fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18 #include "volumes.h"
  19 #include "check-integrity.h"
  20 #include "locking.h"
  21 #include "rcu-string.h"
  22 #include "backref.h"
  23 #include "transaction.h"
  24
  25 static struct kmem_cache *extent_state_cache;
  26 static struct kmem_cache *extent_buffer_cache;
  27 static struct bio_set *btrfs_bioset;
  28
  29 static inline bool extent_state_in_tree(const struct extent_state *state)
  30 {
  31         return !RB_EMPTY_NODE(&state->rb_node);
  32 }
  33
  34 #ifdef CONFIG_BTRFS_DEBUG
  35 static LIST_HEAD(buffers);
  36 static LIST_HEAD(states);
  37
  38 static DEFINE_SPINLOCK(leak_lock);
  39
  40 static inline
  41 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  42 {
  43         unsigned long flags;
  44
  45         spin_lock_irqsave(&leak_lock, flags);
  46         list_add(new, head);
  47         spin_unlock_irqrestore(&leak_lock, flags);
  48 }
  49
  50 static inline
  51 void btrfs_leak_debug_del(struct list_head *entry)
  52 {
  53         unsigned long flags;
  54
  55         spin_lock_irqsave(&leak_lock, flags);
  56         list_del(entry);
  57         spin_unlock_irqrestore(&leak_lock, flags);
  58 }
  59
  60 static inline
  61 void btrfs_leak_debug_check(void)
  62 {
  63         struct extent_state *state;
  64         struct extent_buffer *eb;
  65
  66         while (!list_empty(&states)) {
  67                 state = list_entry(states.next, struct extent_state, leak_list);
  68                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  69                        state->start, state->end, state->state,
  70                        extent_state_in_tree(state),
  71                        atomic_read(&state->refs));
  72                 list_del(&state->leak_list);
  73                 kmem_cache_free(extent_state_cache, state);
  74         }
  75
  76         while (!list_empty(&buffers)) {
  77                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  78                 pr_err("BTRFS: buffer leak start %llu len %lu refs %d\n",
  79                        eb->start, eb->len, atomic_read(&eb->refs));
  80                 list_del(&eb->leak_list);
  81                 kmem_cache_free(extent_buffer_cache, eb);
  82         }
  83 }
  84
  85 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  86         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  87 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  88                 struct extent_io_tree *tree, u64 start, u64 end)
  89 {
  90         struct inode *inode;
  91         u64 isize;
  92
  93         if (!tree->mapping)
  94                 return;
  95
  96         inode = tree->mapping->host;
  97         isize = i_size_read(inode);
  98         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  99                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
 100                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
 101                                 caller, btrfs_ino(inode), isize, start, end);
 102         }
 103 }
 104 #else
 105 #define btrfs_leak_debug_add(new, head) do {} while (0)
 106 #define btrfs_leak_debug_del(entry)     do {} while (0)
 107 #define btrfs_leak_debug_check()        do {} while (0)
 108 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 109 #endif
 110
 111 #define BUFFER_LRU_MAX 64
 112
 113 struct tree_entry {
 114         u64 start;
 115         u64 end;
 116         struct rb_node rb_node;
 117 };
 118
 119 struct extent_page_data {
 120         struct bio *bio;
 121         struct extent_io_tree *tree;
 122         get_extent_t *get_extent;
 123         unsigned long bio_flags;
 124
 125         /* tells writepage not to lock the state bits for this range
 126          * it still does the unlocking
 127          */
 128         unsigned int extent_locked:1;
 129
 130         /* tells the submit_bio code to use a WRITE_SYNC */
 131         unsigned int sync_io:1;
 132 };
 133
 134 static void add_extent_changeset(struct extent_state *state, unsigned bits,
 135                                  struct extent_changeset *changeset,
 136                                  int set)
 137 {
 138         int ret;
 139
 140         if (!changeset)
 141                 return;
 142         if (set && (state->state & bits) == bits)
 143                 return;
 144         if (!set && (state->state & bits) == 0)
 145                 return;
 146         changeset->bytes_changed += state->end - state->start + 1;
 147         ret = ulist_add(changeset->range_changed, state->start, state->end,
 148                         GFP_ATOMIC);
 149         /* ENOMEM */
 150         BUG_ON(ret < 0);
 151 }
 152
 153 static noinline void flush_write_bio(void *data);
 154 static inline struct btrfs_fs_info *
 155 tree_fs_info(struct extent_io_tree *tree)
 156 {
 157         if (!tree->mapping)
 158                 return NULL;
 159         return btrfs_sb(tree->mapping->host->i_sb);
 160 }
 161
 162 int __init extent_io_init(void)
 163 {
 164         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 165                         sizeof(struct extent_state), 0,
 166                         SLAB_MEM_SPREAD, NULL);
 167         if (!extent_state_cache)
 168                 return -ENOMEM;
 169
 170         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 171                         sizeof(struct extent_buffer), 0,
 172                         SLAB_MEM_SPREAD, NULL);
 173         if (!extent_buffer_cache)
 174                 goto free_state_cache;
 175
 176         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 177                                      offsetof(struct btrfs_io_bio, bio));
 178         if (!btrfs_bioset)
 179                 goto free_buffer_cache;
 180
 181         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 182                 goto free_bioset;
 183
 184         return 0;
 185
 186 free_bioset:
 187         bioset_free(btrfs_bioset);
 188         btrfs_bioset = NULL;
 189
 190 free_buffer_cache:
 191         kmem_cache_destroy(extent_buffer_cache);
 192         extent_buffer_cache = NULL;
 193
 194 free_state_cache:
 195         kmem_cache_destroy(extent_state_cache);
 196         extent_state_cache = NULL;
 197         return -ENOMEM;
 198 }
 199
 200 void extent_io_exit(void)
 201 {
 202         btrfs_leak_debug_check();
 203
 204         /*
 205          * Make sure all delayed rcu free are flushed before we
 206          * destroy caches.
 207          */
 208         rcu_barrier();
 209         kmem_cache_destroy(extent_state_cache);
 210         kmem_cache_destroy(extent_buffer_cache);
 211         if (btrfs_bioset)
 212                 bioset_free(btrfs_bioset);
 213 }
 214
 215 void extent_io_tree_init(struct extent_io_tree *tree,
 216                          struct address_space *mapping)
 217 {
 218         tree->state = RB_ROOT;
 219         tree->ops = NULL;
 220         tree->dirty_bytes = 0;
 221         spin_lock_init(&tree->lock);
 222         tree->mapping = mapping;
 223 }
 224
 225 static struct extent_state *alloc_extent_state(gfp_t mask)
 226 {
 227         struct extent_state *state;
 228
 229         state = kmem_cache_alloc(extent_state_cache, mask);
 230         if (!state)
 231                 return state;
 232         state->state = 0;
 233         state->failrec = NULL;
 234         RB_CLEAR_NODE(&state->rb_node);
 235         btrfs_leak_debug_add(&state->leak_list, &states);
 236         atomic_set(&state->refs, 1);
 237         init_waitqueue_head(&state->wq);
 238         trace_alloc_extent_state(state, mask, _RET_IP_);
 239         return state;
 240 }
 241
 242 void free_extent_state(struct extent_state *state)
 243 {
 244         if (!state)
 245                 return;
 246         if (atomic_dec_and_test(&state->refs)) {
 247                 WARN_ON(extent_state_in_tree(state));
 248                 btrfs_leak_debug_del(&state->leak_list);
 249                 trace_free_extent_state(state, _RET_IP_);
 250                 kmem_cache_free(extent_state_cache, state);
 251         }
 252 }
 253
 254 static struct rb_node *tree_insert(struct rb_root *root,
 255                                    struct rb_node *search_start,
 256                                    u64 offset,
 257                                    struct rb_node *node,
 258                                    struct rb_node ***p_in,
 259                                    struct rb_node **parent_in)
 260 {
 261         struct rb_node **p;
 262         struct rb_node *parent = NULL;
 263         struct tree_entry *entry;
 264
 265         if (p_in && parent_in) {
 266                 p = *p_in;
 267                 parent = *parent_in;
 268                 goto do_insert;
 269         }
 270
 271         p = search_start ? &search_start : &root->rb_node;
 272         while (*p) {
 273                 parent = *p;
 274                 entry = rb_entry(parent, struct tree_entry, rb_node);
 275
 276                 if (offset < entry->start)
 277                         p = &(*p)->rb_left;
 278                 else if (offset > entry->end)
 279                         p = &(*p)->rb_right;
 280                 else
 281                         return parent;
 282         }
 283
 284 do_insert:
 285         rb_link_node(node, parent, p);
 286         rb_insert_color(node, root);
 287         return NULL;
 288 }
 289
 290 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 291                                       struct rb_node **prev_ret,
 292                                       struct rb_node **next_ret,
 293                                       struct rb_node ***p_ret,
 294                                       struct rb_node **parent_ret)
 295 {
 296         struct rb_root *root = &tree->state;
 297         struct rb_node **n = &root->rb_node;
 298         struct rb_node *prev = NULL;
 299         struct rb_node *orig_prev = NULL;
 300         struct tree_entry *entry;
 301         struct tree_entry *prev_entry = NULL;
 302
 303         while (*n) {
 304                 prev = *n;
 305                 entry = rb_entry(prev, struct tree_entry, rb_node);
 306                 prev_entry = entry;
 307
 308                 if (offset < entry->start)
 309                         n = &(*n)->rb_left;
 310                 else if (offset > entry->end)
 311                         n = &(*n)->rb_right;
 312                 else
 313                         return *n;
 314         }
 315
 316         if (p_ret)
 317                 *p_ret = n;
 318         if (parent_ret)
 319                 *parent_ret = prev;
 320
 321         if (prev_ret) {
 322                 orig_prev = prev;
 323                 while (prev && offset > prev_entry->end) {
 324                         prev = rb_next(prev);
 325                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 326                 }
 327                 *prev_ret = prev;
 328                 prev = orig_prev;
 329         }
 330
 331         if (next_ret) {
 332                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 333                 while (prev && offset < prev_entry->start) {
 334                         prev = rb_prev(prev);
 335                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 336                 }
 337                 *next_ret = prev;
 338         }
 339         return NULL;
 340 }
 341
 342 static inline struct rb_node *
 343 tree_search_for_insert(struct extent_io_tree *tree,
 344                        u64 offset,
 345                        struct rb_node ***p_ret,
 346                        struct rb_node **parent_ret)
 347 {
 348         struct rb_node *prev = NULL;
 349         struct rb_node *ret;
 350
 351         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 352         if (!ret)
 353                 return prev;
 354         return ret;
 355 }
 356
 357 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 358                                           u64 offset)
 359 {
 360         return tree_search_for_insert(tree, offset, NULL, NULL);
 361 }
 362
 363 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 364                      struct extent_state *other)
 365 {
 366         if (tree->ops && tree->ops->merge_extent_hook)
 367                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 368                                              other);
 369 }
 370
 371 /*
 372  * utility function to look for merge candidates inside a given range.
 373  * Any extents with matching state are merged together into a single
 374  * extent in the tree.  Extents with EXTENT_IO in their state field
 375  * are not merged because the end_io handlers need to be able to do
 376  * operations on them without sleeping (or doing allocations/splits).
 377  *
 378  * This should be called with the tree lock held.
 379  */
 380 static void merge_state(struct extent_io_tree *tree,
 381                         struct extent_state *state)
 382 {
 383         struct extent_state *other;
 384         struct rb_node *other_node;
 385
 386         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 387                 return;
 388
 389         other_node = rb_prev(&state->rb_node);
 390         if (other_node) {
 391                 other = rb_entry(other_node, struct extent_state, rb_node);
 392                 if (other->end == state->start - 1 &&
 393                     other->state == state->state) {
 394                         merge_cb(tree, state, other);
 395                         state->start = other->start;
 396                         rb_erase(&other->rb_node, &tree->state);
 397                         RB_CLEAR_NODE(&other->rb_node);
 398                         free_extent_state(other);
 399                 }
 400         }
 401         other_node = rb_next(&state->rb_node);
 402         if (other_node) {
 403                 other = rb_entry(other_node, struct extent_state, rb_node);
 404                 if (other->start == state->end + 1 &&
 405                     other->state == state->state) {
 406                         merge_cb(tree, state, other);
 407                         state->end = other->end;
 408                         rb_erase(&other->rb_node, &tree->state);
 409                         RB_CLEAR_NODE(&other->rb_node);
 410                         free_extent_state(other);
 411                 }
 412         }
 413 }
 414
 415 static void set_state_cb(struct extent_io_tree *tree,
 416                          struct extent_state *state, unsigned *bits)
 417 {
 418         if (tree->ops && tree->ops->set_bit_hook)
 419                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 420 }
 421
 422 static void clear_state_cb(struct extent_io_tree *tree,
 423                            struct extent_state *state, unsigned *bits)
 424 {
 425         if (tree->ops && tree->ops->clear_bit_hook)
 426                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 427 }
 428
 429 static void set_state_bits(struct extent_io_tree *tree,
 430                            struct extent_state *state, unsigned *bits,
 431                            struct extent_changeset *changeset);
 432
 433 /*
 434  * insert an extent_state struct into the tree.  'bits' are set on the
 435  * struct before it is inserted.
 436  *
 437  * This may return -EEXIST if the extent is already there, in which case the
 438  * state struct is freed.
 439  *
 440  * The tree lock is not taken internally.  This is a utility function and
 441  * probably isn't what you want to call (see set/clear_extent_bit).
 442  */
 443 static int insert_state(struct extent_io_tree *tree,
 444                         struct extent_state *state, u64 start, u64 end,
 445                         struct rb_node ***p,
 446                         struct rb_node **parent,
 447                         unsigned *bits, struct extent_changeset *changeset)
 448 {
 449         struct rb_node *node;
 450
 451         if (end < start)
 452                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 453                        end, start);
 454         state->start = start;
 455         state->end = end;
 456
 457         set_state_bits(tree, state, bits, changeset);
 458
 459         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 460         if (node) {
 461                 struct extent_state *found;
 462                 found = rb_entry(node, struct extent_state, rb_node);
 463                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
 464                        found->start, found->end, start, end);
 465                 return -EEXIST;
 466         }
 467         merge_state(tree, state);
 468         return 0;
 469 }
 470
 471 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 472                      u64 split)
 473 {
 474         if (tree->ops && tree->ops->split_extent_hook)
 475                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 476 }
 477
 478 /*
 479  * split a given extent state struct in two, inserting the preallocated
 480  * struct 'prealloc' as the newly created second half.  'split' indicates an
 481  * offset inside 'orig' where it should be split.
 482  *
 483  * Before calling,
 484  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 485  * are two extent state structs in the tree:
 486  * prealloc: [orig->start, split - 1]
 487  * orig: [ split, orig->end ]
 488  *
 489  * The tree locks are not taken by this function. They need to be held
 490  * by the caller.
 491  */
 492 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 493                        struct extent_state *prealloc, u64 split)
 494 {
 495         struct rb_node *node;
 496
 497         split_cb(tree, orig, split);
 498
 499         prealloc->start = orig->start;
 500         prealloc->end = split - 1;
 501         prealloc->state = orig->state;
 502         orig->start = split;
 503
 504         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 505                            &prealloc->rb_node, NULL, NULL);
 506         if (node) {
 507                 free_extent_state(prealloc);
 508                 return -EEXIST;
 509         }
 510         return 0;
 511 }
 512
 513 static struct extent_state *next_state(struct extent_state *state)
 514 {
 515         struct rb_node *next = rb_next(&state->rb_node);
 516         if (next)
 517                 return rb_entry(next, struct extent_state, rb_node);
 518         else
 519                 return NULL;
 520 }
 521
 522 /*
 523  * utility function to clear some bits in an extent state struct.
 524  * it will optionally wake up any one waiting on this state (wake == 1).
 525  *
 526  * If no bits are set on the state struct after clearing things, the
 527  * struct is freed and removed from the tree
 528  */
 529 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 530                                             struct extent_state *state,
 531                                             unsigned *bits, int wake,
 532                                             struct extent_changeset *changeset)
 533 {
 534         struct extent_state *next;
 535         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 536
 537         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 538                 u64 range = state->end - state->start + 1;
 539                 WARN_ON(range > tree->dirty_bytes);
 540                 tree->dirty_bytes -= range;
 541         }
 542         clear_state_cb(tree, state, bits);
 543         add_extent_changeset(state, bits_to_clear, changeset, 0);
 544         state->state &= ~bits_to_clear;
 545         if (wake)
 546                 wake_up(&state->wq);
 547         if (state->state == 0) {
 548                 next = next_state(state);
 549                 if (extent_state_in_tree(state)) {
 550                         rb_erase(&state->rb_node, &tree->state);
 551                         RB_CLEAR_NODE(&state->rb_node);
 552                         free_extent_state(state);
 553                 } else {
 554                         WARN_ON(1);
 555                 }
 556         } else {
 557                 merge_state(tree, state);
 558                 next = next_state(state);
 559         }
 560         return next;
 561 }
 562
 563 static struct extent_state *
 564 alloc_extent_state_atomic(struct extent_state *prealloc)
 565 {
 566         if (!prealloc)
 567                 prealloc = alloc_extent_state(GFP_ATOMIC);
 568
 569         return prealloc;
 570 }
 571
 572 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 573 {
 574         btrfs_panic(tree_fs_info(tree), err,
 575                     "Locking error: Extent tree was modified by another thread while locked.");
 576 }
 577
 578 /*
 579  * clear some bits on a range in the tree.  This may require splitting
 580  * or inserting elements in the tree, so the gfp mask is used to
 581  * indicate which allocations or sleeping are allowed.
 582  *
 583  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 584  * the given range from the tree regardless of state (ie for truncate).
 585  *
 586  * the range [start, end] is inclusive.
 587  *
 588  * This takes the tree lock, and returns 0 on success and < 0 on error.
 589  */
 590 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 591                               unsigned bits, int wake, int delete,
 592                               struct extent_state **cached_state,
 593                               gfp_t mask, struct extent_changeset *changeset)
 594 {
 595         struct extent_state *state;
 596         struct extent_state *cached;
 597         struct extent_state *prealloc = NULL;
 598         struct rb_node *node;
 599         u64 last_end;
 600         int err;
 601         int clear = 0;
 602
 603         btrfs_debug_check_extent_io_range(tree, start, end);
 604
 605         if (bits & EXTENT_DELALLOC)
 606                 bits |= EXTENT_NORESERVE;
 607
 608         if (delete)
 609                 bits |= ~EXTENT_CTLBITS;
 610         bits |= EXTENT_FIRST_DELALLOC;
 611
 612         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 613                 clear = 1;
 614 again:
 615         if (!prealloc && gfpflags_allow_blocking(mask)) {
 616                 /*
 617                  * Don't care for allocation failure here because we might end
 618                  * up not needing the pre-allocated extent state at all, which
 619                  * is the case if we only have in the tree extent states that
 620                  * cover our input range and don't cover too any other range.
 621                  * If we end up needing a new extent state we allocate it later.
 622                  */
 623                 prealloc = alloc_extent_state(mask);
 624         }
 625
 626         spin_lock(&tree->lock);
 627         if (cached_state) {
 628                 cached = *cached_state;
 629
 630                 if (clear) {
 631                         *cached_state = NULL;
 632                         cached_state = NULL;
 633                 }
 634
 635                 if (cached && extent_state_in_tree(cached) &&
 636                     cached->start <= start && cached->end > start) {
 637                         if (clear)
 638                                 atomic_dec(&cached->refs);
 639                         state = cached;
 640                         goto hit_next;
 641                 }
 642                 if (clear)
 643                         free_extent_state(cached);
 644         }
 645         /*
 646          * this search will find the extents that end after
 647          * our range starts
 648          */
 649         node = tree_search(tree, start);
 650         if (!node)
 651                 goto out;
 652         state = rb_entry(node, struct extent_state, rb_node);
 653 hit_next:
 654         if (state->start > end)
 655                 goto out;
 656         WARN_ON(state->end < start);
 657         last_end = state->end;
 658
 659         /* the state doesn't have the wanted bits, go ahead */
 660         if (!(state->state & bits)) {
 661                 state = next_state(state);
 662                 goto next;
 663         }
 664
 665         /*
 666          *     | ---- desired range ---- |
 667          *  | state | or
 668          *  | ------------- state -------------- |
 669          *
 670          * We need to split the extent we found, and may flip
 671          * bits on second half.
 672          *
 673          * If the extent we found extends past our range, we
 674          * just split and search again.  It'll get split again
 675          * the next time though.
 676          *
 677          * If the extent we found is inside our range, we clear
 678          * the desired bit on it.
 679          */
 680
 681         if (state->start < start) {
 682                 prealloc = alloc_extent_state_atomic(prealloc);
 683                 BUG_ON(!prealloc);
 684                 err = split_state(tree, state, prealloc, start);
 685                 if (err)
 686                         extent_io_tree_panic(tree, err);
 687
 688                 prealloc = NULL;
 689                 if (err)
 690                         goto out;
 691                 if (state->end <= end) {
 692                         state = clear_state_bit(tree, state, &bits, wake,
 693                                                 changeset);
 694                         goto next;
 695                 }
 696                 goto search_again;
 697         }
 698         /*
 699          * | ---- desired range ---- |
 700          *                        | state |
 701          * We need to split the extent, and clear the bit
 702          * on the first half
 703          */
 704         if (state->start <= end && state->end > end) {
 705                 prealloc = alloc_extent_state_atomic(prealloc);
 706                 BUG_ON(!prealloc);
 707                 err = split_state(tree, state, prealloc, end + 1);
 708                 if (err)
 709                         extent_io_tree_panic(tree, err);
 710
 711                 if (wake)
 712                         wake_up(&state->wq);
 713
 714                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 715
 716                 prealloc = NULL;
 717                 goto out;
 718         }
 719
 720         state = clear_state_bit(tree, state, &bits, wake, changeset);
 721 next:
 722         if (last_end == (u64)-1)
 723                 goto out;
 724         start = last_end + 1;
 725         if (start <= end && state && !need_resched())
 726                 goto hit_next;
 727
 728 search_again:
 729         if (start > end)
 730                 goto out;
 731         spin_unlock(&tree->lock);
 732         if (gfpflags_allow_blocking(mask))
 733                 cond_resched();
 734         goto again;
 735
 736 out:
 737         spin_unlock(&tree->lock);
 738         if (prealloc)
 739                 free_extent_state(prealloc);
 740
 741         return 0;
 742
 743 }
 744
 745 static void wait_on_state(struct extent_io_tree *tree,
 746                           struct extent_state *state)
 747                 __releases(tree->lock)
 748                 __acquires(tree->lock)
 749 {
 750         DEFINE_WAIT(wait);
 751         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 752         spin_unlock(&tree->lock);
 753         schedule();
 754         spin_lock(&tree->lock);
 755         finish_wait(&state->wq, &wait);
 756 }
 757
 758 /*
 759  * waits for one or more bits to clear on a range in the state tree.
 760  * The range [start, end] is inclusive.
 761  * The tree lock is taken by this function
 762  */
 763 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 764                             unsigned long bits)
 765 {
 766         struct extent_state *state;
 767         struct rb_node *node;
 768
 769         btrfs_debug_check_extent_io_range(tree, start, end);
 770
 771         spin_lock(&tree->lock);
 772 again:
 773         while (1) {
 774                 /*
 775                  * this search will find all the extents that end after
 776                  * our range starts
 777                  */
 778                 node = tree_search(tree, start);
 779 process_node:
 780                 if (!node)
 781                         break;
 782
 783                 state = rb_entry(node, struct extent_state, rb_node);
 784
 785                 if (state->start > end)
 786                         goto out;
 787
 788                 if (state->state & bits) {
 789                         start = state->start;
 790                         atomic_inc(&state->refs);
 791                         wait_on_state(tree, state);
 792                         free_extent_state(state);
 793                         goto again;
 794                 }
 795                 start = state->end + 1;
 796
 797                 if (start > end)
 798                         break;
 799
 800                 if (!cond_resched_lock(&tree->lock)) {
 801                         node = rb_next(node);
 802                         goto process_node;
 803                 }
 804         }
 805 out:
 806         spin_unlock(&tree->lock);
 807 }
 808
 809 static void set_state_bits(struct extent_io_tree *tree,
 810                            struct extent_state *state,
 811                            unsigned *bits, struct extent_changeset *changeset)
 812 {
 813         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 814
 815         set_state_cb(tree, state, bits);
 816         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 817                 u64 range = state->end - state->start + 1;
 818                 tree->dirty_bytes += range;
 819         }
 820         add_extent_changeset(state, bits_to_set, changeset, 1);
 821         state->state |= bits_to_set;
 822 }
 823
 824 static void cache_state_if_flags(struct extent_state *state,
 825                                  struct extent_state **cached_ptr,
 826                                  unsigned flags)
 827 {
 828         if (cached_ptr && !(*cached_ptr)) {
 829                 if (!flags || (state->state & flags)) {
 830                         *cached_ptr = state;
 831                         atomic_inc(&state->refs);
 832                 }
 833         }
 834 }
 835
 836 static void cache_state(struct extent_state *state,
 837                         struct extent_state **cached_ptr)
 838 {
 839         return cache_state_if_flags(state, cached_ptr,
 840                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 841 }
 842
 843 /*
 844  * set some bits on a range in the tree.  This may require allocations or
 845  * sleeping, so the gfp mask is used to indicate what is allowed.
 846  *
 847  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 848  * part of the range already has the desired bits set.  The start of the
 849  * existing range is returned in failed_start in this case.
 850  *
 851  * [start, end] is inclusive This takes the tree lock.
 852  */
 853
 854 static int __must_check
 855 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 856                  unsigned bits, unsigned exclusive_bits,
 857                  u64 *failed_start, struct extent_state **cached_state,
 858                  gfp_t mask, struct extent_changeset *changeset)
 859 {
 860         struct extent_state *state;
 861         struct extent_state *prealloc = NULL;
 862         struct rb_node *node;
 863         struct rb_node **p;
 864         struct rb_node *parent;
 865         int err = 0;
 866         u64 last_start;
 867         u64 last_end;
 868
 869         btrfs_debug_check_extent_io_range(tree, start, end);
 870
 871         bits |= EXTENT_FIRST_DELALLOC;
 872 again:
 873         if (!prealloc && gfpflags_allow_blocking(mask)) {
 874                 /*
 875                  * Don't care for allocation failure here because we might end
 876                  * up not needing the pre-allocated extent state at all, which
 877                  * is the case if we only have in the tree extent states that
 878                  * cover our input range and don't cover too any other range.
 879                  * If we end up needing a new extent state we allocate it later.
 880                  */
 881                 prealloc = alloc_extent_state(mask);
 882         }
 883
 884         spin_lock(&tree->lock);
 885         if (cached_state && *cached_state) {
 886                 state = *cached_state;
 887                 if (state->start <= start && state->end > start &&
 888                     extent_state_in_tree(state)) {
 889                         node = &state->rb_node;
 890                         goto hit_next;
 891                 }
 892         }
 893         /*
 894          * this search will find all the extents that end after
 895          * our range starts.
 896          */
 897         node = tree_search_for_insert(tree, start, &p, &parent);
 898         if (!node) {
 899                 prealloc = alloc_extent_state_atomic(prealloc);
 900                 BUG_ON(!prealloc);
 901                 err = insert_state(tree, prealloc, start, end,
 902                                    &p, &parent, &bits, changeset);
 903                 if (err)
 904                         extent_io_tree_panic(tree, err);
 905
 906                 cache_state(prealloc, cached_state);
 907                 prealloc = NULL;
 908                 goto out;
 909         }
 910         state = rb_entry(node, struct extent_state, rb_node);
 911 hit_next:
 912         last_start = state->start;
 913         last_end = state->end;
 914
 915         /*
 916          * | ---- desired range ---- |
 917          * | state |
 918          *
 919          * Just lock what we found and keep going
 920          */
 921         if (state->start == start && state->end <= end) {
 922                 if (state->state & exclusive_bits) {
 923                         *failed_start = state->start;
 924                         err = -EEXIST;
 925                         goto out;
 926                 }
 927
 928                 set_state_bits(tree, state, &bits, changeset);
 929                 cache_state(state, cached_state);
 930                 merge_state(tree, state);
 931                 if (last_end == (u64)-1)
 932                         goto out;
 933                 start = last_end + 1;
 934                 state = next_state(state);
 935                 if (start < end && state && state->start == start &&
 936                     !need_resched())
 937                         goto hit_next;
 938                 goto search_again;
 939         }
 940
 941         /*
 942          *     | ---- desired range ---- |
 943          * | state |
 944          *   or
 945          * | ------------- state -------------- |
 946          *
 947          * We need to split the extent we found, and may flip bits on
 948          * second half.
 949          *
 950          * If the extent we found extends past our
 951          * range, we just split and search again.  It'll get split
 952          * again the next time though.
 953          *
 954          * If the extent we found is inside our range, we set the
 955          * desired bit on it.
 956          */
 957         if (state->start < start) {
 958                 if (state->state & exclusive_bits) {
 959                         *failed_start = start;
 960                         err = -EEXIST;
 961                         goto out;
 962                 }
 963
 964                 prealloc = alloc_extent_state_atomic(prealloc);
 965                 BUG_ON(!prealloc);
 966                 err = split_state(tree, state, prealloc, start);
 967                 if (err)
 968                         extent_io_tree_panic(tree, err);
 969
 970                 prealloc = NULL;
 971                 if (err)
 972                         goto out;
 973                 if (state->end <= end) {
 974                         set_state_bits(tree, state, &bits, changeset);
 975                         cache_state(state, cached_state);
 976                         merge_state(tree, state);
 977                         if (last_end == (u64)-1)
 978                                 goto out;
 979                         start = last_end + 1;
 980                         state = next_state(state);
 981                         if (start < end && state && state->start == start &&
 982                             !need_resched())
 983                                 goto hit_next;
 984                 }
 985                 goto search_again;
 986         }
 987         /*
 988          * | ---- desired range ---- |
 989          *     | state | or               | state |
 990          *
 991          * There's a hole, we need to insert something in it and
 992          * ignore the extent we found.
 993          */
 994         if (state->start > start) {
 995                 u64 this_end;
 996                 if (end < last_start)
 997                         this_end = end;
 998                 else
 999                         this_end = last_start - 1;
1000
1001                 prealloc = alloc_extent_state_atomic(prealloc);
1002                 BUG_ON(!prealloc);
1003
1004                 /*
1005                  * Avoid to free 'prealloc' if it can be merged with
1006                  * the later extent.
1007                  */
1008                 err = insert_state(tree, prealloc, start, this_end,
1009                                    NULL, NULL, &bits, changeset);
1010                 if (err)
1011                         extent_io_tree_panic(tree, err);
1012
1013                 cache_state(prealloc, cached_state);
1014                 prealloc = NULL;
1015                 start = this_end + 1;
1016                 goto search_again;
1017         }
1018         /*
1019          * | ---- desired range ---- |
1020          *                        | state |
1021          * We need to split the extent, and set the bit
1022          * on the first half
1023          */
1024         if (state->start <= end && state->end > end) {
1025                 if (state->state & exclusive_bits) {
1026                         *failed_start = start;
1027                         err = -EEXIST;
1028                         goto out;
1029                 }
1030
1031                 prealloc = alloc_extent_state_atomic(prealloc);
1032                 BUG_ON(!prealloc);
1033                 err = split_state(tree, state, prealloc, end + 1);
1034                 if (err)
1035                         extent_io_tree_panic(tree, err);
1036
1037                 set_state_bits(tree, prealloc, &bits, changeset);
1038                 cache_state(prealloc, cached_state);
1039                 merge_state(tree, prealloc);
1040                 prealloc = NULL;
1041                 goto out;
1042         }
1043
1044 search_again:
1045         if (start > end)
1046                 goto out;
1047         spin_unlock(&tree->lock);
1048         if (gfpflags_allow_blocking(mask))
1049                 cond_resched();
1050         goto again;
1051
1052 out:
1053         spin_unlock(&tree->lock);
1054         if (prealloc)
1055                 free_extent_state(prealloc);
1056
1057         return err;
1058
1059 }
1060
1061 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1062                    unsigned bits, u64 * failed_start,
1063                    struct extent_state **cached_state, gfp_t mask)
1064 {
1065         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1066                                 cached_state, mask, NULL);
1067 }
1068
1069
1070 /**
1071  * convert_extent_bit - convert all bits in a given range from one bit to
1072  *                      another
1073  * @tree:       the io tree to search
1074  * @start:      the start offset in bytes
1075  * @end:        the end offset in bytes (inclusive)
1076  * @bits:       the bits to set in this range
1077  * @clear_bits: the bits to clear in this range
1078  * @cached_state:       state that we're going to cache
1079  *
1080  * This will go through and set bits for the given range.  If any states exist
1081  * already in this range they are set with the given bit and cleared of the
1082  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1083  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1084  * boundary bits like LOCK.
1085  *
1086  * All allocations are done with GFP_NOFS.
1087  */
1088 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1089                        unsigned bits, unsigned clear_bits,
1090                        struct extent_state **cached_state)
1091 {
1092         struct extent_state *state;
1093         struct extent_state *prealloc = NULL;
1094         struct rb_node *node;
1095         struct rb_node **p;
1096         struct rb_node *parent;
1097         int err = 0;
1098         u64 last_start;
1099         u64 last_end;
1100         bool first_iteration = true;
1101
1102         btrfs_debug_check_extent_io_range(tree, start, end);
1103
1104 again:
1105         if (!prealloc) {
1106                 /*
1107                  * Best effort, don't worry if extent state allocation fails
1108                  * here for the first iteration. We might have a cached state
1109                  * that matches exactly the target range, in which case no
1110                  * extent state allocations are needed. We'll only know this
1111                  * after locking the tree.
1112                  */
1113                 prealloc = alloc_extent_state(GFP_NOFS);
1114                 if (!prealloc && !first_iteration)
1115                         return -ENOMEM;
1116         }
1117
1118         spin_lock(&tree->lock);
1119         if (cached_state && *cached_state) {
1120                 state = *cached_state;
1121                 if (state->start <= start && state->end > start &&
1122                     extent_state_in_tree(state)) {
1123                         node = &state->rb_node;
1124                         goto hit_next;
1125                 }
1126         }
1127
1128         /*
1129          * this search will find all the extents that end after
1130          * our range starts.
1131          */
1132         node = tree_search_for_insert(tree, start, &p, &parent);
1133         if (!node) {
1134                 prealloc = alloc_extent_state_atomic(prealloc);
1135                 if (!prealloc) {
1136                         err = -ENOMEM;
1137                         goto out;
1138                 }
1139                 err = insert_state(tree, prealloc, start, end,
1140                                    &p, &parent, &bits, NULL);
1141                 if (err)
1142                         extent_io_tree_panic(tree, err);
1143                 cache_state(prealloc, cached_state);
1144                 prealloc = NULL;
1145                 goto out;
1146         }
1147         state = rb_entry(node, struct extent_state, rb_node);
1148 hit_next:
1149         last_start = state->start;
1150         last_end = state->end;
1151
1152         /*
1153          * | ---- desired range ---- |
1154          * | state |
1155          *
1156          * Just lock what we found and keep going
1157          */
1158         if (state->start == start && state->end <= end) {
1159                 set_state_bits(tree, state, &bits, NULL);
1160                 cache_state(state, cached_state);
1161                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1162                 if (last_end == (u64)-1)
1163                         goto out;
1164                 start = last_end + 1;
1165                 if (start < end && state && state->start == start &&
1166                     !need_resched())
1167                         goto hit_next;
1168                 goto search_again;
1169         }
1170
1171         /*
1172          *     | ---- desired range ---- |
1173          * | state |
1174          *   or
1175          * | ------------- state -------------- |
1176          *
1177          * We need to split the extent we found, and may flip bits on
1178          * second half.
1179          *
1180          * If the extent we found extends past our
1181          * range, we just split and search again.  It'll get split
1182          * again the next time though.
1183          *
1184          * If the extent we found is inside our range, we set the
1185          * desired bit on it.
1186          */
1187         if (state->start < start) {
1188                 prealloc = alloc_extent_state_atomic(prealloc);
1189                 if (!prealloc) {
1190                         err = -ENOMEM;
1191                         goto out;
1192                 }
1193                 err = split_state(tree, state, prealloc, start);
1194                 if (err)
1195                         extent_io_tree_panic(tree, err);
1196                 prealloc = NULL;
1197                 if (err)
1198                         goto out;
1199                 if (state->end <= end) {
1200                         set_state_bits(tree, state, &bits, NULL);
1201                         cache_state(state, cached_state);
1202                         state = clear_state_bit(tree, state, &clear_bits, 0,
1203                                                 NULL);
1204                         if (last_end == (u64)-1)
1205                                 goto out;
1206                         start = last_end + 1;
1207                         if (start < end && state && state->start == start &&
1208                             !need_resched())
1209                                 goto hit_next;
1210                 }
1211                 goto search_again;
1212         }
1213         /*
1214          * | ---- desired range ---- |
1215          *     | state | or               | state |
1216          *
1217          * There's a hole, we need to insert something in it and
1218          * ignore the extent we found.
1219          */
1220         if (state->start > start) {
1221                 u64 this_end;
1222                 if (end < last_start)
1223                         this_end = end;
1224                 else
1225                         this_end = last_start - 1;
1226
1227                 prealloc = alloc_extent_state_atomic(prealloc);
1228                 if (!prealloc) {
1229                         err = -ENOMEM;
1230                         goto out;
1231                 }
1232
1233                 /*
1234                  * Avoid to free 'prealloc' if it can be merged with
1235                  * the later extent.
1236                  */
1237                 err = insert_state(tree, prealloc, start, this_end,
1238                                    NULL, NULL, &bits, NULL);
1239                 if (err)
1240                         extent_io_tree_panic(tree, err);
1241                 cache_state(prealloc, cached_state);
1242                 prealloc = NULL;
1243                 start = this_end + 1;
1244                 goto search_again;
1245         }
1246         /*
1247          * | ---- desired range ---- |
1248          *                        | state |
1249          * We need to split the extent, and set the bit
1250          * on the first half
1251          */
1252         if (state->start <= end && state->end > end) {
1253                 prealloc = alloc_extent_state_atomic(prealloc);
1254                 if (!prealloc) {
1255                         err = -ENOMEM;
1256                         goto out;
1257                 }
1258
1259                 err = split_state(tree, state, prealloc, end + 1);
1260                 if (err)
1261                         extent_io_tree_panic(tree, err);
1262
1263                 set_state_bits(tree, prealloc, &bits, NULL);
1264                 cache_state(prealloc, cached_state);
1265                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1266                 prealloc = NULL;
1267                 goto out;
1268         }
1269
1270 search_again:
1271         if (start > end)
1272                 goto out;
1273         spin_unlock(&tree->lock);
1274         cond_resched();
1275         first_iteration = false;
1276         goto again;
1277
1278 out:
1279         spin_unlock(&tree->lock);
1280         if (prealloc)
1281                 free_extent_state(prealloc);
1282
1283         return err;
1284 }
1285
1286 /* wrappers around set/clear extent bit */
1287 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1288                            unsigned bits, struct extent_changeset *changeset)
1289 {
1290         /*
1291          * We don't support EXTENT_LOCKED yet, as current changeset will
1292          * record any bits changed, so for EXTENT_LOCKED case, it will
1293          * either fail with -EEXIST or changeset will record the whole
1294          * range.
1295          */
1296         BUG_ON(bits & EXTENT_LOCKED);
1297
1298         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1299                                 changeset);
1300 }
1301
1302 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1303                      unsigned bits, int wake, int delete,
1304                      struct extent_state **cached, gfp_t mask)
1305 {
1306         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1307                                   cached, mask, NULL);
1308 }
1309
1310 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1311                 unsigned bits, struct extent_changeset *changeset)
1312 {
1313         /*
1314          * Don't support EXTENT_LOCKED case, same reason as
1315          * set_record_extent_bits().
1316          */
1317         BUG_ON(bits & EXTENT_LOCKED);
1318
1319         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1320                                   changeset);
1321 }
1322
1323 /*
1324  * either insert or lock state struct between start and end use mask to tell
1325  * us if waiting is desired.
1326  */
1327 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1328                      struct extent_state **cached_state)
1329 {
1330         int err;
1331         u64 failed_start;
1332
1333         while (1) {
1334                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1335                                        EXTENT_LOCKED, &failed_start,
1336                                        cached_state, GFP_NOFS, NULL);
1337                 if (err == -EEXIST) {
1338                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1339                         start = failed_start;
1340                 } else
1341                         break;
1342                 WARN_ON(start > end);
1343         }
1344         return err;
1345 }
1346
1347 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1348 {
1349         int err;
1350         u64 failed_start;
1351
1352         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1353                                &failed_start, NULL, GFP_NOFS, NULL);
1354         if (err == -EEXIST) {
1355                 if (failed_start > start)
1356                         clear_extent_bit(tree, start, failed_start - 1,
1357                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1358                 return 0;
1359         }
1360         return 1;
1361 }
1362
1363 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1364 {
1365         unsigned long index = start >> PAGE_SHIFT;
1366         unsigned long end_index = end >> PAGE_SHIFT;
1367         struct page *page;
1368
1369         while (index <= end_index) {
1370                 page = find_get_page(inode->i_mapping, index);
1371                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1372                 clear_page_dirty_for_io(page);
1373                 put_page(page);
1374                 index++;
1375         }
1376 }
1377
1378 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1379 {
1380         unsigned long index = start >> PAGE_SHIFT;
1381         unsigned long end_index = end >> PAGE_SHIFT;
1382         struct page *page;
1383
1384         while (index <= end_index) {
1385                 page = find_get_page(inode->i_mapping, index);
1386                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1387                 __set_page_dirty_nobuffers(page);
1388                 account_page_redirty(page);
1389                 put_page(page);
1390                 index++;
1391         }
1392 }
1393
1394 /*
1395  * helper function to set both pages and extents in the tree writeback
1396  */
1397 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1398 {
1399         unsigned long index = start >> PAGE_SHIFT;
1400         unsigned long end_index = end >> PAGE_SHIFT;
1401         struct page *page;
1402
1403         while (index <= end_index) {
1404                 page = find_get_page(tree->mapping, index);
1405                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1406                 set_page_writeback(page);
1407                 put_page(page);
1408                 index++;
1409         }
1410 }
1411
1412 /* find the first state struct with 'bits' set after 'start', and
1413  * return it.  tree->lock must be held.  NULL will returned if
1414  * nothing was found after 'start'
1415  */
1416 static struct extent_state *
1417 find_first_extent_bit_state(struct extent_io_tree *tree,
1418                             u64 start, unsigned bits)
1419 {
1420         struct rb_node *node;
1421         struct extent_state *state;
1422
1423         /*
1424          * this search will find all the extents that end after
1425          * our range starts.
1426          */
1427         node = tree_search(tree, start);
1428         if (!node)
1429                 goto out;
1430
1431         while (1) {
1432                 state = rb_entry(node, struct extent_state, rb_node);
1433                 if (state->end >= start && (state->state & bits))
1434                         return state;
1435
1436                 node = rb_next(node);
1437                 if (!node)
1438                         break;
1439         }
1440 out:
1441         return NULL;
1442 }
1443
1444 /*
1445  * find the first offset in the io tree with 'bits' set. zero is
1446  * returned if we find something, and *start_ret and *end_ret are
1447  * set to reflect the state struct that was found.
1448  *
1449  * If nothing was found, 1 is returned. If found something, return 0.
1450  */
1451 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1452                           u64 *start_ret, u64 *end_ret, unsigned bits,
1453                           struct extent_state **cached_state)
1454 {
1455         struct extent_state *state;
1456         struct rb_node *n;
1457         int ret = 1;
1458
1459         spin_lock(&tree->lock);
1460         if (cached_state && *cached_state) {
1461                 state = *cached_state;
1462                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1463                         n = rb_next(&state->rb_node);
1464                         while (n) {
1465                                 state = rb_entry(n, struct extent_state,
1466                                                  rb_node);
1467                                 if (state->state & bits)
1468                                         goto got_it;
1469                                 n = rb_next(n);
1470                         }
1471                         free_extent_state(*cached_state);
1472                         *cached_state = NULL;
1473                         goto out;
1474                 }
1475                 free_extent_state(*cached_state);
1476                 *cached_state = NULL;
1477         }
1478
1479         state = find_first_extent_bit_state(tree, start, bits);
1480 got_it:
1481         if (state) {
1482                 cache_state_if_flags(state, cached_state, 0);
1483                 *start_ret = state->start;
1484                 *end_ret = state->end;
1485                 ret = 0;
1486         }
1487 out:
1488         spin_unlock(&tree->lock);
1489         return ret;
1490 }
1491
1492 /*
1493  * find a contiguous range of bytes in the file marked as delalloc, not
1494  * more than 'max_bytes'.  start and end are used to return the range,
1495  *
1496  * 1 is returned if we find something, 0 if nothing was in the tree
1497  */
1498 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1499                                         u64 *start, u64 *end, u64 max_bytes,
1500                                         struct extent_state **cached_state)
1501 {
1502         struct rb_node *node;
1503         struct extent_state *state;
1504         u64 cur_start = *start;
1505         u64 found = 0;
1506         u64 total_bytes = 0;
1507
1508         spin_lock(&tree->lock);
1509
1510         /*
1511          * this search will find all the extents that end after
1512          * our range starts.
1513          */
1514         node = tree_search(tree, cur_start);
1515         if (!node) {
1516                 if (!found)
1517                         *end = (u64)-1;
1518                 goto out;
1519         }
1520
1521         while (1) {
1522                 state = rb_entry(node, struct extent_state, rb_node);
1523                 if (found && (state->start != cur_start ||
1524                               (state->state & EXTENT_BOUNDARY))) {
1525                         goto out;
1526                 }
1527                 if (!(state->state & EXTENT_DELALLOC)) {
1528                         if (!found)
1529                                 *end = state->end;
1530                         goto out;
1531                 }
1532                 if (!found) {
1533                         *start = state->start;
1534                         *cached_state = state;
1535                         atomic_inc(&state->refs);
1536                 }
1537                 found++;
1538                 *end = state->end;
1539                 cur_start = state->end + 1;
1540                 node = rb_next(node);
1541                 total_bytes += state->end - state->start + 1;
1542                 if (total_bytes >= max_bytes)
1543                         break;
1544                 if (!node)
1545                         break;
1546         }
1547 out:
1548         spin_unlock(&tree->lock);
1549         return found;
1550 }
1551
1552 static noinline void __unlock_for_delalloc(struct inode *inode,
1553                                            struct page *locked_page,
1554                                            u64 start, u64 end)
1555 {
1556         int ret;
1557         struct page *pages[16];
1558         unsigned long index = start >> PAGE_SHIFT;
1559         unsigned long end_index = end >> PAGE_SHIFT;
1560         unsigned long nr_pages = end_index - index + 1;
1561         int i;
1562
1563         if (index == locked_page->index && end_index == index)
1564                 return;
1565
1566         while (nr_pages > 0) {
1567                 ret = find_get_pages_contig(inode->i_mapping, index,
1568                                      min_t(unsigned long, nr_pages,
1569                                      ARRAY_SIZE(pages)), pages);
1570                 for (i = 0; i < ret; i++) {
1571                         if (pages[i] != locked_page)
1572                                 unlock_page(pages[i]);
1573                         put_page(pages[i]);
1574                 }
1575                 nr_pages -= ret;
1576                 index += ret;
1577                 cond_resched();
1578         }
1579 }
1580
1581 static noinline int lock_delalloc_pages(struct inode *inode,
1582                                         struct page *locked_page,
1583                                         u64 delalloc_start,
1584                                         u64 delalloc_end)
1585 {
1586         unsigned long index = delalloc_start >> PAGE_SHIFT;
1587         unsigned long start_index = index;
1588         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1589         unsigned long pages_locked = 0;
1590         struct page *pages[16];
1591         unsigned long nrpages;
1592         int ret;
1593         int i;
1594
1595         /* the caller is responsible for locking the start index */
1596         if (index == locked_page->index && index == end_index)
1597                 return 0;
1598
1599         /* skip the page at the start index */
1600         nrpages = end_index - index + 1;
1601         while (nrpages > 0) {
1602                 ret = find_get_pages_contig(inode->i_mapping, index,
1603                                      min_t(unsigned long,
1604                                      nrpages, ARRAY_SIZE(pages)), pages);
1605                 if (ret == 0) {
1606                         ret = -EAGAIN;
1607                         goto done;
1608                 }
1609                 /* now we have an array of pages, lock them all */
1610                 for (i = 0; i < ret; i++) {
1611                         /*
1612                          * the caller is taking responsibility for
1613                          * locked_page
1614                          */
1615                         if (pages[i] != locked_page) {
1616                                 lock_page(pages[i]);
1617                                 if (!PageDirty(pages[i]) ||
1618                                     pages[i]->mapping != inode->i_mapping) {
1619                                         ret = -EAGAIN;
1620                                         unlock_page(pages[i]);
1621                                         put_page(pages[i]);
1622                                         goto done;
1623                                 }
1624                         }
1625                         put_page(pages[i]);
1626                         pages_locked++;
1627                 }
1628                 nrpages -= ret;
1629                 index += ret;
1630                 cond_resched();
1631         }
1632         ret = 0;
1633 done:
1634         if (ret && pages_locked) {
1635                 __unlock_for_delalloc(inode, locked_page,
1636                               delalloc_start,
1637                               ((u64)(start_index + pages_locked - 1)) <<
1638                               PAGE_SHIFT);
1639         }
1640         return ret;
1641 }
1642
1643 /*
1644  * find a contiguous range of bytes in the file marked as delalloc, not
1645  * more than 'max_bytes'.  start and end are used to return the range,
1646  *
1647  * 1 is returned if we find something, 0 if nothing was in the tree
1648  */
1649 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1650                                     struct extent_io_tree *tree,
1651                                     struct page *locked_page, u64 *start,
1652                                     u64 *end, u64 max_bytes)
1653 {
1654         u64 delalloc_start;
1655         u64 delalloc_end;
1656         u64 found;
1657         struct extent_state *cached_state = NULL;
1658         int ret;
1659         int loops = 0;
1660
1661 again:
1662         /* step one, find a bunch of delalloc bytes starting at start */
1663         delalloc_start = *start;
1664         delalloc_end = 0;
1665         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1666                                     max_bytes, &cached_state);
1667         if (!found || delalloc_end <= *start) {
1668                 *start = delalloc_start;
1669                 *end = delalloc_end;
1670                 free_extent_state(cached_state);
1671                 return 0;
1672         }
1673
1674         /*
1675          * start comes from the offset of locked_page.  We have to lock
1676          * pages in order, so we can't process delalloc bytes before
1677          * locked_page
1678          */
1679         if (delalloc_start < *start)
1680                 delalloc_start = *start;
1681
1682         /*
1683          * make sure to limit the number of pages we try to lock down
1684          */
1685         if (delalloc_end + 1 - delalloc_start > max_bytes)
1686                 delalloc_end = delalloc_start + max_bytes - 1;
1687
1688         /* step two, lock all the pages after the page that has start */
1689         ret = lock_delalloc_pages(inode, locked_page,
1690                                   delalloc_start, delalloc_end);
1691         if (ret == -EAGAIN) {
1692                 /* some of the pages are gone, lets avoid looping by
1693                  * shortening the size of the delalloc range we're searching
1694                  */
1695                 free_extent_state(cached_state);
1696                 cached_state = NULL;
1697                 if (!loops) {
1698                         max_bytes = PAGE_SIZE;
1699                         loops = 1;
1700                         goto again;
1701                 } else {
1702                         found = 0;
1703                         goto out_failed;
1704                 }
1705         }
1706         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1707
1708         /* step three, lock the state bits for the whole range */
1709         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1710
1711         /* then test to make sure it is all still delalloc */
1712         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1713                              EXTENT_DELALLOC, 1, cached_state);
1714         if (!ret) {
1715                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1716                                      &cached_state, GFP_NOFS);
1717                 __unlock_for_delalloc(inode, locked_page,
1718                               delalloc_start, delalloc_end);
1719                 cond_resched();
1720                 goto again;
1721         }
1722         free_extent_state(cached_state);
1723         *start = delalloc_start;
1724         *end = delalloc_end;
1725 out_failed:
1726         return found;
1727 }
1728
1729 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1730                                  u64 delalloc_end, struct page *locked_page,
1731                                  unsigned clear_bits,
1732                                  unsigned long page_ops)
1733 {
1734         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1735         int ret;
1736         struct page *pages[16];
1737         unsigned long index = start >> PAGE_SHIFT;
1738         unsigned long end_index = end >> PAGE_SHIFT;
1739         unsigned long nr_pages = end_index - index + 1;
1740         int i;
1741
1742         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1743         if (page_ops == 0)
1744                 return;
1745
1746         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1747                 mapping_set_error(inode->i_mapping, -EIO);
1748
1749         while (nr_pages > 0) {
1750                 ret = find_get_pages_contig(inode->i_mapping, index,
1751                                      min_t(unsigned long,
1752                                      nr_pages, ARRAY_SIZE(pages)), pages);
1753                 for (i = 0; i < ret; i++) {
1754
1755                         if (page_ops & PAGE_SET_PRIVATE2)
1756                                 SetPagePrivate2(pages[i]);
1757
1758                         if (pages[i] == locked_page) {
1759                                 put_page(pages[i]);
1760                                 continue;
1761                         }
1762                         if (page_ops & PAGE_CLEAR_DIRTY)
1763                                 clear_page_dirty_for_io(pages[i]);
1764                         if (page_ops & PAGE_SET_WRITEBACK)
1765                                 set_page_writeback(pages[i]);
1766                         if (page_ops & PAGE_SET_ERROR)
1767                                 SetPageError(pages[i]);
1768                         if (page_ops & PAGE_END_WRITEBACK)
1769                                 end_page_writeback(pages[i]);
1770                         if (page_ops & PAGE_UNLOCK)
1771                                 unlock_page(pages[i]);
1772                         put_page(pages[i]);
1773                 }
1774                 nr_pages -= ret;
1775                 index += ret;
1776                 cond_resched();
1777         }
1778 }
1779
1780 /*
1781  * count the number of bytes in the tree that have a given bit(s)
1782  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1783  * cached.  The total number found is returned.
1784  */
1785 u64 count_range_bits(struct extent_io_tree *tree,
1786                      u64 *start, u64 search_end, u64 max_bytes,
1787                      unsigned bits, int contig)
1788 {
1789         struct rb_node *node;
1790         struct extent_state *state;
1791         u64 cur_start = *start;
1792         u64 total_bytes = 0;
1793         u64 last = 0;
1794         int found = 0;
1795
1796         if (WARN_ON(search_end <= cur_start))
1797                 return 0;
1798
1799         spin_lock(&tree->lock);
1800         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1801                 total_bytes = tree->dirty_bytes;
1802                 goto out;
1803         }
1804         /*
1805          * this search will find all the extents that end after
1806          * our range starts.
1807          */
1808         node = tree_search(tree, cur_start);
1809         if (!node)
1810                 goto out;
1811
1812         while (1) {
1813                 state = rb_entry(node, struct extent_state, rb_node);
1814                 if (state->start > search_end)
1815                         break;
1816                 if (contig && found && state->start > last + 1)
1817                         break;
1818                 if (state->end >= cur_start && (state->state & bits) == bits) {
1819                         total_bytes += min(search_end, state->end) + 1 -
1820                                        max(cur_start, state->start);
1821                         if (total_bytes >= max_bytes)
1822                                 break;
1823                         if (!found) {
1824                                 *start = max(cur_start, state->start);
1825                                 found = 1;
1826                         }
1827                         last = state->end;
1828                 } else if (contig && found) {
1829                         break;
1830                 }
1831                 node = rb_next(node);
1832                 if (!node)
1833                         break;
1834         }
1835 out:
1836         spin_unlock(&tree->lock);
1837         return total_bytes;
1838 }
1839
1840 /*
1841  * set the private field for a given byte offset in the tree.  If there isn't
1842  * an extent_state there already, this does nothing.
1843  */
1844 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1845                 struct io_failure_record *failrec)
1846 {
1847         struct rb_node *node;
1848         struct extent_state *state;
1849         int ret = 0;
1850
1851         spin_lock(&tree->lock);
1852         /*
1853          * this search will find all the extents that end after
1854          * our range starts.
1855          */
1856         node = tree_search(tree, start);
1857         if (!node) {
1858                 ret = -ENOENT;
1859                 goto out;
1860         }
1861         state = rb_entry(node, struct extent_state, rb_node);
1862         if (state->start != start) {
1863                 ret = -ENOENT;
1864                 goto out;
1865         }
1866         state->failrec = failrec;
1867 out:
1868         spin_unlock(&tree->lock);
1869         return ret;
1870 }
1871
1872 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1873                 struct io_failure_record **failrec)
1874 {
1875         struct rb_node *node;
1876         struct extent_state *state;
1877         int ret = 0;
1878
1879         spin_lock(&tree->lock);
1880         /*
1881          * this search will find all the extents that end after
1882          * our range starts.
1883          */
1884         node = tree_search(tree, start);
1885         if (!node) {
1886                 ret = -ENOENT;
1887                 goto out;
1888         }
1889         state = rb_entry(node, struct extent_state, rb_node);
1890         if (state->start != start) {
1891                 ret = -ENOENT;
1892                 goto out;
1893         }
1894         *failrec = state->failrec;
1895 out:
1896         spin_unlock(&tree->lock);
1897         return ret;
1898 }
1899
1900 /*
1901  * searches a range in the state tree for a given mask.
1902  * If 'filled' == 1, this returns 1 only if every extent in the tree
1903  * has the bits set.  Otherwise, 1 is returned if any bit in the
1904  * range is found set.
1905  */
1906 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1907                    unsigned bits, int filled, struct extent_state *cached)
1908 {
1909         struct extent_state *state = NULL;
1910         struct rb_node *node;
1911         int bitset = 0;
1912
1913         spin_lock(&tree->lock);
1914         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1915             cached->end > start)
1916                 node = &cached->rb_node;
1917         else
1918                 node = tree_search(tree, start);
1919         while (node && start <= end) {
1920                 state = rb_entry(node, struct extent_state, rb_node);
1921
1922                 if (filled && state->start > start) {
1923                         bitset = 0;
1924                         break;
1925                 }
1926
1927                 if (state->start > end)
1928                         break;
1929
1930                 if (state->state & bits) {
1931                         bitset = 1;
1932                         if (!filled)
1933                                 break;
1934                 } else if (filled) {
1935                         bitset = 0;
1936                         break;
1937                 }
1938
1939                 if (state->end == (u64)-1)
1940                         break;
1941
1942                 start = state->end + 1;
1943                 if (start > end)
1944                         break;
1945                 node = rb_next(node);
1946                 if (!node) {
1947                         if (filled)
1948                                 bitset = 0;
1949                         break;
1950                 }
1951         }
1952         spin_unlock(&tree->lock);
1953         return bitset;
1954 }
1955
1956 /*
1957  * helper function to set a given page up to date if all the
1958  * extents in the tree for that page are up to date
1959  */
1960 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1961 {
1962         u64 start = page_offset(page);
1963         u64 end = start + PAGE_SIZE - 1;
1964         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1965                 SetPageUptodate(page);
1966 }
1967
1968 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1969 {
1970         int ret;
1971         int err = 0;
1972         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1973
1974         set_state_failrec(failure_tree, rec->start, NULL);
1975         ret = clear_extent_bits(failure_tree, rec->start,
1976                                 rec->start + rec->len - 1,
1977                                 EXTENT_LOCKED | EXTENT_DIRTY);
1978         if (ret)
1979                 err = ret;
1980
1981         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1982                                 rec->start + rec->len - 1,
1983                                 EXTENT_DAMAGED);
1984         if (ret && !err)
1985                 err = ret;
1986
1987         kfree(rec);
1988         return err;
1989 }
1990
1991 /*
1992  * this bypasses the standard btrfs submit functions deliberately, as
1993  * the standard behavior is to write all copies in a raid setup. here we only
1994  * want to write the one bad copy. so we do the mapping for ourselves and issue
1995  * submit_bio directly.
1996  * to avoid any synchronization issues, wait for the data after writing, which
1997  * actually prevents the read that triggered the error from finishing.
1998  * currently, there can be no more than two copies of every data bit. thus,
1999  * exactly one rewrite is required.
2000  */
2001 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2002                       struct page *page, unsigned int pg_offset, int mirror_num)
2003 {
2004         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2005         struct bio *bio;
2006         struct btrfs_device *dev;
2007         u64 map_length = 0;
2008         u64 sector;
2009         struct btrfs_bio *bbio = NULL;
2010         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2011         int ret;
2012
2013         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2014         BUG_ON(!mirror_num);
2015
2016         /* we can't repair anything in raid56 yet */
2017         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2018                 return 0;
2019
2020         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2021         if (!bio)
2022                 return -EIO;
2023         bio->bi_iter.bi_size = 0;
2024         map_length = length;
2025
2026         /*
2027          * Avoid races with device replace and make sure our bbio has devices
2028          * associated to its stripes that don't go away while we are doing the
2029          * read repair operation.
2030          */
2031         btrfs_bio_counter_inc_blocked(fs_info);
2032         ret = btrfs_map_block(fs_info, WRITE, logical,
2033                               &map_length, &bbio, mirror_num);
2034         if (ret) {
2035                 btrfs_bio_counter_dec(fs_info);
2036                 bio_put(bio);
2037                 return -EIO;
2038         }
2039         BUG_ON(mirror_num != bbio->mirror_num);
2040         sector = bbio->stripes[mirror_num-1].physical >> 9;
2041         bio->bi_iter.bi_sector = sector;
2042         dev = bbio->stripes[mirror_num-1].dev;
2043         btrfs_put_bbio(bbio);
2044         if (!dev || !dev->bdev || !dev->writeable) {
2045                 btrfs_bio_counter_dec(fs_info);
2046                 bio_put(bio);
2047                 return -EIO;
2048         }
2049         bio->bi_bdev = dev->bdev;
2050         bio_set_op_attrs(bio, REQ_OP_WRITE, WRITE_SYNC);
2051         bio_add_page(bio, page, length, pg_offset);
2052
2053         if (btrfsic_submit_bio_wait(bio)) {
2054                 /* try to remap that extent elsewhere? */
2055                 btrfs_bio_counter_dec(fs_info);
2056                 bio_put(bio);
2057                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2058                 return -EIO;
2059         }
2060
2061         btrfs_info_rl_in_rcu(fs_info,
2062                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2063                                   btrfs_ino(inode), start,
2064                                   rcu_str_deref(dev->name), sector);
2065         btrfs_bio_counter_dec(fs_info);
2066         bio_put(bio);
2067         return 0;
2068 }
2069
2070 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2071                          int mirror_num)
2072 {
2073         u64 start = eb->start;
2074         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2075         int ret = 0;
2076
2077         if (root->fs_info->sb->s_flags & MS_RDONLY)
2078                 return -EROFS;
2079
2080         for (i = 0; i < num_pages; i++) {
2081                 struct page *p = eb->pages[i];
2082
2083                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2084                                         PAGE_SIZE, start, p,
2085                                         start - page_offset(p), mirror_num);
2086                 if (ret)
2087                         break;
2088                 start += PAGE_SIZE;
2089         }
2090
2091         return ret;
2092 }
2093
2094 /*
2095  * each time an IO finishes, we do a fast check in the IO failure tree
2096  * to see if we need to process or clean up an io_failure_record
2097  */
2098 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2099                      unsigned int pg_offset)
2100 {
2101         u64 private;
2102         struct io_failure_record *failrec;
2103         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2104         struct extent_state *state;
2105         int num_copies;
2106         int ret;
2107
2108         private = 0;
2109         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2110                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2111         if (!ret)
2112                 return 0;
2113
2114         ret = get_state_failrec(&BTRFS_I(inode)->io_failure_tree, start,
2115                         &failrec);
2116         if (ret)
2117                 return 0;
2118
2119         BUG_ON(!failrec->this_mirror);
2120
2121         if (failrec->in_validation) {
2122                 /* there was no real error, just free the record */
2123                 btrfs_debug(fs_info,
2124                         "clean_io_failure: freeing dummy error at %llu",
2125                         failrec->start);
2126                 goto out;
2127         }
2128         if (fs_info->sb->s_flags & MS_RDONLY)
2129                 goto out;
2130
2131         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2132         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2133                                             failrec->start,
2134                                             EXTENT_LOCKED);
2135         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2136
2137         if (state && state->start <= failrec->start &&
2138             state->end >= failrec->start + failrec->len - 1) {
2139                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2140                                               failrec->len);
2141                 if (num_copies > 1)  {
2142                         repair_io_failure(inode, start, failrec->len,
2143                                           failrec->logical, page,
2144                                           pg_offset, failrec->failed_mirror);
2145                 }
2146         }
2147
2148 out:
2149         free_io_failure(inode, failrec);
2150
2151         return 0;
2152 }
2153
2154 /*
2155  * Can be called when
2156  * - hold extent lock
2157  * - under ordered extent
2158  * - the inode is freeing
2159  */
2160 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2161 {
2162         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2163         struct io_failure_record *failrec;
2164         struct extent_state *state, *next;
2165
2166         if (RB_EMPTY_ROOT(&failure_tree->state))
2167                 return;
2168
2169         spin_lock(&failure_tree->lock);
2170         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2171         while (state) {
2172                 if (state->start > end)
2173                         break;
2174
2175                 ASSERT(state->end <= end);
2176
2177                 next = next_state(state);
2178
2179                 failrec = state->failrec;
2180                 free_extent_state(state);
2181                 kfree(failrec);
2182
2183                 state = next;
2184         }
2185         spin_unlock(&failure_tree->lock);
2186 }
2187
2188 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2189                 struct io_failure_record **failrec_ret)
2190 {
2191         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2192         struct io_failure_record *failrec;
2193         struct extent_map *em;
2194         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2195         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2196         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2197         int ret;
2198         u64 logical;
2199
2200         ret = get_state_failrec(failure_tree, start, &failrec);
2201         if (ret) {
2202                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2203                 if (!failrec)
2204                         return -ENOMEM;
2205
2206                 failrec->start = start;
2207                 failrec->len = end - start + 1;
2208                 failrec->this_mirror = 0;
2209                 failrec->bio_flags = 0;
2210                 failrec->in_validation = 0;
2211
2212                 read_lock(&em_tree->lock);
2213                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2214                 if (!em) {
2215                         read_unlock(&em_tree->lock);
2216                         kfree(failrec);
2217                         return -EIO;
2218                 }
2219
2220                 if (em->start > start || em->start + em->len <= start) {
2221                         free_extent_map(em);
2222                         em = NULL;
2223                 }
2224                 read_unlock(&em_tree->lock);
2225                 if (!em) {
2226                         kfree(failrec);
2227                         return -EIO;
2228                 }
2229
2230                 logical = start - em->start;
2231                 logical = em->block_start + logical;
2232                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2233                         logical = em->block_start;
2234                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2235                         extent_set_compress_type(&failrec->bio_flags,
2236                                                  em->compress_type);
2237                 }
2238
2239                 btrfs_debug(fs_info,
2240                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2241                         logical, start, failrec->len);
2242
2243                 failrec->logical = logical;
2244                 free_extent_map(em);
2245
2246                 /* set the bits in the private failure tree */
2247                 ret = set_extent_bits(failure_tree, start, end,
2248                                         EXTENT_LOCKED | EXTENT_DIRTY);
2249                 if (ret >= 0)
2250                         ret = set_state_failrec(failure_tree, start, failrec);
2251                 /* set the bits in the inode's tree */
2252                 if (ret >= 0)
2253                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2254                 if (ret < 0) {
2255                         kfree(failrec);
2256                         return ret;
2257                 }
2258         } else {
2259                 btrfs_debug(fs_info,
2260                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2261                         failrec->logical, failrec->start, failrec->len,
2262                         failrec->in_validation);
2263                 /*
2264                  * when data can be on disk more than twice, add to failrec here
2265                  * (e.g. with a list for failed_mirror) to make
2266                  * clean_io_failure() clean all those errors at once.
2267                  */
2268         }
2269
2270         *failrec_ret = failrec;
2271
2272         return 0;
2273 }
2274
2275 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2276                            struct io_failure_record *failrec, int failed_mirror)
2277 {
2278         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2279         int num_copies;
2280
2281         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2282         if (num_copies == 1) {
2283                 /*
2284                  * we only have a single copy of the data, so don't bother with
2285                  * all the retry and error correction code that follows. no
2286                  * matter what the error is, it is very likely to persist.
2287                  */
2288                 btrfs_debug(fs_info,
2289                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2290                         num_copies, failrec->this_mirror, failed_mirror);
2291                 return 0;
2292         }
2293
2294         /*
2295          * there are two premises:
2296          *      a) deliver good data to the caller
2297          *      b) correct the bad sectors on disk
2298          */
2299         if (failed_bio->bi_vcnt > 1) {
2300                 /*
2301                  * to fulfill b), we need to know the exact failing sectors, as
2302                  * we don't want to rewrite any more than the failed ones. thus,
2303                  * we need separate read requests for the failed bio
2304                  *
2305                  * if the following BUG_ON triggers, our validation request got
2306                  * merged. we need separate requests for our algorithm to work.
2307                  */
2308                 BUG_ON(failrec->in_validation);
2309                 failrec->in_validation = 1;
2310                 failrec->this_mirror = failed_mirror;
2311         } else {
2312                 /*
2313                  * we're ready to fulfill a) and b) alongside. get a good copy
2314                  * of the failed sector and if we succeed, we have setup
2315                  * everything for repair_io_failure to do the rest for us.
2316                  */
2317                 if (failrec->in_validation) {
2318                         BUG_ON(failrec->this_mirror != failed_mirror);
2319                         failrec->in_validation = 0;
2320                         failrec->this_mirror = 0;
2321                 }
2322                 failrec->failed_mirror = failed_mirror;
2323                 failrec->this_mirror++;
2324                 if (failrec->this_mirror == failed_mirror)
2325                         failrec->this_mirror++;
2326         }
2327
2328         if (failrec->this_mirror > num_copies) {
2329                 btrfs_debug(fs_info,
2330                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2331                         num_copies, failrec->this_mirror, failed_mirror);
2332                 return 0;
2333         }
2334
2335         return 1;
2336 }
2337
2338
2339 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2340                                     struct io_failure_record *failrec,
2341                                     struct page *page, int pg_offset, int icsum,
2342                                     bio_end_io_t *endio_func, void *data)
2343 {
2344         struct bio *bio;
2345         struct btrfs_io_bio *btrfs_failed_bio;
2346         struct btrfs_io_bio *btrfs_bio;
2347
2348         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2349         if (!bio)
2350                 return NULL;
2351
2352         bio->bi_end_io = endio_func;
2353         bio->bi_iter.bi_sector = failrec->logical >> 9;
2354         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2355         bio->bi_iter.bi_size = 0;
2356         bio->bi_private = data;
2357
2358         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2359         if (btrfs_failed_bio->csum) {
2360                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2361                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2362
2363                 btrfs_bio = btrfs_io_bio(bio);
2364                 btrfs_bio->csum = btrfs_bio->csum_inline;
2365                 icsum *= csum_size;
2366                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2367                        csum_size);
2368         }
2369
2370         bio_add_page(bio, page, failrec->len, pg_offset);
2371
2372         return bio;
2373 }
2374
2375 /*
2376  * this is a generic handler for readpage errors (default
2377  * readpage_io_failed_hook). if other copies exist, read those and write back
2378  * good data to the failed position. does not investigate in remapping the
2379  * failed extent elsewhere, hoping the device will be smart enough to do this as
2380  * needed
2381  */
2382
2383 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2384                               struct page *page, u64 start, u64 end,
2385                               int failed_mirror)
2386 {
2387         struct io_failure_record *failrec;
2388         struct inode *inode = page->mapping->host;
2389         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2390         struct bio *bio;
2391         int read_mode;
2392         int ret;
2393
2394         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2395
2396         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2397         if (ret)
2398                 return ret;
2399
2400         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2401         if (!ret) {
2402                 free_io_failure(inode, failrec);
2403                 return -EIO;
2404         }
2405
2406         if (failed_bio->bi_vcnt > 1)
2407                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2408         else
2409                 read_mode = READ_SYNC;
2410
2411         phy_offset >>= inode->i_sb->s_blocksize_bits;
2412         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2413                                       start - page_offset(page),
2414                                       (int)phy_offset, failed_bio->bi_end_io,
2415                                       NULL);
2416         if (!bio) {
2417                 free_io_failure(inode, failrec);
2418                 return -EIO;
2419         }
2420         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2421
2422         btrfs_debug(btrfs_sb(inode->i_sb),
2423                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2424                 read_mode, failrec->this_mirror, failrec->in_validation);
2425
2426         ret = tree->ops->submit_bio_hook(inode, bio, failrec->this_mirror,
2427                                          failrec->bio_flags, 0);
2428         if (ret) {
2429                 free_io_failure(inode, failrec);
2430                 bio_put(bio);
2431         }
2432
2433         return ret;
2434 }
2435
2436 /* lots and lots of room for performance fixes in the end_bio funcs */
2437
2438 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2439 {
2440         int uptodate = (err == 0);
2441         struct extent_io_tree *tree;
2442         int ret = 0;
2443
2444         tree = &BTRFS_I(page->mapping->host)->io_tree;
2445
2446         if (tree->ops && tree->ops->writepage_end_io_hook) {
2447                 ret = tree->ops->writepage_end_io_hook(page, start,
2448                                                end, NULL, uptodate);
2449                 if (ret)
2450                         uptodate = 0;
2451         }
2452
2453         if (!uptodate) {
2454                 ClearPageUptodate(page);
2455                 SetPageError(page);
2456                 ret = ret < 0 ? ret : -EIO;
2457                 mapping_set_error(page->mapping, ret);
2458         }
2459 }
2460
2461 /*
2462  * after a writepage IO is done, we need to:
2463  * clear the uptodate bits on error
2464  * clear the writeback bits in the extent tree for this IO
2465  * end_page_writeback if the page has no more pending IO
2466  *
2467  * Scheduling is not allowed, so the extent state tree is expected
2468  * to have one and only one object corresponding to this IO.
2469  */
2470 static void end_bio_extent_writepage(struct bio *bio)
2471 {
2472         struct bio_vec *bvec;
2473         u64 start;
2474         u64 end;
2475         int i;
2476
2477         bio_for_each_segment_all(bvec, bio, i) {
2478                 struct page *page = bvec->bv_page;
2479
2480                 /* We always issue full-page reads, but if some block
2481                  * in a page fails to read, blk_update_request() will
2482                  * advance bv_offset and adjust bv_len to compensate.
2483                  * Print a warning for nonzero offsets, and an error
2484                  * if they don't add up to a full page.  */
2485                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2486                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2487                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2488                                    "partial page write in btrfs with offset %u and length %u",
2489                                         bvec->bv_offset, bvec->bv_len);
2490                         else
2491                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2492                                    "incomplete page write in btrfs with offset %u and length %u",
2493                                         bvec->bv_offset, bvec->bv_len);
2494                 }
2495
2496                 start = page_offset(page);
2497                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2498
2499                 end_extent_writepage(page, bio->bi_error, start, end);
2500                 end_page_writeback(page);
2501         }
2502
2503         bio_put(bio);
2504 }
2505
2506 static void
2507 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2508                               int uptodate)
2509 {
2510         struct extent_state *cached = NULL;
2511         u64 end = start + len - 1;
2512
2513         if (uptodate && tree->track_uptodate)
2514                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2515         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2516 }
2517
2518 /*
2519  * after a readpage IO is done, we need to:
2520  * clear the uptodate bits on error
2521  * set the uptodate bits if things worked
2522  * set the page up to date if all extents in the tree are uptodate
2523  * clear the lock bit in the extent tree
2524  * unlock the page if there are no other extents locked for it
2525  *
2526  * Scheduling is not allowed, so the extent state tree is expected
2527  * to have one and only one object corresponding to this IO.
2528  */
2529 static void end_bio_extent_readpage(struct bio *bio)
2530 {
2531         struct bio_vec *bvec;
2532         int uptodate = !bio->bi_error;
2533         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2534         struct extent_io_tree *tree;
2535         u64 offset = 0;
2536         u64 start;
2537         u64 end;
2538         u64 len;
2539         u64 extent_start = 0;
2540         u64 extent_len = 0;
2541         int mirror;
2542         int ret;
2543         int i;
2544
2545         bio_for_each_segment_all(bvec, bio, i) {
2546                 struct page *page = bvec->bv_page;
2547                 struct inode *inode = page->mapping->host;
2548                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2549
2550                 btrfs_debug(fs_info,
2551                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2552                         (u64)bio->bi_iter.bi_sector, bio->bi_error,
2553                         io_bio->mirror_num);
2554                 tree = &BTRFS_I(inode)->io_tree;
2555
2556                 /* We always issue full-page reads, but if some block
2557                  * in a page fails to read, blk_update_request() will
2558                  * advance bv_offset and adjust bv_len to compensate.
2559                  * Print a warning for nonzero offsets, and an error
2560                  * if they don't add up to a full page.  */
2561                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2562                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2563                                 btrfs_err(fs_info,
2564                                         "partial page read in btrfs with offset %u and length %u",
2565                                         bvec->bv_offset, bvec->bv_len);
2566                         else
2567                                 btrfs_info(fs_info,
2568                                         "incomplete page read in btrfs with offset %u and length %u",
2569                                         bvec->bv_offset, bvec->bv_len);
2570                 }
2571
2572                 start = page_offset(page);
2573                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2574                 len = bvec->bv_len;
2575
2576                 mirror = io_bio->mirror_num;
2577                 if (likely(uptodate && tree->ops &&
2578                            tree->ops->readpage_end_io_hook)) {
2579                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2580                                                               page, start, end,
2581                                                               mirror);
2582                         if (ret)
2583                                 uptodate = 0;
2584                         else
2585                                 clean_io_failure(inode, start, page, 0);
2586                 }
2587
2588                 if (likely(uptodate))
2589                         goto readpage_ok;
2590
2591                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2592                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2593                         if (!ret && !bio->bi_error)
2594                                 uptodate = 1;
2595                 } else {
2596                         /*
2597                          * The generic bio_readpage_error handles errors the
2598                          * following way: If possible, new read requests are
2599                          * created and submitted and will end up in
2600                          * end_bio_extent_readpage as well (if we're lucky, not
2601                          * in the !uptodate case). In that case it returns 0 and
2602                          * we just go on with the next page in our bio. If it
2603                          * can't handle the error it will return -EIO and we
2604                          * remain responsible for that page.
2605                          */
2606                         ret = bio_readpage_error(bio, offset, page, start, end,
2607                                                  mirror);
2608                         if (ret == 0) {
2609                                 uptodate = !bio->bi_error;
2610                                 offset += len;
2611                                 continue;
2612                         }
2613                 }
2614 readpage_ok:
2615                 if (likely(uptodate)) {
2616                         loff_t i_size = i_size_read(inode);
2617                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2618                         unsigned off;
2619
2620                         /* Zero out the end if this page straddles i_size */
2621                         off = i_size & (PAGE_SIZE-1);
2622                         if (page->index == end_index && off)
2623                                 zero_user_segment(page, off, PAGE_SIZE);
2624                         SetPageUptodate(page);
2625                 } else {
2626                         ClearPageUptodate(page);
2627                         SetPageError(page);
2628                 }
2629                 unlock_page(page);
2630                 offset += len;
2631
2632                 if (unlikely(!uptodate)) {
2633                         if (extent_len) {
2634                                 endio_readpage_release_extent(tree,
2635                                                               extent_start,
2636                                                               extent_len, 1);
2637                                 extent_start = 0;
2638                                 extent_len = 0;
2639                         }
2640                         endio_readpage_release_extent(tree, start,
2641                                                       end - start + 1, 0);
2642                 } else if (!extent_len) {
2643                         extent_start = start;
2644                         extent_len = end + 1 - start;
2645                 } else if (extent_start + extent_len == start) {
2646                         extent_len += end + 1 - start;
2647                 } else {
2648                         endio_readpage_release_extent(tree, extent_start,
2649                                                       extent_len, uptodate);
2650                         extent_start = start;
2651                         extent_len = end + 1 - start;
2652                 }
2653         }
2654
2655         if (extent_len)
2656                 endio_readpage_release_extent(tree, extent_start, extent_len,
2657                                               uptodate);
2658         if (io_bio->end_io)
2659                 io_bio->end_io(io_bio, bio->bi_error);
2660         bio_put(bio);
2661 }
2662
2663 /*
2664  * this allocates from the btrfs_bioset.  We're returning a bio right now
2665  * but you can call btrfs_io_bio for the appropriate container_of magic
2666  */
2667 struct bio *
2668 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2669                 gfp_t gfp_flags)
2670 {
2671         struct btrfs_io_bio *btrfs_bio;
2672         struct bio *bio;
2673
2674         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2675
2676         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2677                 while (!bio && (nr_vecs /= 2)) {
2678                         bio = bio_alloc_bioset(gfp_flags,
2679                                                nr_vecs, btrfs_bioset);
2680                 }
2681         }
2682
2683         if (bio) {
2684                 bio->bi_bdev = bdev;
2685                 bio->bi_iter.bi_sector = first_sector;
2686                 btrfs_bio = btrfs_io_bio(bio);
2687                 btrfs_bio->csum = NULL;
2688                 btrfs_bio->csum_allocated = NULL;
2689                 btrfs_bio->end_io = NULL;
2690         }
2691         return bio;
2692 }
2693
2694 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2695 {
2696         struct btrfs_io_bio *btrfs_bio;
2697         struct bio *new;
2698
2699         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2700         if (new) {
2701                 btrfs_bio = btrfs_io_bio(new);
2702                 btrfs_bio->csum = NULL;
2703                 btrfs_bio->csum_allocated = NULL;
2704                 btrfs_bio->end_io = NULL;
2705         }
2706         return new;
2707 }
2708
2709 /* this also allocates from the btrfs_bioset */
2710 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2711 {
2712         struct btrfs_io_bio *btrfs_bio;
2713         struct bio *bio;
2714
2715         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2716         if (bio) {
2717                 btrfs_bio = btrfs_io_bio(bio);
2718                 btrfs_bio->csum = NULL;
2719                 btrfs_bio->csum_allocated = NULL;
2720                 btrfs_bio->end_io = NULL;
2721         }
2722         return bio;
2723 }
2724
2725
2726 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2727                                        unsigned long bio_flags)
2728 {
2729         int ret = 0;
2730         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2731         struct page *page = bvec->bv_page;
2732         struct extent_io_tree *tree = bio->bi_private;
2733         u64 start;
2734
2735         start = page_offset(page) + bvec->bv_offset;
2736
2737         bio->bi_private = NULL;
2738         bio_get(bio);
2739
2740         if (tree->ops && tree->ops->submit_bio_hook)
2741                 ret = tree->ops->submit_bio_hook(page->mapping->host, bio,
2742                                            mirror_num, bio_flags, start);
2743         else
2744                 btrfsic_submit_bio(bio);
2745
2746         bio_put(bio);
2747         return ret;
2748 }
2749
2750 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2751                      unsigned long offset, size_t size, struct bio *bio,
2752                      unsigned long bio_flags)
2753 {
2754         int ret = 0;
2755         if (tree->ops && tree->ops->merge_bio_hook)
2756                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2757                                                 bio_flags);
2758         return ret;
2759
2760 }
2761
2762 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2763                               struct writeback_control *wbc,
2764                               struct page *page, sector_t sector,
2765                               size_t size, unsigned long offset,
2766                               struct block_device *bdev,
2767                               struct bio **bio_ret,
2768                               unsigned long max_pages,
2769                               bio_end_io_t end_io_func,
2770                               int mirror_num,
2771                               unsigned long prev_bio_flags,
2772                               unsigned long bio_flags,
2773                               bool force_bio_submit)
2774 {
2775         int ret = 0;
2776         struct bio *bio;
2777         int contig = 0;
2778         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2779         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2780
2781         if (bio_ret && *bio_ret) {
2782                 bio = *bio_ret;
2783                 if (old_compressed)
2784                         contig = bio->bi_iter.bi_sector == sector;
2785                 else
2786                         contig = bio_end_sector(bio) == sector;
2787
2788                 if (prev_bio_flags != bio_flags || !contig ||
2789                     force_bio_submit ||
2790                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2791                     bio_add_page(bio, page, page_size, offset) < page_size) {
2792                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2793                         if (ret < 0) {
2794                                 *bio_ret = NULL;
2795                                 return ret;
2796                         }
2797                         bio = NULL;
2798                 } else {
2799                         if (wbc)
2800                                 wbc_account_io(wbc, page, page_size);
2801                         return 0;
2802                 }
2803         }
2804
2805         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2806                         GFP_NOFS | __GFP_HIGH);
2807         if (!bio)
2808                 return -ENOMEM;
2809
2810         bio_add_page(bio, page, page_size, offset);
2811         bio->bi_end_io = end_io_func;
2812         bio->bi_private = tree;
2813         bio_set_op_attrs(bio, op, op_flags);
2814         if (wbc) {
2815                 wbc_init_bio(wbc, bio);
2816                 wbc_account_io(wbc, page, page_size);
2817         }
2818
2819         if (bio_ret)
2820                 *bio_ret = bio;
2821         else
2822                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2823
2824         return ret;
2825 }
2826
2827 static void attach_extent_buffer_page(struct extent_buffer *eb,
2828                                       struct page *page)
2829 {
2830         if (!PagePrivate(page)) {
2831                 SetPagePrivate(page);
2832                 get_page(page);
2833                 set_page_private(page, (unsigned long)eb);
2834         } else {
2835                 WARN_ON(page->private != (unsigned long)eb);
2836         }
2837 }
2838
2839 void set_page_extent_mapped(struct page *page)
2840 {
2841         if (!PagePrivate(page)) {
2842                 SetPagePrivate(page);
2843                 get_page(page);
2844                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2845         }
2846 }
2847
2848 static struct extent_map *
2849 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2850                  u64 start, u64 len, get_extent_t *get_extent,
2851                  struct extent_map **em_cached)
2852 {
2853         struct extent_map *em;
2854
2855         if (em_cached && *em_cached) {
2856                 em = *em_cached;
2857                 if (extent_map_in_tree(em) && start >= em->start &&
2858                     start < extent_map_end(em)) {
2859                         atomic_inc(&em->refs);
2860                         return em;
2861                 }
2862
2863                 free_extent_map(em);
2864                 *em_cached = NULL;
2865         }
2866
2867         em = get_extent(inode, page, pg_offset, start, len, 0);
2868         if (em_cached && !IS_ERR_OR_NULL(em)) {
2869                 BUG_ON(*em_cached);
2870                 atomic_inc(&em->refs);
2871                 *em_cached = em;
2872         }
2873         return em;
2874 }
2875 /*
2876  * basic readpage implementation.  Locked extent state structs are inserted
2877  * into the tree that are removed when the IO is done (by the end_io
2878  * handlers)
2879  * XXX JDM: This needs looking at to ensure proper page locking
2880  * return 0 on success, otherwise return error
2881  */
2882 static int __do_readpage(struct extent_io_tree *tree,
2883                          struct page *page,
2884                          get_extent_t *get_extent,
2885                          struct extent_map **em_cached,
2886                          struct bio **bio, int mirror_num,
2887                          unsigned long *bio_flags, int read_flags,
2888                          u64 *prev_em_start)
2889 {
2890         struct inode *inode = page->mapping->host;
2891         u64 start = page_offset(page);
2892         u64 page_end = start + PAGE_SIZE - 1;
2893         u64 end;
2894         u64 cur = start;
2895         u64 extent_offset;
2896         u64 last_byte = i_size_read(inode);
2897         u64 block_start;
2898         u64 cur_end;
2899         sector_t sector;
2900         struct extent_map *em;
2901         struct block_device *bdev;
2902         int ret = 0;
2903         int nr = 0;
2904         size_t pg_offset = 0;
2905         size_t iosize;
2906         size_t disk_io_size;
2907         size_t blocksize = inode->i_sb->s_blocksize;
2908         unsigned long this_bio_flag = 0;
2909
2910         set_page_extent_mapped(page);
2911
2912         end = page_end;
2913         if (!PageUptodate(page)) {
2914                 if (cleancache_get_page(page) == 0) {
2915                         BUG_ON(blocksize != PAGE_SIZE);
2916                         unlock_extent(tree, start, end);
2917                         goto out;
2918                 }
2919         }
2920
2921         if (page->index == last_byte >> PAGE_SHIFT) {
2922                 char *userpage;
2923                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2924
2925                 if (zero_offset) {
2926                         iosize = PAGE_SIZE - zero_offset;
2927                         userpage = kmap_atomic(page);
2928                         memset(userpage + zero_offset, 0, iosize);
2929                         flush_dcache_page(page);
2930                         kunmap_atomic(userpage);
2931                 }
2932         }
2933         while (cur <= end) {
2934                 unsigned long pnr = (last_byte >> PAGE_SHIFT) + 1;
2935                 bool force_bio_submit = false;
2936
2937                 if (cur >= last_byte) {
2938                         char *userpage;
2939                         struct extent_state *cached = NULL;
2940
2941                         iosize = PAGE_SIZE - pg_offset;
2942                         userpage = kmap_atomic(page);
2943                         memset(userpage + pg_offset, 0, iosize);
2944                         flush_dcache_page(page);
2945                         kunmap_atomic(userpage);
2946                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2947                                             &cached, GFP_NOFS);
2948                         unlock_extent_cached(tree, cur,
2949                                              cur + iosize - 1,
2950                                              &cached, GFP_NOFS);
2951                         break;
2952                 }
2953                 em = __get_extent_map(inode, page, pg_offset, cur,
2954                                       end - cur + 1, get_extent, em_cached);
2955                 if (IS_ERR_OR_NULL(em)) {
2956                         SetPageError(page);
2957                         unlock_extent(tree, cur, end);
2958                         break;
2959                 }
2960                 extent_offset = cur - em->start;
2961                 BUG_ON(extent_map_end(em) <= cur);
2962                 BUG_ON(end < cur);
2963
2964                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2965                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2966                         extent_set_compress_type(&this_bio_flag,
2967                                                  em->compress_type);
2968                 }
2969
2970                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2971                 cur_end = min(extent_map_end(em) - 1, end);
2972                 iosize = ALIGN(iosize, blocksize);
2973                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2974                         disk_io_size = em->block_len;
2975                         sector = em->block_start >> 9;
2976                 } else {
2977                         sector = (em->block_start + extent_offset) >> 9;
2978                         disk_io_size = iosize;
2979                 }
2980                 bdev = em->bdev;
2981                 block_start = em->block_start;
2982                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2983                         block_start = EXTENT_MAP_HOLE;
2984
2985                 /*
2986                  * If we have a file range that points to a compressed extent
2987                  * and it's followed by a consecutive file range that points to
2988                  * to the same compressed extent (possibly with a different
2989                  * offset and/or length, so it either points to the whole extent
2990                  * or only part of it), we must make sure we do not submit a
2991                  * single bio to populate the pages for the 2 ranges because
2992                  * this makes the compressed extent read zero out the pages
2993                  * belonging to the 2nd range. Imagine the following scenario:
2994                  *
2995                  *  File layout
2996                  *  [0 - 8K]                     [8K - 24K]
2997                  *    |                               |
2998                  *    |                               |
2999                  * points to extent X,         points to extent X,
3000                  * offset 4K, length of 8K     offset 0, length 16K
3001                  *
3002                  * [extent X, compressed length = 4K uncompressed length = 16K]
3003                  *
3004                  * If the bio to read the compressed extent covers both ranges,
3005                  * it will decompress extent X into the pages belonging to the
3006                  * first range and then it will stop, zeroing out the remaining
3007                  * pages that belong to the other range that points to extent X.
3008                  * So here we make sure we submit 2 bios, one for the first
3009                  * range and another one for the third range. Both will target
3010                  * the same physical extent from disk, but we can't currently
3011                  * make the compressed bio endio callback populate the pages
3012                  * for both ranges because each compressed bio is tightly
3013                  * coupled with a single extent map, and each range can have
3014                  * an extent map with a different offset value relative to the
3015                  * uncompressed data of our extent and different lengths. This
3016                  * is a corner case so we prioritize correctness over
3017                  * non-optimal behavior (submitting 2 bios for the same extent).
3018                  */
3019                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3020                     prev_em_start && *prev_em_start != (u64)-1 &&
3021                     *prev_em_start != em->orig_start)
3022                         force_bio_submit = true;
3023
3024                 if (prev_em_start)
3025                         *prev_em_start = em->orig_start;
3026
3027                 free_extent_map(em);
3028                 em = NULL;
3029
3030                 /* we've found a hole, just zero and go on */
3031                 if (block_start == EXTENT_MAP_HOLE) {
3032                         char *userpage;
3033                         struct extent_state *cached = NULL;
3034
3035                         userpage = kmap_atomic(page);
3036                         memset(userpage + pg_offset, 0, iosize);
3037                         flush_dcache_page(page);
3038                         kunmap_atomic(userpage);
3039
3040                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3041                                             &cached, GFP_NOFS);
3042                         unlock_extent_cached(tree, cur,
3043                                              cur + iosize - 1,
3044                                              &cached, GFP_NOFS);
3045                         cur = cur + iosize;
3046                         pg_offset += iosize;
3047                         continue;
3048                 }
3049                 /* the get_extent function already copied into the page */
3050                 if (test_range_bit(tree, cur, cur_end,
3051                                    EXTENT_UPTODATE, 1, NULL)) {
3052                         check_page_uptodate(tree, page);
3053                         unlock_extent(tree, cur, cur + iosize - 1);
3054                         cur = cur + iosize;
3055                         pg_offset += iosize;
3056                         continue;
3057                 }
3058                 /* we have an inline extent but it didn't get marked up
3059                  * to date.  Error out
3060                  */
3061                 if (block_start == EXTENT_MAP_INLINE) {
3062                         SetPageError(page);
3063                         unlock_extent(tree, cur, cur + iosize - 1);
3064                         cur = cur + iosize;
3065                         pg_offset += iosize;
3066                         continue;
3067                 }
3068
3069                 pnr -= page->index;
3070                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3071                                          page, sector, disk_io_size, pg_offset,
3072                                          bdev, bio, pnr,
3073                                          end_bio_extent_readpage, mirror_num,
3074                                          *bio_flags,
3075                                          this_bio_flag,
3076                                          force_bio_submit);
3077                 if (!ret) {
3078                         nr++;
3079                         *bio_flags = this_bio_flag;
3080                 } else {
3081                         SetPageError(page);
3082                         unlock_extent(tree, cur, cur + iosize - 1);
3083                         goto out;
3084                 }
3085                 cur = cur + iosize;
3086                 pg_offset += iosize;
3087         }
3088 out:
3089         if (!nr) {
3090                 if (!PageError(page))
3091                         SetPageUptodate(page);
3092                 unlock_page(page);
3093         }
3094         return ret;
3095 }
3096
3097 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3098                                              struct page *pages[], int nr_pages,
3099                                              u64 start, u64 end,
3100                                              get_extent_t *get_extent,
3101                                              struct extent_map **em_cached,
3102                                              struct bio **bio, int mirror_num,
3103                                              unsigned long *bio_flags,
3104                                              u64 *prev_em_start)
3105 {
3106         struct inode *inode;
3107         struct btrfs_ordered_extent *ordered;
3108         int index;
3109
3110         inode = pages[0]->mapping->host;
3111         while (1) {
3112                 lock_extent(tree, start, end);
3113                 ordered = btrfs_lookup_ordered_range(inode, start,
3114                                                      end - start + 1);
3115                 if (!ordered)
3116                         break;
3117                 unlock_extent(tree, start, end);
3118                 btrfs_start_ordered_extent(inode, ordered, 1);
3119                 btrfs_put_ordered_extent(ordered);
3120         }
3121
3122         for (index = 0; index < nr_pages; index++) {
3123                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3124                               mirror_num, bio_flags, 0, prev_em_start);
3125                 put_page(pages[index]);
3126         }
3127 }
3128
3129 static void __extent_readpages(struct extent_io_tree *tree,
3130                                struct page *pages[],
3131                                int nr_pages, get_extent_t *get_extent,
3132                                struct extent_map **em_cached,
3133                                struct bio **bio, int mirror_num,
3134                                unsigned long *bio_flags,
3135                                u64 *prev_em_start)
3136 {
3137         u64 start = 0;
3138         u64 end = 0;
3139         u64 page_start;
3140         int index;
3141         int first_index = 0;
3142
3143         for (index = 0; index < nr_pages; index++) {
3144                 page_start = page_offset(pages[index]);
3145                 if (!end) {
3146                         start = page_start;
3147                         end = start + PAGE_SIZE - 1;
3148                         first_index = index;
3149                 } else if (end + 1 == page_start) {
3150                         end += PAGE_SIZE;
3151                 } else {
3152                         __do_contiguous_readpages(tree, &pages[first_index],
3153                                                   index - first_index, start,
3154                                                   end, get_extent, em_cached,
3155                                                   bio, mirror_num, bio_flags,
3156                                                   prev_em_start);
3157                         start = page_start;
3158                         end = start + PAGE_SIZE - 1;
3159                         first_index = index;
3160                 }
3161         }
3162
3163         if (end)
3164                 __do_contiguous_readpages(tree, &pages[first_index],
3165                                           index - first_index, start,
3166                                           end, get_extent, em_cached, bio,
3167                                           mirror_num, bio_flags,
3168                                           prev_em_start);
3169 }
3170
3171 static int __extent_read_full_page(struct extent_io_tree *tree,
3172                                    struct page *page,
3173                                    get_extent_t *get_extent,
3174                                    struct bio **bio, int mirror_num,
3175                                    unsigned long *bio_flags, int read_flags)
3176 {
3177         struct inode *inode = page->mapping->host;
3178         struct btrfs_ordered_extent *ordered;
3179         u64 start = page_offset(page);
3180         u64 end = start + PAGE_SIZE - 1;
3181         int ret;
3182
3183         while (1) {
3184                 lock_extent(tree, start, end);
3185                 ordered = btrfs_lookup_ordered_range(inode, start,
3186                                                 PAGE_SIZE);
3187                 if (!ordered)
3188                         break;
3189                 unlock_extent(tree, start, end);
3190                 btrfs_start_ordered_extent(inode, ordered, 1);
3191                 btrfs_put_ordered_extent(ordered);
3192         }
3193
3194         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3195                             bio_flags, read_flags, NULL);
3196         return ret;
3197 }
3198
3199 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3200                             get_extent_t *get_extent, int mirror_num)
3201 {
3202         struct bio *bio = NULL;
3203         unsigned long bio_flags = 0;
3204         int ret;
3205
3206         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3207                                       &bio_flags, 0);
3208         if (bio)
3209                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3210         return ret;
3211 }
3212
3213 static void update_nr_written(struct page *page, struct writeback_control *wbc,
3214                               unsigned long nr_written)
3215 {
3216         wbc->nr_to_write -= nr_written;
3217 }
3218
3219 /*
3220  * helper for __extent_writepage, doing all of the delayed allocation setup.
3221  *
3222  * This returns 1 if our fill_delalloc function did all the work required
3223  * to write the page (copy into inline extent).  In this case the IO has
3224  * been started and the page is already unlocked.
3225  *
3226  * This returns 0 if all went well (page still locked)
3227  * This returns < 0 if there were errors (page still locked)
3228  */
3229 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3230                               struct page *page, struct writeback_control *wbc,
3231                               struct extent_page_data *epd,
3232                               u64 delalloc_start,
3233                               unsigned long *nr_written)
3234 {
3235         struct extent_io_tree *tree = epd->tree;
3236         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3237         u64 nr_delalloc;
3238         u64 delalloc_to_write = 0;
3239         u64 delalloc_end = 0;
3240         int ret;
3241         int page_started = 0;
3242
3243         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3244                 return 0;
3245
3246         while (delalloc_end < page_end) {
3247                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3248                                                page,
3249                                                &delalloc_start,
3250                                                &delalloc_end,
3251                                                BTRFS_MAX_EXTENT_SIZE);
3252                 if (nr_delalloc == 0) {
3253                         delalloc_start = delalloc_end + 1;
3254                         continue;
3255                 }
3256                 ret = tree->ops->fill_delalloc(inode, page,
3257                                                delalloc_start,
3258                                                delalloc_end,
3259                                                &page_started,
3260                                                nr_written);
3261                 /* File system has been set read-only */
3262                 if (ret) {
3263                         SetPageError(page);
3264                         /* fill_delalloc should be return < 0 for error
3265                          * but just in case, we use > 0 here meaning the
3266                          * IO is started, so we don't want to return > 0
3267                          * unless things are going well.
3268                          */
3269                         ret = ret < 0 ? ret : -EIO;
3270                         goto done;
3271                 }
3272                 /*
3273                  * delalloc_end is already one less than the total length, so
3274                  * we don't subtract one from PAGE_SIZE
3275                  */
3276                 delalloc_to_write += (delalloc_end - delalloc_start +
3277                                       PAGE_SIZE) >> PAGE_SHIFT;
3278                 delalloc_start = delalloc_end + 1;
3279         }
3280         if (wbc->nr_to_write < delalloc_to_write) {
3281                 int thresh = 8192;
3282
3283                 if (delalloc_to_write < thresh * 2)
3284                         thresh = delalloc_to_write;
3285                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3286                                          thresh);
3287         }
3288
3289         /* did the fill delalloc function already unlock and start
3290          * the IO?
3291          */
3292         if (page_started) {
3293                 /*
3294                  * we've unlocked the page, so we can't update
3295                  * the mapping's writeback index, just update
3296                  * nr_to_write.
3297                  */
3298                 wbc->nr_to_write -= *nr_written;
3299                 return 1;
3300         }
3301
3302         ret = 0;
3303
3304 done:
3305         return ret;
3306 }
3307
3308 /*
3309  * helper for __extent_writepage.  This calls the writepage start hooks,
3310  * and does the loop to map the page into extents and bios.
3311  *
3312  * We return 1 if the IO is started and the page is unlocked,
3313  * 0 if all went well (page still locked)
3314  * < 0 if there were errors (page still locked)
3315  */
3316 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3317                                  struct page *page,
3318                                  struct writeback_control *wbc,
3319                                  struct extent_page_data *epd,
3320                                  loff_t i_size,
3321                                  unsigned long nr_written,
3322                                  int write_flags, int *nr_ret)
3323 {
3324         struct extent_io_tree *tree = epd->tree;
3325         u64 start = page_offset(page);
3326         u64 page_end = start + PAGE_SIZE - 1;
3327         u64 end;
3328         u64 cur = start;
3329         u64 extent_offset;
3330         u64 block_start;
3331         u64 iosize;
3332         sector_t sector;
3333         struct extent_state *cached_state = NULL;
3334         struct extent_map *em;
3335         struct block_device *bdev;
3336         size_t pg_offset = 0;
3337         size_t blocksize;
3338         int ret = 0;
3339         int nr = 0;
3340         bool compressed;
3341
3342         if (tree->ops && tree->ops->writepage_start_hook) {
3343                 ret = tree->ops->writepage_start_hook(page, start,
3344                                                       page_end);
3345                 if (ret) {
3346                         /* Fixup worker will requeue */
3347                         if (ret == -EBUSY)
3348                                 wbc->pages_skipped++;
3349                         else
3350                                 redirty_page_for_writepage(wbc, page);
3351
3352                         update_nr_written(page, wbc, nr_written);
3353                         unlock_page(page);
3354                         ret = 1;
3355                         goto done_unlocked;
3356                 }
3357         }
3358
3359         /*
3360          * we don't want to touch the inode after unlocking the page,
3361          * so we update the mapping writeback index now
3362          */
3363         update_nr_written(page, wbc, nr_written + 1);
3364
3365         end = page_end;
3366         if (i_size <= start) {
3367                 if (tree->ops && tree->ops->writepage_end_io_hook)
3368                         tree->ops->writepage_end_io_hook(page, start,
3369                                                          page_end, NULL, 1);
3370                 goto done;
3371         }
3372
3373         blocksize = inode->i_sb->s_blocksize;
3374
3375         while (cur <= end) {
3376                 u64 em_end;
3377                 unsigned long max_nr;
3378
3379                 if (cur >= i_size) {
3380                         if (tree->ops && tree->ops->writepage_end_io_hook)
3381                                 tree->ops->writepage_end_io_hook(page, cur,
3382                                                          page_end, NULL, 1);
3383                         break;
3384                 }
3385                 em = epd->get_extent(inode, page, pg_offset, cur,
3386                                      end - cur + 1, 1);
3387                 if (IS_ERR_OR_NULL(em)) {
3388                         SetPageError(page);
3389                         ret = PTR_ERR_OR_ZERO(em);
3390                         break;
3391                 }
3392
3393                 extent_offset = cur - em->start;
3394                 em_end = extent_map_end(em);
3395                 BUG_ON(em_end <= cur);
3396                 BUG_ON(end < cur);
3397                 iosize = min(em_end - cur, end - cur + 1);
3398                 iosize = ALIGN(iosize, blocksize);
3399                 sector = (em->block_start + extent_offset) >> 9;
3400                 bdev = em->bdev;
3401                 block_start = em->block_start;
3402                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3403                 free_extent_map(em);
3404                 em = NULL;
3405
3406                 /*
3407                  * compressed and inline extents are written through other
3408                  * paths in the FS
3409                  */
3410                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3411                     block_start == EXTENT_MAP_INLINE) {
3412                         /*
3413                          * end_io notification does not happen here for
3414                          * compressed extents
3415                          */
3416                         if (!compressed && tree->ops &&
3417                             tree->ops->writepage_end_io_hook)
3418                                 tree->ops->writepage_end_io_hook(page, cur,
3419                                                          cur + iosize - 1,
3420                                                          NULL, 1);
3421                         else if (compressed) {
3422                                 /* we don't want to end_page_writeback on
3423                                  * a compressed extent.  this happens
3424                                  * elsewhere
3425                                  */
3426                                 nr++;
3427                         }
3428
3429                         cur += iosize;
3430                         pg_offset += iosize;
3431                         continue;
3432                 }
3433
3434                 max_nr = (i_size >> PAGE_SHIFT) + 1;
3435
3436                 set_range_writeback(tree, cur, cur + iosize - 1);
3437                 if (!PageWriteback(page)) {
3438                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3439                                    "page %lu not writeback, cur %llu end %llu",
3440                                page->index, cur, end);
3441                 }
3442
3443                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3444                                          page, sector, iosize, pg_offset,
3445                                          bdev, &epd->bio, max_nr,
3446                                          end_bio_extent_writepage,
3447                                          0, 0, 0, false);
3448                 if (ret)
3449                         SetPageError(page);
3450
3451                 cur = cur + iosize;
3452                 pg_offset += iosize;
3453                 nr++;
3454         }
3455 done:
3456         *nr_ret = nr;
3457
3458 done_unlocked:
3459
3460         /* drop our reference on any cached states */
3461         free_extent_state(cached_state);
3462         return ret;
3463 }
3464
3465 /*
3466  * the writepage semantics are similar to regular writepage.  extent
3467  * records are inserted to lock ranges in the tree, and as dirty areas
3468  * are found, they are marked writeback.  Then the lock bits are removed
3469  * and the end_io handler clears the writeback ranges
3470  */
3471 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3472                               void *data)
3473 {
3474         struct inode *inode = page->mapping->host;
3475         struct extent_page_data *epd = data;
3476         u64 start = page_offset(page);
3477         u64 page_end = start + PAGE_SIZE - 1;
3478         int ret;
3479         int nr = 0;
3480         size_t pg_offset = 0;
3481         loff_t i_size = i_size_read(inode);
3482         unsigned long end_index = i_size >> PAGE_SHIFT;
3483         int write_flags = 0;
3484         unsigned long nr_written = 0;
3485
3486         if (wbc->sync_mode == WB_SYNC_ALL)
3487                 write_flags = WRITE_SYNC;
3488
3489         trace___extent_writepage(page, inode, wbc);
3490
3491         WARN_ON(!PageLocked(page));
3492
3493         ClearPageError(page);
3494
3495         pg_offset = i_size & (PAGE_SIZE - 1);
3496         if (page->index > end_index ||
3497            (page->index == end_index && !pg_offset)) {
3498                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3499                 unlock_page(page);
3500                 return 0;
3501         }
3502
3503         if (page->index == end_index) {
3504                 char *userpage;
3505
3506                 userpage = kmap_atomic(page);
3507                 memset(userpage + pg_offset, 0,
3508                        PAGE_SIZE - pg_offset);
3509                 kunmap_atomic(userpage);
3510                 flush_dcache_page(page);
3511         }
3512
3513         pg_offset = 0;
3514
3515         set_page_extent_mapped(page);
3516
3517         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3518         if (ret == 1)
3519                 goto done_unlocked;
3520         if (ret)
3521                 goto done;
3522
3523         ret = __extent_writepage_io(inode, page, wbc, epd,
3524                                     i_size, nr_written, write_flags, &nr);
3525         if (ret == 1)
3526                 goto done_unlocked;
3527
3528 done:
3529         if (nr == 0) {
3530                 /* make sure the mapping tag for page dirty gets cleared */
3531                 set_page_writeback(page);
3532                 end_page_writeback(page);
3533         }
3534         if (PageError(page)) {
3535                 ret = ret < 0 ? ret : -EIO;
3536                 end_extent_writepage(page, ret, start, page_end);
3537         }
3538         unlock_page(page);
3539         return ret;
3540
3541 done_unlocked:
3542         return 0;
3543 }
3544
3545 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3546 {
3547         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3548                        TASK_UNINTERRUPTIBLE);
3549 }
3550
3551 static noinline_for_stack int
3552 lock_extent_buffer_for_io(struct extent_buffer *eb,
3553                           struct btrfs_fs_info *fs_info,
3554                           struct extent_page_data *epd)
3555 {
3556         unsigned long i, num_pages;
3557         int flush = 0;
3558         int ret = 0;
3559
3560         if (!btrfs_try_tree_write_lock(eb)) {
3561                 flush = 1;
3562                 flush_write_bio(epd);
3563                 btrfs_tree_lock(eb);
3564         }
3565
3566         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3567                 btrfs_tree_unlock(eb);
3568                 if (!epd->sync_io)
3569                         return 0;
3570                 if (!flush) {
3571                         flush_write_bio(epd);
3572                         flush = 1;
3573                 }
3574                 while (1) {
3575                         wait_on_extent_buffer_writeback(eb);
3576                         btrfs_tree_lock(eb);
3577                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3578                                 break;
3579                         btrfs_tree_unlock(eb);
3580                 }
3581         }
3582
3583         /*
3584          * We need to do this to prevent races in people who check if the eb is
3585          * under IO since we can end up having no IO bits set for a short period
3586          * of time.
3587          */
3588         spin_lock(&eb->refs_lock);
3589         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3590                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3591                 spin_unlock(&eb->refs_lock);
3592                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3593                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3594                                      -eb->len,
3595                                      fs_info->dirty_metadata_batch);
3596                 ret = 1;
3597         } else {
3598                 spin_unlock(&eb->refs_lock);
3599         }
3600
3601         btrfs_tree_unlock(eb);
3602
3603         if (!ret)
3604                 return ret;
3605
3606         num_pages = num_extent_pages(eb->start, eb->len);
3607         for (i = 0; i < num_pages; i++) {
3608                 struct page *p = eb->pages[i];
3609
3610                 if (!trylock_page(p)) {
3611                         if (!flush) {
3612                                 flush_write_bio(epd);
3613                                 flush = 1;
3614                         }
3615                         lock_page(p);
3616                 }
3617         }
3618
3619         return ret;
3620 }
3621
3622 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3623 {
3624         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3625         smp_mb__after_atomic();
3626         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3627 }
3628
3629 static void set_btree_ioerr(struct page *page)
3630 {
3631         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3632
3633         SetPageError(page);
3634         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3635                 return;
3636
3637         /*
3638          * If writeback for a btree extent that doesn't belong to a log tree
3639          * failed, increment the counter transaction->eb_write_errors.
3640          * We do this because while the transaction is running and before it's
3641          * committing (when we call filemap_fdata[write|wait]_range against
3642          * the btree inode), we might have
3643          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3644          * returns an error or an error happens during writeback, when we're
3645          * committing the transaction we wouldn't know about it, since the pages
3646          * can be no longer dirty nor marked anymore for writeback (if a
3647          * subsequent modification to the extent buffer didn't happen before the
3648          * transaction commit), which makes filemap_fdata[write|wait]_range not
3649          * able to find the pages tagged with SetPageError at transaction
3650          * commit time. So if this happens we must abort the transaction,
3651          * otherwise we commit a super block with btree roots that point to
3652          * btree nodes/leafs whose content on disk is invalid - either garbage
3653          * or the content of some node/leaf from a past generation that got
3654          * cowed or deleted and is no longer valid.
3655          *
3656          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3657          * not be enough - we need to distinguish between log tree extents vs
3658          * non-log tree extents, and the next filemap_fdatawait_range() call
3659          * will catch and clear such errors in the mapping - and that call might
3660          * be from a log sync and not from a transaction commit. Also, checking
3661          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3662          * not done and would not be reliable - the eb might have been released
3663          * from memory and reading it back again means that flag would not be
3664          * set (since it's a runtime flag, not persisted on disk).
3665          *
3666          * Using the flags below in the btree inode also makes us achieve the
3667          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3668          * writeback for all dirty pages and before filemap_fdatawait_range()
3669          * is called, the writeback for all dirty pages had already finished
3670          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3671          * filemap_fdatawait_range() would return success, as it could not know
3672          * that writeback errors happened (the pages were no longer tagged for
3673          * writeback).
3674          */
3675         switch (eb->log_index) {
3676         case -1:
3677                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3678                 break;
3679         case 0:
3680                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3681                 break;
3682         case 1:
3683                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3684                 break;
3685         default:
3686                 BUG(); /* unexpected, logic error */
3687         }
3688 }
3689
3690 static void end_bio_extent_buffer_writepage(struct bio *bio)
3691 {
3692         struct bio_vec *bvec;
3693         struct extent_buffer *eb;
3694         int i, done;
3695
3696         bio_for_each_segment_all(bvec, bio, i) {
3697                 struct page *page = bvec->bv_page;
3698
3699                 eb = (struct extent_buffer *)page->private;
3700                 BUG_ON(!eb);
3701                 done = atomic_dec_and_test(&eb->io_pages);
3702
3703                 if (bio->bi_error ||
3704                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3705                         ClearPageUptodate(page);
3706                         set_btree_ioerr(page);
3707                 }
3708
3709                 end_page_writeback(page);
3710
3711                 if (!done)
3712                         continue;
3713
3714                 end_extent_buffer_writeback(eb);
3715         }
3716
3717         bio_put(bio);
3718 }
3719
3720 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3721                         struct btrfs_fs_info *fs_info,
3722                         struct writeback_control *wbc,
3723                         struct extent_page_data *epd)
3724 {
3725         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3726         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3727         u64 offset = eb->start;
3728         u32 nritems;
3729         unsigned long i, num_pages;
3730         unsigned long bio_flags = 0;
3731         unsigned long start, end;
3732         int write_flags = (epd->sync_io ? WRITE_SYNC : 0) | REQ_META;
3733         int ret = 0;
3734
3735         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3736         num_pages = num_extent_pages(eb->start, eb->len);
3737         atomic_set(&eb->io_pages, num_pages);
3738         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3739                 bio_flags = EXTENT_BIO_TREE_LOG;
3740
3741         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3742         nritems = btrfs_header_nritems(eb);
3743         if (btrfs_header_level(eb) > 0) {
3744                 end = btrfs_node_key_ptr_offset(nritems);
3745
3746                 memset_extent_buffer(eb, 0, end, eb->len - end);
3747         } else {
3748                 /*
3749                  * leaf:
3750                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3751                  */
3752                 start = btrfs_item_nr_offset(nritems);
3753                 end = btrfs_leaf_data(eb) +
3754                       leaf_data_end(fs_info->tree_root, eb);
3755                 memset_extent_buffer(eb, 0, start, end - start);
3756         }
3757
3758         for (i = 0; i < num_pages; i++) {
3759                 struct page *p = eb->pages[i];
3760
3761                 clear_page_dirty_for_io(p);
3762                 set_page_writeback(p);
3763                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3764                                          p, offset >> 9, PAGE_SIZE, 0, bdev,
3765                                          &epd->bio, -1,
3766                                          end_bio_extent_buffer_writepage,
3767                                          0, epd->bio_flags, bio_flags, false);
3768                 epd->bio_flags = bio_flags;
3769                 if (ret) {
3770                         set_btree_ioerr(p);
3771                         end_page_writeback(p);
3772                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3773                                 end_extent_buffer_writeback(eb);
3774                         ret = -EIO;
3775                         break;
3776                 }
3777                 offset += PAGE_SIZE;
3778                 update_nr_written(p, wbc, 1);
3779                 unlock_page(p);
3780         }
3781
3782         if (unlikely(ret)) {
3783                 for (; i < num_pages; i++) {
3784                         struct page *p = eb->pages[i];
3785                         clear_page_dirty_for_io(p);
3786                         unlock_page(p);
3787                 }
3788         }
3789
3790         return ret;
3791 }
3792
3793 int btree_write_cache_pages(struct address_space *mapping,
3794                                    struct writeback_control *wbc)
3795 {
3796         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3797         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3798         struct extent_buffer *eb, *prev_eb = NULL;
3799         struct extent_page_data epd = {
3800                 .bio = NULL,
3801                 .tree = tree,
3802                 .extent_locked = 0,
3803                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3804                 .bio_flags = 0,
3805         };
3806         int ret = 0;
3807         int done = 0;
3808         int nr_to_write_done = 0;
3809         struct pagevec pvec;
3810         int nr_pages;
3811         pgoff_t index;
3812         pgoff_t end;            /* Inclusive */
3813         int scanned = 0;
3814         int tag;
3815
3816         pagevec_init(&pvec, 0);
3817         if (wbc->range_cyclic) {
3818                 index = mapping->writeback_index; /* Start from prev offset */
3819                 end = -1;
3820         } else {
3821                 index = wbc->range_start >> PAGE_SHIFT;
3822                 end = wbc->range_end >> PAGE_SHIFT;
3823                 scanned = 1;
3824         }
3825         if (wbc->sync_mode == WB_SYNC_ALL)
3826                 tag = PAGECACHE_TAG_TOWRITE;
3827         else
3828                 tag = PAGECACHE_TAG_DIRTY;
3829 retry:
3830         if (wbc->sync_mode == WB_SYNC_ALL)
3831                 tag_pages_for_writeback(mapping, index, end);
3832         while (!done && !nr_to_write_done && (index <= end) &&
3833                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3834                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3835                 unsigned i;
3836
3837                 scanned = 1;
3838                 for (i = 0; i < nr_pages; i++) {
3839                         struct page *page = pvec.pages[i];
3840
3841                         if (!PagePrivate(page))
3842                                 continue;
3843
3844                         if (!wbc->range_cyclic && page->index > end) {
3845                                 done = 1;
3846                                 break;
3847                         }
3848
3849                         spin_lock(&mapping->private_lock);
3850                         if (!PagePrivate(page)) {
3851                                 spin_unlock(&mapping->private_lock);
3852                                 continue;
3853                         }
3854
3855                         eb = (struct extent_buffer *)page->private;
3856
3857                         /*
3858                          * Shouldn't happen and normally this would be a BUG_ON
3859                          * but no sense in crashing the users box for something
3860                          * we can survive anyway.
3861                          */
3862                         if (WARN_ON(!eb)) {
3863                                 spin_unlock(&mapping->private_lock);
3864                                 continue;
3865                         }
3866
3867                         if (eb == prev_eb) {
3868                                 spin_unlock(&mapping->private_lock);
3869                                 continue;
3870                         }
3871
3872                         ret = atomic_inc_not_zero(&eb->refs);
3873                         spin_unlock(&mapping->private_lock);
3874                         if (!ret)
3875                                 continue;
3876
3877                         prev_eb = eb;
3878                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3879                         if (!ret) {
3880                                 free_extent_buffer(eb);
3881                                 continue;
3882                         }
3883
3884                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3885                         if (ret) {
3886                                 done = 1;
3887                                 free_extent_buffer(eb);
3888                                 break;
3889                         }
3890                         free_extent_buffer(eb);
3891
3892                         /*
3893                          * the filesystem may choose to bump up nr_to_write.
3894                          * We have to make sure to honor the new nr_to_write
3895                          * at any time
3896                          */
3897                         nr_to_write_done = wbc->nr_to_write <= 0;
3898                 }
3899                 pagevec_release(&pvec);
3900                 cond_resched();
3901         }
3902         if (!scanned && !done) {
3903                 /*
3904                  * We hit the last page and there is more work to be done: wrap
3905                  * back to the start of the file
3906                  */
3907                 scanned = 1;
3908                 index = 0;
3909                 goto retry;
3910         }
3911         flush_write_bio(&epd);
3912         return ret;
3913 }
3914
3915 /**
3916  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3917  * @mapping: address space structure to write
3918  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3919  * @writepage: function called for each page
3920  * @data: data passed to writepage function
3921  *
3922  * If a page is already under I/O, write_cache_pages() skips it, even
3923  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3924  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3925  * and msync() need to guarantee that all the data which was dirty at the time
3926  * the call was made get new I/O started against them.  If wbc->sync_mode is
3927  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3928  * existing IO to complete.
3929  */
3930 static int extent_write_cache_pages(struct extent_io_tree *tree,
3931                              struct address_space *mapping,
3932                              struct writeback_control *wbc,
3933                              writepage_t writepage, void *data,
3934                              void (*flush_fn)(void *))
3935 {
3936         struct inode *inode = mapping->host;
3937         int ret = 0;
3938         int done = 0;
3939         int nr_to_write_done = 0;
3940         struct pagevec pvec;
3941         int nr_pages;
3942         pgoff_t index;
3943         pgoff_t end;            /* Inclusive */
3944         pgoff_t done_index;
3945         int range_whole = 0;
3946         int scanned = 0;
3947         int tag;
3948
3949         /*
3950          * We have to hold onto the inode so that ordered extents can do their
3951          * work when the IO finishes.  The alternative to this is failing to add
3952          * an ordered extent if the igrab() fails there and that is a huge pain
3953          * to deal with, so instead just hold onto the inode throughout the
3954          * writepages operation.  If it fails here we are freeing up the inode
3955          * anyway and we'd rather not waste our time writing out stuff that is
3956          * going to be truncated anyway.
3957          */
3958         if (!igrab(inode))
3959                 return 0;
3960
3961         pagevec_init(&pvec, 0);
3962         if (wbc->range_cyclic) {
3963                 index = mapping->writeback_index; /* Start from prev offset */
3964                 end = -1;
3965         } else {
3966                 index = wbc->range_start >> PAGE_SHIFT;
3967                 end = wbc->range_end >> PAGE_SHIFT;
3968                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3969                         range_whole = 1;
3970                 scanned = 1;
3971         }
3972         if (wbc->sync_mode == WB_SYNC_ALL)
3973                 tag = PAGECACHE_TAG_TOWRITE;
3974         else
3975                 tag = PAGECACHE_TAG_DIRTY;
3976 retry:
3977         if (wbc->sync_mode == WB_SYNC_ALL)
3978                 tag_pages_for_writeback(mapping, index, end);
3979         done_index = index;
3980         while (!done && !nr_to_write_done && (index <= end) &&
3981                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3982                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3983                 unsigned i;
3984
3985                 scanned = 1;
3986                 for (i = 0; i < nr_pages; i++) {
3987                         struct page *page = pvec.pages[i];
3988
3989                         done_index = page->index;
3990                         /*
3991                          * At this point we hold neither mapping->tree_lock nor
3992                          * lock on the page itself: the page may be truncated or
3993                          * invalidated (changing page->mapping to NULL), or even
3994                          * swizzled back from swapper_space to tmpfs file
3995                          * mapping
3996                          */
3997                         if (!trylock_page(page)) {
3998                                 flush_fn(data);
3999                                 lock_page(page);
4000                         }
4001
4002                         if (unlikely(page->mapping != mapping)) {
4003                                 unlock_page(page);
4004                                 continue;
4005                         }
4006
4007                         if (!wbc->range_cyclic && page->index > end) {
4008                                 done = 1;
4009                                 unlock_page(page);
4010                                 continue;
4011                         }
4012
4013                         if (wbc->sync_mode != WB_SYNC_NONE) {
4014                                 if (PageWriteback(page))
4015                                         flush_fn(data);
4016                                 wait_on_page_writeback(page);
4017                         }
4018
4019                         if (PageWriteback(page) ||
4020                             !clear_page_dirty_for_io(page)) {
4021                                 unlock_page(page);
4022                                 continue;
4023                         }
4024
4025                         ret = (*writepage)(page, wbc, data);
4026
4027                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4028                                 unlock_page(page);
4029                                 ret = 0;
4030                         }
4031                         if (ret < 0) {
4032                                 /*
4033                                  * done_index is set past this page,
4034                                  * so media errors will not choke
4035                                  * background writeout for the entire
4036                                  * file. This has consequences for
4037                                  * range_cyclic semantics (ie. it may
4038                                  * not be suitable for data integrity
4039                                  * writeout).
4040                                  */
4041                                 done_index = page->index + 1;
4042                                 done = 1;
4043                                 break;
4044                         }
4045
4046                         /*
4047                          * the filesystem may choose to bump up nr_to_write.
4048                          * We have to make sure to honor the new nr_to_write
4049                          * at any time
4050                          */
4051                         nr_to_write_done = wbc->nr_to_write <= 0;
4052                 }
4053                 pagevec_release(&pvec);
4054                 cond_resched();
4055         }
4056         if (!scanned && !done) {
4057                 /*
4058                  * We hit the last page and there is more work to be done: wrap
4059                  * back to the start of the file
4060                  */
4061                 scanned = 1;
4062                 index = 0;
4063                 goto retry;
4064         }
4065
4066         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4067                 mapping->writeback_index = done_index;
4068
4069         btrfs_add_delayed_iput(inode);
4070         return ret;
4071 }
4072
4073 static void flush_epd_write_bio(struct extent_page_data *epd)
4074 {
4075         if (epd->bio) {
4076                 int ret;
4077
4078                 bio_set_op_attrs(epd->bio, REQ_OP_WRITE,
4079                                  epd->sync_io ? WRITE_SYNC : 0);
4080
4081                 ret = submit_one_bio(epd->bio, 0, epd->bio_flags);
4082                 BUG_ON(ret < 0); /* -ENOMEM */
4083                 epd->bio = NULL;
4084         }
4085 }
4086
4087 static noinline void flush_write_bio(void *data)
4088 {
4089         struct extent_page_data *epd = data;
4090         flush_epd_write_bio(epd);
4091 }
4092
4093 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4094                           get_extent_t *get_extent,
4095                           struct writeback_control *wbc)
4096 {
4097         int ret;
4098         struct extent_page_data epd = {
4099                 .bio = NULL,
4100                 .tree = tree,
4101                 .get_extent = get_extent,
4102                 .extent_locked = 0,
4103                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4104                 .bio_flags = 0,
4105         };
4106
4107         ret = __extent_writepage(page, wbc, &epd);
4108
4109         flush_epd_write_bio(&epd);
4110         return ret;
4111 }
4112
4113 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4114                               u64 start, u64 end, get_extent_t *get_extent,
4115                               int mode)
4116 {
4117         int ret = 0;
4118         struct address_space *mapping = inode->i_mapping;
4119         struct page *page;
4120         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4121                 PAGE_SHIFT;
4122
4123         struct extent_page_data epd = {
4124                 .bio = NULL,
4125                 .tree = tree,
4126                 .get_extent = get_extent,
4127                 .extent_locked = 1,
4128                 .sync_io = mode == WB_SYNC_ALL,
4129                 .bio_flags = 0,
4130         };
4131         struct writeback_control wbc_writepages = {
4132                 .sync_mode      = mode,
4133                 .nr_to_write    = nr_pages * 2,
4134                 .range_start    = start,
4135                 .range_end      = end + 1,
4136         };
4137
4138         while (start <= end) {
4139                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4140                 if (clear_page_dirty_for_io(page))
4141                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4142                 else {
4143                         if (tree->ops && tree->ops->writepage_end_io_hook)
4144                                 tree->ops->writepage_end_io_hook(page, start,
4145                                                  start + PAGE_SIZE - 1,
4146                                                  NULL, 1);
4147                         unlock_page(page);
4148                 }
4149                 put_page(page);
4150                 start += PAGE_SIZE;
4151         }
4152
4153         flush_epd_write_bio(&epd);
4154         return ret;
4155 }
4156
4157 int extent_writepages(struct extent_io_tree *tree,
4158                       struct address_space *mapping,
4159                       get_extent_t *get_extent,
4160                       struct writeback_control *wbc)
4161 {
4162         int ret = 0;
4163         struct extent_page_data epd = {
4164                 .bio = NULL,
4165                 .tree = tree,
4166                 .get_extent = get_extent,
4167                 .extent_locked = 0,
4168                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4169                 .bio_flags = 0,
4170         };
4171
4172         ret = extent_write_cache_pages(tree, mapping, wbc,
4173                                        __extent_writepage, &epd,
4174                                        flush_write_bio);
4175         flush_epd_write_bio(&epd);
4176         return ret;
4177 }
4178
4179 int extent_readpages(struct extent_io_tree *tree,
4180                      struct address_space *mapping,
4181                      struct list_head *pages, unsigned nr_pages,
4182                      get_extent_t get_extent)
4183 {
4184         struct bio *bio = NULL;
4185         unsigned page_idx;
4186         unsigned long bio_flags = 0;
4187         struct page *pagepool[16];
4188         struct page *page;
4189         struct extent_map *em_cached = NULL;
4190         int nr = 0;
4191         u64 prev_em_start = (u64)-1;
4192
4193         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4194                 page = list_entry(pages->prev, struct page, lru);
4195
4196                 prefetchw(&page->flags);
4197                 list_del(&page->lru);
4198                 if (add_to_page_cache_lru(page, mapping,
4199                                         page->index,
4200                                         readahead_gfp_mask(mapping))) {
4201                         put_page(page);
4202                         continue;
4203                 }
4204
4205                 pagepool[nr++] = page;
4206                 if (nr < ARRAY_SIZE(pagepool))
4207                         continue;
4208                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4209                                    &bio, 0, &bio_flags, &prev_em_start);
4210                 nr = 0;
4211         }
4212         if (nr)
4213                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4214                                    &bio, 0, &bio_flags, &prev_em_start);
4215
4216         if (em_cached)
4217                 free_extent_map(em_cached);
4218
4219         BUG_ON(!list_empty(pages));
4220         if (bio)
4221                 return submit_one_bio(bio, 0, bio_flags);
4222         return 0;
4223 }
4224
4225 /*
4226  * basic invalidatepage code, this waits on any locked or writeback
4227  * ranges corresponding to the page, and then deletes any extent state
4228  * records from the tree
4229  */
4230 int extent_invalidatepage(struct extent_io_tree *tree,
4231                           struct page *page, unsigned long offset)
4232 {
4233         struct extent_state *cached_state = NULL;
4234         u64 start = page_offset(page);
4235         u64 end = start + PAGE_SIZE - 1;
4236         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4237
4238         start += ALIGN(offset, blocksize);
4239         if (start > end)
4240                 return 0;
4241
4242         lock_extent_bits(tree, start, end, &cached_state);
4243         wait_on_page_writeback(page);
4244         clear_extent_bit(tree, start, end,
4245                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4246                          EXTENT_DO_ACCOUNTING,
4247                          1, 1, &cached_state, GFP_NOFS);
4248         return 0;
4249 }
4250
4251 /*
4252  * a helper for releasepage, this tests for areas of the page that
4253  * are locked or under IO and drops the related state bits if it is safe
4254  * to drop the page.
4255  */
4256 static int try_release_extent_state(struct extent_map_tree *map,
4257                                     struct extent_io_tree *tree,
4258                                     struct page *page, gfp_t mask)
4259 {
4260         u64 start = page_offset(page);
4261         u64 end = start + PAGE_SIZE - 1;
4262         int ret = 1;
4263
4264         if (test_range_bit(tree, start, end,
4265                            EXTENT_IOBITS, 0, NULL))
4266                 ret = 0;
4267         else {
4268                 if ((mask & GFP_NOFS) == GFP_NOFS)
4269                         mask = GFP_NOFS;
4270                 /*
4271                  * at this point we can safely clear everything except the
4272                  * locked bit and the nodatasum bit
4273                  */
4274                 ret = clear_extent_bit(tree, start, end,
4275                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4276                                  0, 0, NULL, mask);
4277
4278                 /* if clear_extent_bit failed for enomem reasons,
4279                  * we can't allow the release to continue.
4280                  */
4281                 if (ret < 0)
4282                         ret = 0;
4283                 else
4284                         ret = 1;
4285         }
4286         return ret;
4287 }
4288
4289 /*
4290  * a helper for releasepage.  As long as there are no locked extents
4291  * in the range corresponding to the page, both state records and extent
4292  * map records are removed
4293  */
4294 int try_release_extent_mapping(struct extent_map_tree *map,
4295                                struct extent_io_tree *tree, struct page *page,
4296                                gfp_t mask)
4297 {
4298         struct extent_map *em;
4299         u64 start = page_offset(page);
4300         u64 end = start + PAGE_SIZE - 1;
4301
4302         if (gfpflags_allow_blocking(mask) &&
4303             page->mapping->host->i_size > SZ_16M) {
4304                 u64 len;
4305                 while (start <= end) {
4306                         len = end - start + 1;
4307                         write_lock(&map->lock);
4308                         em = lookup_extent_mapping(map, start, len);
4309                         if (!em) {
4310                                 write_unlock(&map->lock);
4311                                 break;
4312                         }
4313                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4314                             em->start != start) {
4315                                 write_unlock(&map->lock);
4316                                 free_extent_map(em);
4317                                 break;
4318                         }
4319                         if (!test_range_bit(tree, em->start,
4320                                             extent_map_end(em) - 1,
4321                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4322                                             0, NULL)) {
4323                                 remove_extent_mapping(map, em);
4324                                 /* once for the rb tree */
4325                                 free_extent_map(em);
4326                         }
4327                         start = extent_map_end(em);
4328                         write_unlock(&map->lock);
4329
4330                         /* once for us */
4331                         free_extent_map(em);
4332                 }
4333         }
4334         return try_release_extent_state(map, tree, page, mask);
4335 }
4336
4337 /*
4338  * helper function for fiemap, which doesn't want to see any holes.
4339  * This maps until we find something past 'last'
4340  */
4341 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4342                                                 u64 offset,
4343                                                 u64 last,
4344                                                 get_extent_t *get_extent)
4345 {
4346         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4347         struct extent_map *em;
4348         u64 len;
4349
4350         if (offset >= last)
4351                 return NULL;
4352
4353         while (1) {
4354                 len = last - offset;
4355                 if (len == 0)
4356                         break;
4357                 len = ALIGN(len, sectorsize);
4358                 em = get_extent(inode, NULL, 0, offset, len, 0);
4359                 if (IS_ERR_OR_NULL(em))
4360                         return em;
4361
4362                 /* if this isn't a hole return it */
4363                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4364                     em->block_start != EXTENT_MAP_HOLE) {
4365                         return em;
4366                 }
4367
4368                 /* this is a hole, advance to the next extent */
4369                 offset = extent_map_end(em);
4370                 free_extent_map(em);
4371                 if (offset >= last)
4372                         break;
4373         }
4374         return NULL;
4375 }
4376
4377 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4378                 __u64 start, __u64 len, get_extent_t *get_extent)
4379 {
4380         int ret = 0;
4381         u64 off = start;
4382         u64 max = start + len;
4383         u32 flags = 0;
4384         u32 found_type;
4385         u64 last;
4386         u64 last_for_get_extent = 0;
4387         u64 disko = 0;
4388         u64 isize = i_size_read(inode);
4389         struct btrfs_key found_key;
4390         struct extent_map *em = NULL;
4391         struct extent_state *cached_state = NULL;
4392         struct btrfs_path *path;
4393         struct btrfs_root *root = BTRFS_I(inode)->root;
4394         int end = 0;
4395         u64 em_start = 0;
4396         u64 em_len = 0;
4397         u64 em_end = 0;
4398
4399         if (len == 0)
4400                 return -EINVAL;
4401
4402         path = btrfs_alloc_path();
4403         if (!path)
4404                 return -ENOMEM;
4405         path->leave_spinning = 1;
4406
4407         start = round_down(start, BTRFS_I(inode)->root->sectorsize);
4408         len = round_up(max, BTRFS_I(inode)->root->sectorsize) - start;
4409
4410         /*
4411          * lookup the last file extent.  We're not using i_size here
4412          * because there might be preallocation past i_size
4413          */
4414         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
4415                                        0);
4416         if (ret < 0) {
4417                 btrfs_free_path(path);
4418                 return ret;
4419         } else {
4420                 WARN_ON(!ret);
4421                 if (ret == 1)
4422                         ret = 0;
4423         }
4424
4425         path->slots[0]--;
4426         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4427         found_type = found_key.type;
4428
4429         /* No extents, but there might be delalloc bits */
4430         if (found_key.objectid != btrfs_ino(inode) ||
4431             found_type != BTRFS_EXTENT_DATA_KEY) {
4432                 /* have to trust i_size as the end */
4433                 last = (u64)-1;
4434                 last_for_get_extent = isize;
4435         } else {
4436                 /*
4437                  * remember the start of the last extent.  There are a
4438                  * bunch of different factors that go into the length of the
4439                  * extent, so its much less complex to remember where it started
4440                  */
4441                 last = found_key.offset;
4442                 last_for_get_extent = last + 1;
4443         }
4444         btrfs_release_path(path);
4445
4446         /*
4447          * we might have some extents allocated but more delalloc past those
4448          * extents.  so, we trust isize unless the start of the last extent is
4449          * beyond isize
4450          */
4451         if (last < isize) {
4452                 last = (u64)-1;
4453                 last_for_get_extent = isize;
4454         }
4455
4456         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4457                          &cached_state);
4458
4459         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4460                                    get_extent);
4461         if (!em)
4462                 goto out;
4463         if (IS_ERR(em)) {
4464                 ret = PTR_ERR(em);
4465                 goto out;
4466         }
4467
4468         while (!end) {
4469                 u64 offset_in_extent = 0;
4470
4471                 /* break if the extent we found is outside the range */
4472                 if (em->start >= max || extent_map_end(em) < off)
4473                         break;
4474
4475                 /*
4476                  * get_extent may return an extent that starts before our
4477                  * requested range.  We have to make sure the ranges
4478                  * we return to fiemap always move forward and don't
4479                  * overlap, so adjust the offsets here
4480                  */
4481                 em_start = max(em->start, off);
4482
4483                 /*
4484                  * record the offset from the start of the extent
4485                  * for adjusting the disk offset below.  Only do this if the
4486                  * extent isn't compressed since our in ram offset may be past
4487                  * what we have actually allocated on disk.
4488                  */
4489                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4490                         offset_in_extent = em_start - em->start;
4491                 em_end = extent_map_end(em);
4492                 em_len = em_end - em_start;
4493                 disko = 0;
4494                 flags = 0;
4495
4496                 /*
4497                  * bump off for our next call to get_extent
4498                  */
4499                 off = extent_map_end(em);
4500                 if (off >= max)
4501                         end = 1;
4502
4503                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4504                         end = 1;
4505                         flags |= FIEMAP_EXTENT_LAST;
4506                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4507                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4508                                   FIEMAP_EXTENT_NOT_ALIGNED);
4509                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4510                         flags |= (FIEMAP_EXTENT_DELALLOC |
4511                                   FIEMAP_EXTENT_UNKNOWN);
4512                 } else if (fieinfo->fi_extents_max) {
4513                         struct btrfs_trans_handle *trans;
4514
4515                         u64 bytenr = em->block_start -
4516                                 (em->start - em->orig_start);
4517
4518                         disko = em->block_start + offset_in_extent;
4519
4520                         /*
4521                          * We need a trans handle to get delayed refs
4522                          */
4523                         trans = btrfs_join_transaction(root);
4524                         /*
4525                          * It's OK if we can't start a trans we can still check
4526                          * from commit_root
4527                          */
4528                         if (IS_ERR(trans))
4529                                 trans = NULL;
4530
4531                         /*
4532                          * As btrfs supports shared space, this information
4533                          * can be exported to userspace tools via
4534                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4535                          * then we're just getting a count and we can skip the
4536                          * lookup stuff.
4537                          */
4538                         ret = btrfs_check_shared(trans, root->fs_info,
4539                                                  root->objectid,
4540                                                  btrfs_ino(inode), bytenr);
4541                         if (trans)
4542                                 btrfs_end_transaction(trans, root);
4543                         if (ret < 0)
4544                                 goto out_free;
4545                         if (ret)
4546                                 flags |= FIEMAP_EXTENT_SHARED;
4547                         ret = 0;
4548                 }
4549                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4550                         flags |= FIEMAP_EXTENT_ENCODED;
4551                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4552                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4553
4554                 free_extent_map(em);
4555                 em = NULL;
4556                 if ((em_start >= last) || em_len == (u64)-1 ||
4557                    (last == (u64)-1 && isize <= em_end)) {
4558                         flags |= FIEMAP_EXTENT_LAST;
4559                         end = 1;
4560                 }
4561
4562                 /* now scan forward to see if this is really the last extent. */
4563                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4564                                            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 = fiemap_fill_next_extent(fieinfo, 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         free_extent_map(em);
4583 out:
4584         btrfs_free_path(path);
4585         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4586                              &cached_state, GFP_NOFS);
4587         return ret;
4588 }
4589
4590 static void __free_extent_buffer(struct extent_buffer *eb)
4591 {
4592         btrfs_leak_debug_del(&eb->leak_list);
4593         kmem_cache_free(extent_buffer_cache, eb);
4594 }
4595
4596 int extent_buffer_under_io(struct extent_buffer *eb)
4597 {
4598         return (atomic_read(&eb->io_pages) ||
4599                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4600                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4601 }
4602
4603 /*
4604  * Helper for releasing extent buffer page.
4605  */
4606 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4607 {
4608         unsigned long index;
4609         struct page *page;
4610         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4611
4612         BUG_ON(extent_buffer_under_io(eb));
4613
4614         index = num_extent_pages(eb->start, eb->len);
4615         if (index == 0)
4616                 return;
4617
4618         do {
4619                 index--;
4620                 page = eb->pages[index];
4621                 if (!page)
4622                         continue;
4623                 if (mapped)
4624                         spin_lock(&page->mapping->private_lock);
4625                 /*
4626                  * We do this since we'll remove the pages after we've
4627                  * removed the eb from the radix tree, so we could race
4628                  * and have this page now attached to the new eb.  So
4629                  * only clear page_private if it's still connected to
4630                  * this eb.
4631                  */
4632                 if (PagePrivate(page) &&
4633                     page->private == (unsigned long)eb) {
4634                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4635                         BUG_ON(PageDirty(page));
4636                         BUG_ON(PageWriteback(page));
4637                         /*
4638                          * We need to make sure we haven't be attached
4639                          * to a new eb.
4640                          */
4641                         ClearPagePrivate(page);
4642                         set_page_private(page, 0);
4643                         /* One for the page private */
4644                         put_page(page);
4645                 }
4646
4647                 if (mapped)
4648                         spin_unlock(&page->mapping->private_lock);
4649
4650                 /* One for when we allocated the page */
4651                 put_page(page);
4652         } while (index != 0);
4653 }
4654
4655 /*
4656  * Helper for releasing the extent buffer.
4657  */
4658 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4659 {
4660         btrfs_release_extent_buffer_page(eb);
4661         __free_extent_buffer(eb);
4662 }
4663
4664 static struct extent_buffer *
4665 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4666                       unsigned long len)
4667 {
4668         struct extent_buffer *eb = NULL;
4669
4670         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4671         eb->start = start;
4672         eb->len = len;
4673         eb->fs_info = fs_info;
4674         eb->bflags = 0;
4675         rwlock_init(&eb->lock);
4676         atomic_set(&eb->write_locks, 0);
4677         atomic_set(&eb->read_locks, 0);
4678         atomic_set(&eb->blocking_readers, 0);
4679         atomic_set(&eb->blocking_writers, 0);
4680         atomic_set(&eb->spinning_readers, 0);
4681         atomic_set(&eb->spinning_writers, 0);
4682         eb->lock_nested = 0;
4683         init_waitqueue_head(&eb->write_lock_wq);
4684         init_waitqueue_head(&eb->read_lock_wq);
4685
4686         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4687
4688         spin_lock_init(&eb->refs_lock);
4689         atomic_set(&eb->refs, 1);
4690         atomic_set(&eb->io_pages, 0);
4691
4692         /*
4693          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4694          */
4695         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4696                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4697         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4698
4699         return eb;
4700 }
4701
4702 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4703 {
4704         unsigned long i;
4705         struct page *p;
4706         struct extent_buffer *new;
4707         unsigned long num_pages = num_extent_pages(src->start, src->len);
4708
4709         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4710         if (new == NULL)
4711                 return NULL;
4712
4713         for (i = 0; i < num_pages; i++) {
4714                 p = alloc_page(GFP_NOFS);
4715                 if (!p) {
4716                         btrfs_release_extent_buffer(new);
4717                         return NULL;
4718                 }
4719                 attach_extent_buffer_page(new, p);
4720                 WARN_ON(PageDirty(p));
4721                 SetPageUptodate(p);
4722                 new->pages[i] = p;
4723         }
4724
4725         copy_extent_buffer(new, src, 0, 0, src->len);
4726         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4727         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4728
4729         return new;
4730 }
4731
4732 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4733                                                   u64 start, unsigned long len)
4734 {
4735         struct extent_buffer *eb;
4736         unsigned long num_pages;
4737         unsigned long i;
4738
4739         num_pages = num_extent_pages(start, len);
4740
4741         eb = __alloc_extent_buffer(fs_info, start, len);
4742         if (!eb)
4743                 return NULL;
4744
4745         for (i = 0; i < num_pages; i++) {
4746                 eb->pages[i] = alloc_page(GFP_NOFS);
4747                 if (!eb->pages[i])
4748                         goto err;
4749         }
4750         set_extent_buffer_uptodate(eb);
4751         btrfs_set_header_nritems(eb, 0);
4752         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4753
4754         return eb;
4755 err:
4756         for (; i > 0; i--)
4757                 __free_page(eb->pages[i - 1]);
4758         __free_extent_buffer(eb);
4759         return NULL;
4760 }
4761
4762 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4763                                                 u64 start, u32 nodesize)
4764 {
4765         unsigned long len;
4766
4767         if (!fs_info) {
4768                 /*
4769                  * Called only from tests that don't always have a fs_info
4770                  * available
4771                  */
4772                 len = nodesize;
4773         } else {
4774                 len = fs_info->tree_root->nodesize;
4775         }
4776
4777         return __alloc_dummy_extent_buffer(fs_info, start, len);
4778 }
4779
4780 static void check_buffer_tree_ref(struct extent_buffer *eb)
4781 {
4782         int refs;
4783         /* the ref bit is tricky.  We have to make sure it is set
4784          * if we have the buffer dirty.   Otherwise the
4785          * code to free a buffer can end up dropping a dirty
4786          * page
4787          *
4788          * Once the ref bit is set, it won't go away while the
4789          * buffer is dirty or in writeback, and it also won't
4790          * go away while we have the reference count on the
4791          * eb bumped.
4792          *
4793          * We can't just set the ref bit without bumping the
4794          * ref on the eb because free_extent_buffer might
4795          * see the ref bit and try to clear it.  If this happens
4796          * free_extent_buffer might end up dropping our original
4797          * ref by mistake and freeing the page before we are able
4798          * to add one more ref.
4799          *
4800          * So bump the ref count first, then set the bit.  If someone
4801          * beat us to it, drop the ref we added.
4802          */
4803         refs = atomic_read(&eb->refs);
4804         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4805                 return;
4806
4807         spin_lock(&eb->refs_lock);
4808         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4809                 atomic_inc(&eb->refs);
4810         spin_unlock(&eb->refs_lock);
4811 }
4812
4813 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4814                 struct page *accessed)
4815 {
4816         unsigned long num_pages, i;
4817
4818         check_buffer_tree_ref(eb);
4819
4820         num_pages = num_extent_pages(eb->start, eb->len);
4821         for (i = 0; i < num_pages; i++) {
4822                 struct page *p = eb->pages[i];
4823
4824                 if (p != accessed)
4825                         mark_page_accessed(p);
4826         }
4827 }
4828
4829 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4830                                          u64 start)
4831 {
4832         struct extent_buffer *eb;
4833
4834         rcu_read_lock();
4835         eb = radix_tree_lookup(&fs_info->buffer_radix,
4836                                start >> PAGE_SHIFT);
4837         if (eb && atomic_inc_not_zero(&eb->refs)) {
4838                 rcu_read_unlock();
4839                 /*
4840                  * Lock our eb's refs_lock to avoid races with
4841                  * free_extent_buffer. When we get our eb it might be flagged
4842                  * with EXTENT_BUFFER_STALE and another task running
4843                  * free_extent_buffer might have seen that flag set,
4844                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4845                  * writeback flags not set) and it's still in the tree (flag
4846                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4847                  * of decrementing the extent buffer's reference count twice.
4848                  * So here we could race and increment the eb's reference count,
4849                  * clear its stale flag, mark it as dirty and drop our reference
4850                  * before the other task finishes executing free_extent_buffer,
4851                  * which would later result in an attempt to free an extent
4852                  * buffer that is dirty.
4853                  */
4854                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4855                         spin_lock(&eb->refs_lock);
4856                         spin_unlock(&eb->refs_lock);
4857                 }
4858                 mark_extent_buffer_accessed(eb, NULL);
4859                 return eb;
4860         }
4861         rcu_read_unlock();
4862
4863         return NULL;
4864 }
4865
4866 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4867 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4868                                         u64 start, u32 nodesize)
4869 {
4870         struct extent_buffer *eb, *exists = NULL;
4871         int ret;
4872
4873         eb = find_extent_buffer(fs_info, start);
4874         if (eb)
4875                 return eb;
4876         eb = alloc_dummy_extent_buffer(fs_info, start, nodesize);
4877         if (!eb)
4878                 return NULL;
4879         eb->fs_info = fs_info;
4880 again:
4881         ret = radix_tree_preload(GFP_NOFS);
4882         if (ret)
4883                 goto free_eb;
4884         spin_lock(&fs_info->buffer_lock);
4885         ret = radix_tree_insert(&fs_info->buffer_radix,
4886                                 start >> PAGE_SHIFT, eb);
4887         spin_unlock(&fs_info->buffer_lock);
4888         radix_tree_preload_end();
4889         if (ret == -EEXIST) {
4890                 exists = find_extent_buffer(fs_info, start);
4891                 if (exists)
4892                         goto free_eb;
4893                 else
4894                         goto again;
4895         }
4896         check_buffer_tree_ref(eb);
4897         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4898
4899         /*
4900          * We will free dummy extent buffer's if they come into
4901          * free_extent_buffer with a ref count of 2, but if we are using this we
4902          * want the buffers to stay in memory until we're done with them, so
4903          * bump the ref count again.
4904          */
4905         atomic_inc(&eb->refs);
4906         return eb;
4907 free_eb:
4908         btrfs_release_extent_buffer(eb);
4909         return exists;
4910 }
4911 #endif
4912
4913 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4914                                           u64 start)
4915 {
4916         unsigned long len = fs_info->tree_root->nodesize;
4917         unsigned long num_pages = num_extent_pages(start, len);
4918         unsigned long i;
4919         unsigned long index = start >> PAGE_SHIFT;
4920         struct extent_buffer *eb;
4921         struct extent_buffer *exists = NULL;
4922         struct page *p;
4923         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4924         int uptodate = 1;
4925         int ret;
4926
4927         if (!IS_ALIGNED(start, fs_info->tree_root->sectorsize)) {
4928                 btrfs_err(fs_info, "bad tree block start %llu", start);
4929                 return ERR_PTR(-EINVAL);
4930         }
4931
4932         eb = find_extent_buffer(fs_info, start);
4933         if (eb)
4934                 return eb;
4935
4936         eb = __alloc_extent_buffer(fs_info, start, len);
4937         if (!eb)
4938                 return ERR_PTR(-ENOMEM);
4939
4940         for (i = 0; i < num_pages; i++, index++) {
4941                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4942                 if (!p) {
4943                         exists = ERR_PTR(-ENOMEM);
4944                         goto free_eb;
4945                 }
4946
4947                 spin_lock(&mapping->private_lock);
4948                 if (PagePrivate(p)) {
4949                         /*
4950                          * We could have already allocated an eb for this page
4951                          * and attached one so lets see if we can get a ref on
4952                          * the existing eb, and if we can we know it's good and
4953                          * we can just return that one, else we know we can just
4954                          * overwrite page->private.
4955                          */
4956                         exists = (struct extent_buffer *)p->private;
4957                         if (atomic_inc_not_zero(&exists->refs)) {
4958                                 spin_unlock(&mapping->private_lock);
4959                                 unlock_page(p);
4960                                 put_page(p);
4961                                 mark_extent_buffer_accessed(exists, p);
4962                                 goto free_eb;
4963                         }
4964                         exists = NULL;
4965
4966                         /*
4967                          * Do this so attach doesn't complain and we need to
4968                          * drop the ref the old guy had.
4969                          */
4970                         ClearPagePrivate(p);
4971                         WARN_ON(PageDirty(p));
4972                         put_page(p);
4973                 }
4974                 attach_extent_buffer_page(eb, p);
4975                 spin_unlock(&mapping->private_lock);
4976                 WARN_ON(PageDirty(p));
4977                 eb->pages[i] = p;
4978                 if (!PageUptodate(p))
4979                         uptodate = 0;
4980
4981                 /*
4982                  * see below about how we avoid a nasty race with release page
4983                  * and why we unlock later
4984                  */
4985         }
4986         if (uptodate)
4987                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4988 again:
4989         ret = radix_tree_preload(GFP_NOFS);
4990         if (ret) {
4991                 exists = ERR_PTR(ret);
4992                 goto free_eb;
4993         }
4994
4995         spin_lock(&fs_info->buffer_lock);
4996         ret = radix_tree_insert(&fs_info->buffer_radix,
4997                                 start >> PAGE_SHIFT, eb);
4998         spin_unlock(&fs_info->buffer_lock);
4999         radix_tree_preload_end();
5000         if (ret == -EEXIST) {
5001                 exists = find_extent_buffer(fs_info, start);
5002                 if (exists)
5003                         goto free_eb;
5004                 else
5005                         goto again;
5006         }
5007         /* add one reference for the tree */
5008         check_buffer_tree_ref(eb);
5009         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5010
5011         /*
5012          * there is a race where release page may have
5013          * tried to find this extent buffer in the radix
5014          * but failed.  It will tell the VM it is safe to
5015          * reclaim the, and it will clear the page private bit.
5016          * We must make sure to set the page private bit properly
5017          * after the extent buffer is in the radix tree so
5018          * it doesn't get lost
5019          */
5020         SetPageChecked(eb->pages[0]);
5021         for (i = 1; i < num_pages; i++) {
5022                 p = eb->pages[i];
5023                 ClearPageChecked(p);
5024                 unlock_page(p);
5025         }
5026         unlock_page(eb->pages[0]);
5027         return eb;
5028
5029 free_eb:
5030         WARN_ON(!atomic_dec_and_test(&eb->refs));
5031         for (i = 0; i < num_pages; i++) {
5032                 if (eb->pages[i])
5033                         unlock_page(eb->pages[i]);
5034         }
5035
5036         btrfs_release_extent_buffer(eb);
5037         return exists;
5038 }
5039
5040 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5041 {
5042         struct extent_buffer *eb =
5043                         container_of(head, struct extent_buffer, rcu_head);
5044
5045         __free_extent_buffer(eb);
5046 }
5047
5048 /* Expects to have eb->eb_lock already held */
5049 static int release_extent_buffer(struct extent_buffer *eb)
5050 {
5051         WARN_ON(atomic_read(&eb->refs) == 0);
5052         if (atomic_dec_and_test(&eb->refs)) {
5053                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5054                         struct btrfs_fs_info *fs_info = eb->fs_info;
5055
5056                         spin_unlock(&eb->refs_lock);
5057
5058                         spin_lock(&fs_info->buffer_lock);
5059                         radix_tree_delete(&fs_info->buffer_radix,
5060                                           eb->start >> PAGE_SHIFT);
5061                         spin_unlock(&fs_info->buffer_lock);
5062                 } else {
5063                         spin_unlock(&eb->refs_lock);
5064                 }
5065
5066                 /* Should be safe to release our pages at this point */
5067                 btrfs_release_extent_buffer_page(eb);
5068 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5069                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5070                         __free_extent_buffer(eb);
5071                         return 1;
5072                 }
5073 #endif
5074                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5075                 return 1;
5076         }
5077         spin_unlock(&eb->refs_lock);
5078
5079         return 0;
5080 }
5081
5082 void free_extent_buffer(struct extent_buffer *eb)
5083 {
5084         int refs;
5085         int old;
5086         if (!eb)
5087                 return;
5088
5089         while (1) {
5090                 refs = atomic_read(&eb->refs);
5091                 if (refs <= 3)
5092                         break;
5093                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5094                 if (old == refs)
5095                         return;
5096         }
5097
5098         spin_lock(&eb->refs_lock);
5099         if (atomic_read(&eb->refs) == 2 &&
5100             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5101                 atomic_dec(&eb->refs);
5102
5103         if (atomic_read(&eb->refs) == 2 &&
5104             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5105             !extent_buffer_under_io(eb) &&
5106             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5107                 atomic_dec(&eb->refs);
5108
5109         /*
5110          * I know this is terrible, but it's temporary until we stop tracking
5111          * the uptodate bits and such for the extent buffers.
5112          */
5113         release_extent_buffer(eb);
5114 }
5115
5116 void free_extent_buffer_stale(struct extent_buffer *eb)
5117 {
5118         if (!eb)
5119                 return;
5120
5121         spin_lock(&eb->refs_lock);
5122         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5123
5124         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5125             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5126                 atomic_dec(&eb->refs);
5127         release_extent_buffer(eb);
5128 }
5129
5130 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5131 {
5132         unsigned long i;
5133         unsigned long num_pages;
5134         struct page *page;
5135
5136         num_pages = num_extent_pages(eb->start, eb->len);
5137
5138         for (i = 0; i < num_pages; i++) {
5139                 page = eb->pages[i];
5140                 if (!PageDirty(page))
5141                         continue;
5142
5143                 lock_page(page);
5144                 WARN_ON(!PagePrivate(page));
5145
5146                 clear_page_dirty_for_io(page);
5147                 spin_lock_irq(&page->mapping->tree_lock);
5148                 if (!PageDirty(page)) {
5149                         radix_tree_tag_clear(&page->mapping->page_tree,
5150                                                 page_index(page),
5151                                                 PAGECACHE_TAG_DIRTY);
5152                 }
5153                 spin_unlock_irq(&page->mapping->tree_lock);
5154                 ClearPageError(page);
5155                 unlock_page(page);
5156         }
5157         WARN_ON(atomic_read(&eb->refs) == 0);
5158 }
5159
5160 int set_extent_buffer_dirty(struct extent_buffer *eb)
5161 {
5162         unsigned long i;
5163         unsigned long num_pages;
5164         int was_dirty = 0;
5165
5166         check_buffer_tree_ref(eb);
5167
5168         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5169
5170         num_pages = num_extent_pages(eb->start, eb->len);
5171         WARN_ON(atomic_read(&eb->refs) == 0);
5172         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5173
5174         for (i = 0; i < num_pages; i++)
5175                 set_page_dirty(eb->pages[i]);
5176         return was_dirty;
5177 }
5178
5179 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5180 {
5181         unsigned long i;
5182         struct page *page;
5183         unsigned long num_pages;
5184
5185         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5186         num_pages = num_extent_pages(eb->start, eb->len);
5187         for (i = 0; i < num_pages; i++) {
5188                 page = eb->pages[i];
5189                 if (page)
5190                         ClearPageUptodate(page);
5191         }
5192 }
5193
5194 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5195 {
5196         unsigned long i;
5197         struct page *page;
5198         unsigned long num_pages;
5199
5200         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5201         num_pages = num_extent_pages(eb->start, eb->len);
5202         for (i = 0; i < num_pages; i++) {
5203                 page = eb->pages[i];
5204                 SetPageUptodate(page);
5205         }
5206 }
5207
5208 int extent_buffer_uptodate(struct extent_buffer *eb)
5209 {
5210         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5211 }
5212
5213 int read_extent_buffer_pages(struct extent_io_tree *tree,
5214                              struct extent_buffer *eb, int wait,
5215                              get_extent_t *get_extent, int mirror_num)
5216 {
5217         unsigned long i;
5218         struct page *page;
5219         int err;
5220         int ret = 0;
5221         int locked_pages = 0;
5222         int all_uptodate = 1;
5223         unsigned long num_pages;
5224         unsigned long num_reads = 0;
5225         struct bio *bio = NULL;
5226         unsigned long bio_flags = 0;
5227
5228         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5229                 return 0;
5230
5231         num_pages = num_extent_pages(eb->start, eb->len);
5232         for (i = 0; i < num_pages; i++) {
5233                 page = eb->pages[i];
5234                 if (wait == WAIT_NONE) {
5235                         if (!trylock_page(page))
5236                                 goto unlock_exit;
5237                 } else {
5238                         lock_page(page);
5239                 }
5240                 locked_pages++;
5241         }
5242         /*
5243          * We need to firstly lock all pages to make sure that
5244          * the uptodate bit of our pages won't be affected by
5245          * clear_extent_buffer_uptodate().
5246          */
5247         for (i = 0; i < num_pages; i++) {
5248                 page = eb->pages[i];
5249                 if (!PageUptodate(page)) {
5250                         num_reads++;
5251                         all_uptodate = 0;
5252                 }
5253         }
5254
5255         if (all_uptodate) {
5256                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5257                 goto unlock_exit;
5258         }
5259
5260         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5261         eb->read_mirror = 0;
5262         atomic_set(&eb->io_pages, num_reads);
5263         for (i = 0; i < num_pages; i++) {
5264                 page = eb->pages[i];
5265
5266                 if (!PageUptodate(page)) {
5267                         if (ret) {
5268                                 atomic_dec(&eb->io_pages);
5269                                 unlock_page(page);
5270                                 continue;
5271                         }
5272
5273                         ClearPageError(page);
5274                         err = __extent_read_full_page(tree, page,
5275                                                       get_extent, &bio,
5276                                                       mirror_num, &bio_flags,
5277                                                       REQ_META);
5278                         if (err) {
5279                                 ret = err;
5280                                 /*
5281                                  * We use &bio in above __extent_read_full_page,
5282                                  * so we ensure that if it returns error, the
5283                                  * current page fails to add itself to bio and
5284                                  * it's been unlocked.
5285                                  *
5286                                  * We must dec io_pages by ourselves.
5287                                  */
5288                                 atomic_dec(&eb->io_pages);
5289                         }
5290                 } else {
5291                         unlock_page(page);
5292                 }
5293         }
5294
5295         if (bio) {
5296                 err = submit_one_bio(bio, mirror_num, bio_flags);
5297                 if (err)
5298                         return err;
5299         }
5300
5301         if (ret || wait != WAIT_COMPLETE)
5302                 return ret;
5303
5304         for (i = 0; i < num_pages; i++) {
5305                 page = eb->pages[i];
5306                 wait_on_page_locked(page);
5307                 if (!PageUptodate(page))
5308                         ret = -EIO;
5309         }
5310
5311         return ret;
5312
5313 unlock_exit:
5314         while (locked_pages > 0) {
5315                 locked_pages--;
5316                 page = eb->pages[locked_pages];
5317                 unlock_page(page);
5318         }
5319         return ret;
5320 }
5321
5322 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5323                         unsigned long start,
5324                         unsigned long len)
5325 {
5326         size_t cur;
5327         size_t offset;
5328         struct page *page;
5329         char *kaddr;
5330         char *dst = (char *)dstv;
5331         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5332         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5333
5334         WARN_ON(start > eb->len);
5335         WARN_ON(start + len > eb->start + eb->len);
5336
5337         offset = (start_offset + start) & (PAGE_SIZE - 1);
5338
5339         while (len > 0) {
5340                 page = eb->pages[i];
5341
5342                 cur = min(len, (PAGE_SIZE - offset));
5343                 kaddr = page_address(page);
5344                 memcpy(dst, kaddr + offset, cur);
5345
5346                 dst += cur;
5347                 len -= cur;
5348                 offset = 0;
5349                 i++;
5350         }
5351 }
5352
5353 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5354                         unsigned long start,
5355                         unsigned long len)
5356 {
5357         size_t cur;
5358         size_t offset;
5359         struct page *page;
5360         char *kaddr;
5361         char __user *dst = (char __user *)dstv;
5362         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5363         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5364         int ret = 0;
5365
5366         WARN_ON(start > eb->len);
5367         WARN_ON(start + len > eb->start + eb->len);
5368
5369         offset = (start_offset + start) & (PAGE_SIZE - 1);
5370
5371         while (len > 0) {
5372                 page = eb->pages[i];
5373
5374                 cur = min(len, (PAGE_SIZE - offset));
5375                 kaddr = page_address(page);
5376                 if (copy_to_user(dst, kaddr + offset, cur)) {
5377                         ret = -EFAULT;
5378                         break;
5379                 }
5380
5381                 dst += cur;
5382                 len -= cur;
5383                 offset = 0;
5384                 i++;
5385         }
5386
5387         return ret;
5388 }
5389
5390 /*
5391  * return 0 if the item is found within a page.
5392  * return 1 if the item spans two pages.
5393  * return -EINVAL otherwise.
5394  */
5395 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5396                                unsigned long min_len, char **map,
5397                                unsigned long *map_start,
5398                                unsigned long *map_len)
5399 {
5400         size_t offset = start & (PAGE_SIZE - 1);
5401         char *kaddr;
5402         struct page *p;
5403         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5404         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5405         unsigned long end_i = (start_offset + start + min_len - 1) >>
5406                 PAGE_SHIFT;
5407
5408         if (i != end_i)
5409                 return 1;
5410
5411         if (i == 0) {
5412                 offset = start_offset;
5413                 *map_start = 0;
5414         } else {
5415                 offset = 0;
5416                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5417         }
5418
5419         if (start + min_len > eb->len) {
5420                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5421                        eb->start, eb->len, start, min_len);
5422                 return -EINVAL;
5423         }
5424
5425         p = eb->pages[i];
5426         kaddr = page_address(p);
5427         *map = kaddr + offset;
5428         *map_len = PAGE_SIZE - offset;
5429         return 0;
5430 }
5431
5432 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5433                           unsigned long start,
5434                           unsigned long len)
5435 {
5436         size_t cur;
5437         size_t offset;
5438         struct page *page;
5439         char *kaddr;
5440         char *ptr = (char *)ptrv;
5441         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5442         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5443         int ret = 0;
5444
5445         WARN_ON(start > eb->len);
5446         WARN_ON(start + len > eb->start + eb->len);
5447
5448         offset = (start_offset + start) & (PAGE_SIZE - 1);
5449
5450         while (len > 0) {
5451                 page = eb->pages[i];
5452
5453                 cur = min(len, (PAGE_SIZE - offset));
5454
5455                 kaddr = page_address(page);
5456                 ret = memcmp(ptr, kaddr + offset, cur);
5457                 if (ret)
5458                         break;
5459
5460                 ptr += cur;
5461                 len -= cur;
5462                 offset = 0;
5463                 i++;
5464         }
5465         return ret;
5466 }
5467
5468 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5469                          unsigned long start, unsigned long len)
5470 {
5471         size_t cur;
5472         size_t offset;
5473         struct page *page;
5474         char *kaddr;
5475         char *src = (char *)srcv;
5476         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5477         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5478
5479         WARN_ON(start > eb->len);
5480         WARN_ON(start + len > eb->start + eb->len);
5481
5482         offset = (start_offset + start) & (PAGE_SIZE - 1);
5483
5484         while (len > 0) {
5485                 page = eb->pages[i];
5486                 WARN_ON(!PageUptodate(page));
5487
5488                 cur = min(len, PAGE_SIZE - offset);
5489                 kaddr = page_address(page);
5490                 memcpy(kaddr + offset, src, cur);
5491
5492                 src += cur;
5493                 len -= cur;
5494                 offset = 0;
5495                 i++;
5496         }
5497 }
5498
5499 void memset_extent_buffer(struct extent_buffer *eb, char c,
5500                           unsigned long start, unsigned long len)
5501 {
5502         size_t cur;
5503         size_t offset;
5504         struct page *page;
5505         char *kaddr;
5506         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5507         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5508
5509         WARN_ON(start > eb->len);
5510         WARN_ON(start + len > eb->start + eb->len);
5511
5512         offset = (start_offset + start) & (PAGE_SIZE - 1);
5513
5514         while (len > 0) {
5515                 page = eb->pages[i];
5516                 WARN_ON(!PageUptodate(page));
5517
5518                 cur = min(len, PAGE_SIZE - offset);
5519                 kaddr = page_address(page);
5520                 memset(kaddr + offset, c, cur);
5521
5522                 len -= cur;
5523                 offset = 0;
5524                 i++;
5525         }
5526 }
5527
5528 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5529                         unsigned long dst_offset, unsigned long src_offset,
5530                         unsigned long len)
5531 {
5532         u64 dst_len = dst->len;
5533         size_t cur;
5534         size_t offset;
5535         struct page *page;
5536         char *kaddr;
5537         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5538         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5539
5540         WARN_ON(src->len != dst_len);
5541
5542         offset = (start_offset + dst_offset) &
5543                 (PAGE_SIZE - 1);
5544
5545         while (len > 0) {
5546                 page = dst->pages[i];
5547                 WARN_ON(!PageUptodate(page));
5548
5549                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5550
5551                 kaddr = page_address(page);
5552                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5553
5554                 src_offset += cur;
5555                 len -= cur;
5556                 offset = 0;
5557                 i++;
5558         }
5559 }
5560
5561 void le_bitmap_set(u8 *map, unsigned int start, int len)
5562 {
5563         u8 *p = map + BIT_BYTE(start);
5564         const unsigned int size = start + len;
5565         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5566         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5567
5568         while (len - bits_to_set >= 0) {
5569                 *p |= mask_to_set;
5570                 len -= bits_to_set;
5571                 bits_to_set = BITS_PER_BYTE;
5572                 mask_to_set = ~(u8)0;
5573                 p++;
5574         }
5575         if (len) {
5576                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5577                 *p |= mask_to_set;
5578         }
5579 }
5580
5581 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5582 {
5583         u8 *p = map + BIT_BYTE(start);
5584         const unsigned int size = start + len;
5585         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5586         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5587
5588         while (len - bits_to_clear >= 0) {
5589                 *p &= ~mask_to_clear;
5590                 len -= bits_to_clear;
5591                 bits_to_clear = BITS_PER_BYTE;
5592                 mask_to_clear = ~(u8)0;
5593                 p++;
5594         }
5595         if (len) {
5596                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5597                 *p &= ~mask_to_clear;
5598         }
5599 }
5600
5601 /*
5602  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5603  * given bit number
5604  * @eb: the extent buffer
5605  * @start: offset of the bitmap item in the extent buffer
5606  * @nr: bit number
5607  * @page_index: return index of the page in the extent buffer that contains the
5608  * given bit number
5609  * @page_offset: return offset into the page given by page_index
5610  *
5611  * This helper hides the ugliness of finding the byte in an extent buffer which
5612  * contains a given bit.
5613  */
5614 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5615                                     unsigned long start, unsigned long nr,
5616                                     unsigned long *page_index,
5617                                     size_t *page_offset)
5618 {
5619         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5620         size_t byte_offset = BIT_BYTE(nr);
5621         size_t offset;
5622
5623         /*
5624          * The byte we want is the offset of the extent buffer + the offset of
5625          * the bitmap item in the extent buffer + the offset of the byte in the
5626          * bitmap item.
5627          */
5628         offset = start_offset + start + byte_offset;
5629
5630         *page_index = offset >> PAGE_SHIFT;
5631         *page_offset = offset & (PAGE_SIZE - 1);
5632 }
5633
5634 /**
5635  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5636  * @eb: the extent buffer
5637  * @start: offset of the bitmap item in the extent buffer
5638  * @nr: bit number to test
5639  */
5640 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5641                            unsigned long nr)
5642 {
5643         u8 *kaddr;
5644         struct page *page;
5645         unsigned long i;
5646         size_t offset;
5647
5648         eb_bitmap_offset(eb, start, nr, &i, &offset);
5649         page = eb->pages[i];
5650         WARN_ON(!PageUptodate(page));
5651         kaddr = page_address(page);
5652         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5653 }
5654
5655 /**
5656  * extent_buffer_bitmap_set - set an area of a bitmap
5657  * @eb: the extent buffer
5658  * @start: offset of the bitmap item in the extent buffer
5659  * @pos: bit number of the first bit
5660  * @len: number of bits to set
5661  */
5662 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5663                               unsigned long pos, unsigned long len)
5664 {
5665         u8 *kaddr;
5666         struct page *page;
5667         unsigned long i;
5668         size_t offset;
5669         const unsigned int size = pos + len;
5670         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5671         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5672
5673         eb_bitmap_offset(eb, start, pos, &i, &offset);
5674         page = eb->pages[i];
5675         WARN_ON(!PageUptodate(page));
5676         kaddr = page_address(page);
5677
5678         while (len >= bits_to_set) {
5679                 kaddr[offset] |= mask_to_set;
5680                 len -= bits_to_set;
5681                 bits_to_set = BITS_PER_BYTE;
5682                 mask_to_set = ~(u8)0;
5683                 if (++offset >= PAGE_SIZE && len > 0) {
5684                         offset = 0;
5685                         page = eb->pages[++i];
5686                         WARN_ON(!PageUptodate(page));
5687                         kaddr = page_address(page);
5688                 }
5689         }
5690         if (len) {
5691                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5692                 kaddr[offset] |= mask_to_set;
5693         }
5694 }
5695
5696
5697 /**
5698  * extent_buffer_bitmap_clear - clear an area of a bitmap
5699  * @eb: the extent buffer
5700  * @start: offset of the bitmap item in the extent buffer
5701  * @pos: bit number of the first bit
5702  * @len: number of bits to clear
5703  */
5704 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5705                                 unsigned long pos, unsigned long len)
5706 {
5707         u8 *kaddr;
5708         struct page *page;
5709         unsigned long i;
5710         size_t offset;
5711         const unsigned int size = pos + len;
5712         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5713         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5714
5715         eb_bitmap_offset(eb, start, pos, &i, &offset);
5716         page = eb->pages[i];
5717         WARN_ON(!PageUptodate(page));
5718         kaddr = page_address(page);
5719
5720         while (len >= bits_to_clear) {
5721                 kaddr[offset] &= ~mask_to_clear;
5722                 len -= bits_to_clear;
5723                 bits_to_clear = BITS_PER_BYTE;
5724                 mask_to_clear = ~(u8)0;
5725                 if (++offset >= PAGE_SIZE && len > 0) {
5726                         offset = 0;
5727                         page = eb->pages[++i];
5728                         WARN_ON(!PageUptodate(page));
5729                         kaddr = page_address(page);
5730                 }
5731         }
5732         if (len) {
5733                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5734                 kaddr[offset] &= ~mask_to_clear;
5735         }
5736 }
5737
5738 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5739 {
5740         unsigned long distance = (src > dst) ? src - dst : dst - src;
5741         return distance < len;
5742 }
5743
5744 static void copy_pages(struct page *dst_page, struct page *src_page,
5745                        unsigned long dst_off, unsigned long src_off,
5746                        unsigned long len)
5747 {
5748         char *dst_kaddr = page_address(dst_page);
5749         char *src_kaddr;
5750         int must_memmove = 0;
5751
5752         if (dst_page != src_page) {
5753                 src_kaddr = page_address(src_page);
5754         } else {
5755                 src_kaddr = dst_kaddr;
5756                 if (areas_overlap(src_off, dst_off, len))
5757                         must_memmove = 1;
5758         }
5759
5760         if (must_memmove)
5761                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5762         else
5763                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5764 }
5765
5766 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5767                            unsigned long src_offset, unsigned long len)
5768 {
5769         size_t cur;
5770         size_t dst_off_in_page;
5771         size_t src_off_in_page;
5772         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5773         unsigned long dst_i;
5774         unsigned long src_i;
5775
5776         if (src_offset + len > dst->len) {
5777                 btrfs_err(dst->fs_info,
5778                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5779                          src_offset, len, dst->len);
5780                 BUG_ON(1);
5781         }
5782         if (dst_offset + len > dst->len) {
5783                 btrfs_err(dst->fs_info,
5784                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5785                          dst_offset, len, dst->len);
5786                 BUG_ON(1);
5787         }
5788
5789         while (len > 0) {
5790                 dst_off_in_page = (start_offset + dst_offset) &
5791                         (PAGE_SIZE - 1);
5792                 src_off_in_page = (start_offset + src_offset) &
5793                         (PAGE_SIZE - 1);
5794
5795                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5796                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5797
5798                 cur = min(len, (unsigned long)(PAGE_SIZE -
5799                                                src_off_in_page));
5800                 cur = min_t(unsigned long, cur,
5801                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5802
5803                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5804                            dst_off_in_page, src_off_in_page, cur);
5805
5806                 src_offset += cur;
5807                 dst_offset += cur;
5808                 len -= cur;
5809         }
5810 }
5811
5812 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5813                            unsigned long src_offset, unsigned long len)
5814 {
5815         size_t cur;
5816         size_t dst_off_in_page;
5817         size_t src_off_in_page;
5818         unsigned long dst_end = dst_offset + len - 1;
5819         unsigned long src_end = src_offset + len - 1;
5820         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5821         unsigned long dst_i;
5822         unsigned long src_i;
5823
5824         if (src_offset + len > dst->len) {
5825                 btrfs_err(dst->fs_info,
5826                           "memmove bogus src_offset %lu move len %lu len %lu",
5827                           src_offset, len, dst->len);
5828                 BUG_ON(1);
5829         }
5830         if (dst_offset + len > dst->len) {
5831                 btrfs_err(dst->fs_info,
5832                           "memmove bogus dst_offset %lu move len %lu len %lu",
5833                           dst_offset, len, dst->len);
5834                 BUG_ON(1);
5835         }
5836         if (dst_offset < src_offset) {
5837                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5838                 return;
5839         }
5840         while (len > 0) {
5841                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5842                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5843
5844                 dst_off_in_page = (start_offset + dst_end) &
5845                         (PAGE_SIZE - 1);
5846                 src_off_in_page = (start_offset + src_end) &
5847                         (PAGE_SIZE - 1);
5848
5849                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5850                 cur = min(cur, dst_off_in_page + 1);
5851                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5852                            dst_off_in_page - cur + 1,
5853                            src_off_in_page - cur + 1, cur);
5854
5855                 dst_end -= cur;
5856                 src_end -= cur;
5857                 len -= cur;
5858         }
5859 }
5860
5861 int try_release_extent_buffer(struct page *page)
5862 {
5863         struct extent_buffer *eb;
5864
5865         /*
5866          * We need to make sure nobody is attaching this page to an eb right
5867          * now.
5868          */
5869         spin_lock(&page->mapping->private_lock);
5870         if (!PagePrivate(page)) {
5871                 spin_unlock(&page->mapping->private_lock);
5872                 return 1;
5873         }
5874
5875         eb = (struct extent_buffer *)page->private;
5876         BUG_ON(!eb);
5877
5878         /*
5879          * This is a little awful but should be ok, we need to make sure that
5880          * the eb doesn't disappear out from under us while we're looking at
5881          * this page.
5882          */
5883         spin_lock(&eb->refs_lock);
5884         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5885                 spin_unlock(&eb->refs_lock);
5886                 spin_unlock(&page->mapping->private_lock);
5887                 return 0;
5888         }
5889         spin_unlock(&page->mapping->private_lock);
5890
5891         /*
5892          * If tree ref isn't set then we know the ref on this eb is a real ref,
5893          * so just return, this page will likely be freed soon anyway.
5894          */
5895         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5896                 spin_unlock(&eb->refs_lock);
5897                 return 0;
5898         }
5899
5900         return release_extent_buffer(eb);
5901 }