082         /*
5083          * I know this is terrible, but it's temporary until we stop tracking
5084          * the uptodate bits and such for the extent buffers.
5085          */
5086         release_extent_buffer(eb);
5087 }
5088
5089 void free_extent_buffer_stale(struct extent_buffer *eb)
5090 {
5091         if (!eb)
5092                 return;
5093
5094         spin_lock(&eb->refs_lock);
5095         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5096
5097         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5098             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5099                 atomic_dec(&eb->refs);
5100         release_extent_buffer(eb);
5101 }
5102
5103 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5104 {
5105         int i;
5106         int num_pages;
5107         struct page *page;
5108
5109         num_pages = num_extent_pages(eb);
5110
5111         for (i = 0; i < num_pages; i++) {
5112                 page = eb->pages[i];
5113                 if (!PageDirty(page))
5114                         continue;
5115
5116                 lock_page(page);
5117                 WARN_ON(!PagePrivate(page));
5118
5119                 clear_page_dirty_for_io(page);
5120                 xa_lock_irq(&page->mapping->i_pages);
5121                 if (!PageDirty(page))
5122                         __xa_clear_mark(&page->mapping->i_pages,
5123                                         page_index(page), PAGECACHE_TAG_DIRTY);
5124                 xa_unlock_irq(&page->mapping->i_pages);
5125                 ClearPageError(page);
5126                 unlock_page(page);
5127         }
5128         WARN_ON(atomic_read(&eb->refs) == 0);
5129 }
5130
5131 bool set_extent_buffer_dirty(struct extent_buffer *eb)
5132 {
5133         int i;
5134         int num_pages;
5135         bool was_dirty;
5136
5137         check_buffer_tree_ref(eb);
5138
5139         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5140
5141         num_pages = num_extent_pages(eb);
5142         WARN_ON(atomic_read(&eb->refs) == 0);
5143         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5144
5145         if (!was_dirty)
5146                 for (i = 0; i < num_pages; i++)
5147                         set_page_dirty(eb->pages[i]);
5148
5149 #ifdef CONFIG_BTRFS_DEBUG
5150         for (i = 0; i < num_pages; i++)
5151                 ASSERT(PageDirty(eb->pages[i]));
5152 #endif
5153
5154         return was_dirty;
5155 }
5156
5157 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5158 {
5159         int i;
5160         struct page *page;
5161         int num_pages;
5162
5163         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5164         num_pages = num_extent_pages(eb);
5165         for (i = 0; i < num_pages; i++) {
5166                 page = eb->pages[i];
5167                 if (page)
5168                         ClearPageUptodate(page);
5169         }
5170 }
5171
5172 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5173 {
5174         int i;
5175         struct page *page;
5176         int num_pages;
5177
5178         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5179         num_pages = num_extent_pages(eb);
5180         for (i = 0; i < num_pages; i++) {
5181                 page = eb->pages[i];
5182                 SetPageUptodate(page);
5183         }
5184 }
5185
5186 int read_extent_buffer_pages(struct extent_io_tree *tree,
5187                              struct extent_buffer *eb, int wait, int mirror_num)
5188 {
5189         int i;
5190         struct page *page;
5191         int err;
5192         int ret = 0;
5193         int locked_pages = 0;
5194         int all_uptodate = 1;
5195         int num_pages;
5196         unsigned long num_reads = 0;
5197         struct bio *bio = NULL;
5198         unsigned long bio_flags = 0;
5199
5200         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5201                 return 0;
5202
5203         num_pages = num_extent_pages(eb);
5204         for (i = 0; i < num_pages; i++) {
5205                 page = eb->pages[i];
5206                 if (wait == WAIT_NONE) {
5207                         if (!trylock_page(page))
5208                                 goto unlock_exit;
5209                 } else {
5210                         lock_page(page);
5211                 }
5212                 locked_pages++;
5213         }
5214         /*
5215          * We need to firstly lock all pages to make sure that
5216          * the uptodate bit of our pages won't be affected by
5217          * clear_extent_buffer_uptodate().
5218          */
5219         for (i = 0; i < num_pages; i++) {
5220                 page = eb->pages[i];
5221                 if (!PageUptodate(page)) {
5222                         num_reads++;
5223                         all_uptodate = 0;
5224                 }
5225         }
5226
5227         if (all_uptodate) {
5228                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5229                 goto unlock_exit;
5230         }
5231
5232         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5233         eb->read_mirror = 0;
5234         atomic_set(&eb->io_pages, num_reads);
5235         for (i = 0; i < num_pages; i++) {
5236                 page = eb->pages[i];
5237
5238                 if (!PageUptodate(page)) {
5239                         if (ret) {
5240                                 atomic_dec(&eb->io_pages);
5241                                 unlock_page(page);
5242                                 continue;
5243                         }
5244
5245                         ClearPageError(page);
5246                         err = __extent_read_full_page(tree, page,
5247                                                       btree_get_extent, &bio,
5248                                                       mirror_num, &bio_flags,
5249                                                       REQ_META);
5250                         if (err) {
5251                                 ret = err;
5252                                 /*
5253                                  * We use &bio in above __extent_read_full_page,
5254                                  * so we ensure that if it returns error, the
5255                                  * current page fails to add itself to bio and
5256                                  * it's been unlocked.
5257                                  *
5258                                  * We must dec io_pages by ourselves.
5259                                  */
5260                                 atomic_dec(&eb->io_pages);
5261                         }
5262                 } else {
5263                         unlock_page(page);
5264                 }
5265         }
5266
5267         if (bio) {
5268                 err = submit_one_bio(bio, mirror_num, bio_flags);
5269                 if (err)
5270                         return err;
5271         }
5272
5273         if (ret || wait != WAIT_COMPLETE)
5274                 return ret;
5275
5276         for (i = 0; i < num_pages; i++) {
5277                 page = eb->pages[i];
5278                 wait_on_page_locked(page);
5279                 if (!PageUptodate(page))
5280                         ret = -EIO;
5281         }
5282
5283         return ret;
5284
5285 unlock_exit:
5286         while (locked_pages > 0) {
5287                 locked_pages--;
5288                 page = eb->pages[locked_pages];
5289                 unlock_page(page);
5290         }
5291         return ret;
5292 }
5293
5294 void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
5295                         unsigned long start, unsigned long len)
5296 {
5297         size_t cur;
5298         size_t offset;
5299         struct page *page;
5300         char *kaddr;
5301         char *dst = (char *)dstv;
5302         size_t start_offset = offset_in_page(eb->start);
5303         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5304
5305         if (start + len > eb->len) {
5306                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5307                      eb->start, eb->len, start, len);
5308                 memset(dst, 0, len);
5309                 return;
5310         }
5311
5312         offset = offset_in_page(start_offset + start);
5313
5314         while (len > 0) {
5315                 page = eb->pages[i];
5316
5317                 cur = min(len, (PAGE_SIZE - offset));
5318                 kaddr = page_address(page);
5319                 memcpy(dst, kaddr + offset, cur);
5320
5321                 dst += cur;
5322                 len -= cur;
5323                 offset = 0;
5324                 i++;
5325         }
5326 }
5327
5328 int read_extent_buffer_to_user(const struct extent_buffer *eb,
5329                                void __user *dstv,
5330                                unsigned long start, unsigned long len)
5331 {
5332         size_t cur;
5333         size_t offset;
5334         struct page *page;
5335         char *kaddr;
5336         char __user *dst = (char __user *)dstv;
5337         size_t start_offset = offset_in_page(eb->start);
5338         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5339         int ret = 0;
5340
5341         WARN_ON(start > eb->len);
5342         WARN_ON(start + len > eb->start + eb->len);
5343
5344         offset = offset_in_page(start_offset + start);
5345
5346         while (len > 0) {
5347                 page = eb->pages[i];
5348
5349                 cur = min(len, (PAGE_SIZE - offset));
5350                 kaddr = page_address(page);
5351                 if (copy_to_user(dst, kaddr + offset, cur)) {
5352                         ret = -EFAULT;
5353                         break;
5354                 }
5355
5356                 dst += cur;
5357                 len -= cur;
5358                 offset = 0;
5359                 i++;
5360         }
5361
5362         return ret;
5363 }
5364
5365 /*
5366  * return 0 if the item is found within a page.
5367  * return 1 if the item spans two pages.
5368  * return -EINVAL otherwise.
5369  */
5370 int map_private_extent_buffer(const struct extent_buffer *eb,
5371                               unsigned long start, unsigned long min_len,
5372                               char **map, unsigned long *map_start,
5373                               unsigned long *map_len)
5374 {
5375         size_t offset;
5376         char *kaddr;
5377         struct page *p;
5378         size_t start_offset = offset_in_page(eb->start);
5379         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5380         unsigned long end_i = (start_offset + start + min_len - 1) >>
5381                 PAGE_SHIFT;
5382
5383         if (start + min_len > eb->len) {
5384                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5385                        eb->start, eb->len, start, min_len);
5386                 return -EINVAL;
5387         }
5388
5389         if (i != end_i)
5390                 return 1;
5391
5392         if (i == 0) {
5393                 offset = start_offset;
5394                 *map_start = 0;
5395         } else {
5396                 offset = 0;
5397                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5398         }
5399
5400         p = eb->pages[i];
5401         kaddr = page_address(p);
5402         *map = kaddr + offset;
5403         *map_len = PAGE_SIZE - offset;
5404         return 0;
5405 }
5406
5407 int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
5408                          unsigned long start, unsigned long len)
5409 {
5410         size_t cur;
5411         size_t offset;
5412         struct page *page;
5413         char *kaddr;
5414         char *ptr = (char *)ptrv;
5415         size_t start_offset = offset_in_page(eb->start);
5416         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5417         int ret = 0;
5418
5419         WARN_ON(start > eb->len);
5420         WARN_ON(start + len > eb->start + eb->len);
5421
5422         offset = offset_in_page(start_offset + start);
5423
5424         while (len > 0) {
5425                 page = eb->pages[i];
5426
5427                 cur = min(len, (PAGE_SIZE - offset));
5428
5429                 kaddr = page_address(page);
5430                 ret = memcmp(ptr, kaddr + offset, cur);
5431                 if (ret)
5432                         break;
5433
5434                 ptr += cur;
5435                 len -= cur;
5436                 offset = 0;
5437                 i++;
5438         }
5439         return ret;
5440 }
5441
5442 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5443                 const void *srcv)
5444 {
5445         char *kaddr;
5446
5447         WARN_ON(!PageUptodate(eb->pages[0]));
5448         kaddr = page_address(eb->pages[0]);
5449         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5450                         BTRFS_FSID_SIZE);
5451 }
5452
5453 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv)
5454 {
5455         char *kaddr;
5456
5457         WARN_ON(!PageUptodate(eb->pages[0]));
5458         kaddr = page_address(eb->pages[0]);
5459         memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5460                         BTRFS_FSID_SIZE);
5461 }
5462
5463 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5464                          unsigned long start, unsigned long len)
5465 {
5466         size_t cur;
5467         size_t offset;
5468         struct page *page;
5469         char *kaddr;
5470         char *src = (char *)srcv;
5471         size_t start_offset = offset_in_page(eb->start);
5472         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5473
5474         WARN_ON(start > eb->len);
5475         WARN_ON(start + len > eb->start + eb->len);
5476
5477         offset = offset_in_page(start_offset + start);
5478
5479         while (len > 0) {
5480                 page = eb->pages[i];
5481                 WARN_ON(!PageUptodate(page));
5482
5483                 cur = min(len, PAGE_SIZE - offset);
5484                 kaddr = page_address(page);
5485                 memcpy(kaddr + offset, src, cur);
5486
5487                 src += cur;
5488                 len -= cur;
5489                 offset = 0;
5490                 i++;
5491         }
5492 }
5493
5494 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5495                 unsigned long len)
5496 {
5497         size_t cur;
5498         size_t offset;
5499         struct page *page;
5500         char *kaddr;
5501         size_t start_offset = offset_in_page(eb->start);
5502         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5503
5504         WARN_ON(start > eb->len);
5505         WARN_ON(start + len > eb->start + eb->len);
5506
5507         offset = offset_in_page(start_offset + start);
5508
5509         while (len > 0) {
5510                 page = eb->pages[i];
5511                 WARN_ON(!PageUptodate(page));
5512
5513                 cur = min(len, PAGE_SIZE - offset);
5514                 kaddr = page_address(page);
5515                 memset(kaddr + offset, 0, cur);
5516
5517                 len -= cur;
5518                 offset = 0;
5519                 i++;
5520         }
5521 }
5522
5523 void copy_extent_buffer_full(struct extent_buffer *dst,
5524                              struct extent_buffer *src)
5525 {
5526         int i;
5527         int num_pages;
5528
5529         ASSERT(dst->len == src->len);
5530
5531         num_pages = num_extent_pages(dst);
5532         for (i = 0; i < num_pages; i++)
5533                 copy_page(page_address(dst->pages[i]),
5534                                 page_address(src->pages[i]));
5535 }
5536
5537 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5538                         unsigned long dst_offset, unsigned long src_offset,
5539                         unsigned long len)
5540 {
5541         u64 dst_len = dst->len;
5542         size_t cur;
5543         size_t offset;
5544         struct page *page;
5545         char *kaddr;
5546         size_t start_offset = offset_in_page(dst->start);
5547         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5548
5549         WARN_ON(src->len != dst_len);
5550
5551         offset = offset_in_page(start_offset + dst_offset);
5552
5553         while (len > 0) {
5554                 page = dst->pages[i];
5555                 WARN_ON(!PageUptodate(page));
5556
5557                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5558
5559                 kaddr = page_address(page);
5560                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5561
5562                 src_offset += cur;
5563                 len -= cur;
5564                 offset = 0;
5565                 i++;
5566         }
5567 }
5568
5569 /*
5570  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5571  * given bit number
5572  * @eb: the extent buffer
5573  * @start: offset of the bitmap item in the extent buffer
5574  * @nr: bit number
5575  * @page_index: return index of the page in the extent buffer that contains the
5576  * given bit number
5577  * @page_offset: return offset into the page given by page_index
5578  *
5579  * This helper hides the ugliness of finding the byte in an extent buffer which
5580  * contains a given bit.
5581  */
5582 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5583                                     unsigned long start, unsigned long nr,
5584                                     unsigned long *page_index,
5585                                     size_t *page_offset)
5586 {
5587         size_t start_offset = offset_in_page(eb->start);
5588         size_t byte_offset = BIT_BYTE(nr);
5589         size_t offset;
5590
5591         /*
5592          * The byte we want is the offset of the extent buffer + the offset of
5593          * the bitmap item in the extent buffer + the offset of the byte in the
5594          * bitmap item.
5595          */
5596         offset = start_offset + start + byte_offset;
5597
5598         *page_index = offset >> PAGE_SHIFT;
5599         *page_offset = offset_in_page(offset);
5600 }
5601
5602 /**
5603  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5604  * @eb: the extent buffer
5605  * @start: offset of the bitmap item in the extent buffer
5606  * @nr: bit number to test
5607  */
5608 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5609                            unsigned long nr)
5610 {
5611         u8 *kaddr;
5612         struct page *page;
5613         unsigned long i;
5614         size_t offset;
5615
5616         eb_bitmap_offset(eb, start, nr, &i, &offset);
5617         page = eb->pages[i];
5618         WARN_ON(!PageUptodate(page));
5619         kaddr = page_address(page);
5620         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5621 }
5622
5623 /**
5624  * extent_buffer_bitmap_set - set an area of a bitmap
5625  * @eb: the extent buffer
5626  * @start: offset of the bitmap item in the extent buffer
5627  * @pos: bit number of the first bit
5628  * @len: number of bits to set
5629  */
5630 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5631                               unsigned long pos, unsigned long len)
5632 {
5633         u8 *kaddr;
5634         struct page *page;
5635         unsigned long i;
5636         size_t offset;
5637         const unsigned int size = pos + len;
5638         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5639         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5640
5641         eb_bitmap_offset(eb, start, pos, &i, &offset);
5642         page = eb->pages[i];
5643         WARN_ON(!PageUptodate(page));
5644         kaddr = page_address(page);
5645
5646         while (len >= bits_to_set) {
5647                 kaddr[offset] |= mask_to_set;
5648                 len -= bits_to_set;
5649                 bits_to_set = BITS_PER_BYTE;
5650                 mask_to_set = ~0;
5651                 if (++offset >= PAGE_SIZE && len > 0) {
5652                         offset = 0;
5653                         page = eb->pages[++i];
5654                         WARN_ON(!PageUptodate(page));
5655                         kaddr = page_address(page);
5656                 }
5657         }
5658         if (len) {
5659                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5660                 kaddr[offset] |= mask_to_set;
5661         }
5662 }
5663
5664
5665 /**
5666  * extent_buffer_bitmap_clear - clear an area of a bitmap
5667  * @eb: the extent buffer
5668  * @start: offset of the bitmap item in the extent buffer
5669  * @pos: bit number of the first bit
5670  * @len: number of bits to clear
5671  */
5672 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5673                                 unsigned long pos, unsigned long len)
5674 {
5675         u8 *kaddr;
5676         struct page *page;
5677         unsigned long i;
5678         size_t offset;
5679         const unsigned int size = pos + len;
5680         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5681         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5682
5683         eb_bitmap_offset(eb, start, pos, &i, &offset);
5684         page = eb->pages[i];
5685         WARN_ON(!PageUptodate(page));
5686         kaddr = page_address(page);
5687
5688         while (len >= bits_to_clear) {
5689                 kaddr[offset] &= ~mask_to_clear;
5690                 len -= bits_to_clear;
5691                 bits_to_clear = BITS_PER_BYTE;
5692                 mask_to_clear = ~0;
5693                 if (++offset >= PAGE_SIZE && len > 0) {
5694                         offset = 0;
5695                         page = eb->pages[++i];
5696                         WARN_ON(!PageUptodate(page));
5697                         kaddr = page_address(page);
5698                 }
5699         }
5700         if (len) {
5701                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5702                 kaddr[offset] &= ~mask_to_clear;
5703         }
5704 }
5705
5706 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5707 {
5708         unsigned long distance = (src > dst) ? src - dst : dst - src;
5709         return distance < len;
5710 }
5711
5712 static void copy_pages(struct page *dst_page, struct page *src_page,
5713                        unsigned long dst_off, unsigned long src_off,
5714                        unsigned long len)
5715 {
5716         char *dst_kaddr = page_address(dst_page);
5717         char *src_kaddr;
5718         int must_memmove = 0;
5719
5720         if (dst_page != src_page) {
5721                 src_kaddr = page_address(src_page);
5722         } else {
5723                 src_kaddr = dst_kaddr;
5724                 if (areas_overlap(src_off, dst_off, len))
5725                         must_memmove = 1;
5726         }
5727
5728         if (must_memmove)
5729                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5730         else
5731                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5732 }
5733
5734 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5735                            unsigned long src_offset, unsigned long len)
5736 {
5737         struct btrfs_fs_info *fs_info = dst->fs_info;
5738         size_t cur;
5739         size_t dst_off_in_page;
5740         size_t src_off_in_page;
5741         size_t start_offset = offset_in_page(dst->start);
5742         unsigned long dst_i;
5743         unsigned long src_i;
5744
5745         if (src_offset + len > dst->len) {
5746                 btrfs_err(fs_info,
5747                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5748                          src_offset, len, dst->len);
5749                 BUG_ON(1);
5750         }
5751         if (dst_offset + len > dst->len) {
5752                 btrfs_err(fs_info,
5753                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5754                          dst_offset, len, dst->len);
5755                 BUG_ON(1);
5756         }
5757
5758         while (len > 0) {
5759                 dst_off_in_page = offset_in_page(start_offset + dst_offset);
5760                 src_off_in_page = offset_in_page(start_offset + src_offset);
5761
5762                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5763                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5764
5765                 cur = min(len, (unsigned long)(PAGE_SIZE -
5766                                                src_off_in_page));
5767                 cur = min_t(unsigned long, cur,
5768                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5769
5770                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5771                            dst_off_in_page, src_off_in_page, cur);
5772
5773                 src_offset += cur;
5774                 dst_offset += cur;
5775                 len -= cur;
5776         }
5777 }
5778
5779 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5780                            unsigned long src_offset, unsigned long len)
5781 {
5782         struct btrfs_fs_info *fs_info = dst->fs_info;
5783         size_t cur;
5784         size_t dst_off_in_page;
5785         size_t src_off_in_page;
5786         unsigned long dst_end = dst_offset + len - 1;
5787         unsigned long src_end = src_offset + len - 1;
5788         size_t start_offset = offset_in_page(dst->start);
5789         unsigned long dst_i;
5790         unsigned long src_i;
5791
5792         if (src_offset + len > dst->len) {
5793                 btrfs_err(fs_info,
5794                           "memmove bogus src_offset %lu move len %lu len %lu",
5795                           src_offset, len, dst->len);
5796                 BUG_ON(1);
5797         }
5798         if (dst_offset + len > dst->len) {
5799                 btrfs_err(fs_info,
5800                           "memmove bogus dst_offset %lu move len %lu len %lu",
5801                           dst_offset, len, dst->len);
5802                 BUG_ON(1);
5803         }
5804         if (dst_offset < src_offset) {
5805                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5806                 return;
5807         }
5808         while (len > 0) {
5809                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5810                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5811
5812                 dst_off_in_page = offset_in_page(start_offset + dst_end);
5813                 src_off_in_page = offset_in_page(start_offset + src_end);
5814
5815                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5816                 cur = min(cur, dst_off_in_page + 1);
5817                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5818                            dst_off_in_page - cur + 1,
5819                            src_off_in_page - cur + 1, cur);
5820
5821                 dst_end -= cur;
5822                 src_end -= cur;
5823                 len -= cur;
5824         }
5825 }
5826
5827 int try_release_extent_buffer(struct page *page)
5828 {
5829         struct extent_buffer *eb;
5830
5831         /*
5832          * We need to make sure nobody is attaching this page to an eb right
5833          * now.
5834          */
5835         spin_lock(&page->mapping->private_lock);
5836         if (!PagePrivate(page)) {
5837                 spin_unlock(&page->mapping->private_lock);
5838                 return 1;
5839         }
5840
5841         eb = (struct extent_buffer *)page->private;
5842         BUG_ON(!eb);
5843
5844         /*
5845          * This is a little awful but should be ok, we need to make sure that
5846          * the eb doesn't disappear out from under us while we're looking at
5847          * this page.
5848          */
5849         spin_lock(&eb->refs_lock);
5850         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5851                 spin_unlock(&eb->refs_lock);
5852                 spin_unlock(&page->mapping->private_lock);
5853                 return 0;
5854         }
5855         spin_unlock(&page->mapping->private_lock);
5856
5857         /*
5858          * If tree ref isn't set then we know the ref on this eb is a real ref,
5859          * so just return, this page will likely be freed soon anyway.
5860          */
5861         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5862                 spin_unlock(&eb->refs_lock);
5863                 return 0;
5864         }
5865
5866         return release_extent_buffer(eb);
5867 }