fs/btrfs/extent_io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/bitops.h>
   3 #include <linux/slab.h>
   4 #include <linux/bio.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/page-flags.h>
   8 #include <linux/spinlock.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/swap.h>
  11 #include <linux/writeback.h>
  12 #include <linux/pagevec.h>
  13 #include <linux/prefetch.h>
  14 #include <linux/cleancache.h>
  15 #include "extent_io.h"
  16 #include "extent_map.h"
  17 #include "ctree.h"
  18 #include "btrfs_inode.h"
  19 #include "volumes.h"
  20 #include "check-integrity.h"
  21 #include "locking.h"
  22 #include "rcu-string.h"
  23 #include "backref.h"
  24 #include "disk-io.h"
  25
  26 static struct kmem_cache *extent_state_cache;
  27 static struct kmem_cache *extent_buffer_cache;
  28 static struct bio_set *btrfs_bioset;
  29
  30 static inline bool extent_state_in_tree(const struct extent_state *state)
  31 {
  32         return !RB_EMPTY_NODE(&state->rb_node);
  33 }
  34
  35 #ifdef CONFIG_BTRFS_DEBUG
  36 static LIST_HEAD(buffers);
  37 static LIST_HEAD(states);
  38
  39 static DEFINE_SPINLOCK(leak_lock);
  40
  41 static inline
  42 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  43 {
  44         unsigned long flags;
  45
  46         spin_lock_irqsave(&leak_lock, flags);
  47         list_add(new, head);
  48         spin_unlock_irqrestore(&leak_lock, flags);
  49 }
  50
  51 static inline
  52 void btrfs_leak_debug_del(struct list_head *entry)
  53 {
  54         unsigned long flags;
  55
  56         spin_lock_irqsave(&leak_lock, flags);
  57         list_del(entry);
  58         spin_unlock_irqrestore(&leak_lock, flags);
  59 }
  60
  61 static inline
  62 void btrfs_leak_debug_check(void)
  63 {
  64         struct extent_state *state;
  65         struct extent_buffer *eb;
  66
  67         while (!list_empty(&states)) {
  68                 state = list_entry(states.next, struct extent_state, leak_list);
  69                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  70                        state->start, state->end, state->state,
  71                        extent_state_in_tree(state),
  72                        refcount_read(&state->refs));
  73                 list_del(&state->leak_list);
  74                 kmem_cache_free(extent_state_cache, state);
  75         }
  76
  77         while (!list_empty(&buffers)) {
  78                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  79                 pr_err("BTRFS: buffer leak start %llu len %lu refs %d\n",
  80                        eb->start, eb->len, atomic_read(&eb->refs));
  81                 list_del(&eb->leak_list);
  82                 kmem_cache_free(extent_buffer_cache, eb);
  83         }
  84 }
  85
  86 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  87         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  88 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  89                 struct extent_io_tree *tree, u64 start, u64 end)
  90 {
  91         if (tree->ops && tree->ops->check_extent_io_range)
  92                 tree->ops->check_extent_io_range(tree->private_data, caller,
  93                                                  start, end);
  94 }
  95 #else
  96 #define btrfs_leak_debug_add(new, head) do {} while (0)
  97 #define btrfs_leak_debug_del(entry)     do {} while (0)
  98 #define btrfs_leak_debug_check()        do {} while (0)
  99 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 100 #endif
 101
 102 #define BUFFER_LRU_MAX 64
 103
 104 struct tree_entry {
 105         u64 start;
 106         u64 end;
 107         struct rb_node rb_node;
 108 };
 109
 110 struct extent_page_data {
 111         struct bio *bio;
 112         struct extent_io_tree *tree;
 113         /* tells writepage not to lock the state bits for this range
 114          * it still does the unlocking
 115          */
 116         unsigned int extent_locked:1;
 117
 118         /* tells the submit_bio code to use REQ_SYNC */
 119         unsigned int sync_io:1;
 120 };
 121
 122 static void add_extent_changeset(struct extent_state *state, unsigned bits,
 123                                  struct extent_changeset *changeset,
 124                                  int set)
 125 {
 126         int ret;
 127
 128         if (!changeset)
 129                 return;
 130         if (set && (state->state & bits) == bits)
 131                 return;
 132         if (!set && (state->state & bits) == 0)
 133                 return;
 134         changeset->bytes_changed += state->end - state->start + 1;
 135         ret = ulist_add(&changeset->range_changed, state->start, state->end,
 136                         GFP_ATOMIC);
 137         /* ENOMEM */
 138         BUG_ON(ret < 0);
 139 }
 140
 141 static void flush_write_bio(struct extent_page_data *epd);
 142
 143 static inline struct btrfs_fs_info *
 144 tree_fs_info(struct extent_io_tree *tree)
 145 {
 146         if (tree->ops)
 147                 return tree->ops->tree_fs_info(tree->private_data);
 148         return NULL;
 149 }
 150
 151 int __init extent_io_init(void)
 152 {
 153         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 154                         sizeof(struct extent_state), 0,
 155                         SLAB_MEM_SPREAD, NULL);
 156         if (!extent_state_cache)
 157                 return -ENOMEM;
 158
 159         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 160                         sizeof(struct extent_buffer), 0,
 161                         SLAB_MEM_SPREAD, NULL);
 162         if (!extent_buffer_cache)
 163                 goto free_state_cache;
 164
 165         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 166                                      offsetof(struct btrfs_io_bio, bio),
 167                                      BIOSET_NEED_BVECS);
 168         if (!btrfs_bioset)
 169                 goto free_buffer_cache;
 170
 171         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 172                 goto free_bioset;
 173
 174         return 0;
 175
 176 free_bioset:
 177         bioset_free(btrfs_bioset);
 178         btrfs_bioset = NULL;
 179
 180 free_buffer_cache:
 181         kmem_cache_destroy(extent_buffer_cache);
 182         extent_buffer_cache = NULL;
 183
 184 free_state_cache:
 185         kmem_cache_destroy(extent_state_cache);
 186         extent_state_cache = NULL;
 187         return -ENOMEM;
 188 }
 189
 190 void extent_io_exit(void)
 191 {
 192         btrfs_leak_debug_check();
 193
 194         /*
 195          * Make sure all delayed rcu free are flushed before we
 196          * destroy caches.
 197          */
 198         rcu_barrier();
 199         kmem_cache_destroy(extent_state_cache);
 200         kmem_cache_destroy(extent_buffer_cache);
 201         if (btrfs_bioset)
 202                 bioset_free(btrfs_bioset);
 203 }
 204
 205 void extent_io_tree_init(struct extent_io_tree *tree,
 206                          void *private_data)
 207 {
 208         tree->state = RB_ROOT;
 209         tree->ops = NULL;
 210         tree->dirty_bytes = 0;
 211         spin_lock_init(&tree->lock);
 212         tree->private_data = private_data;
 213 }
 214
 215 static struct extent_state *alloc_extent_state(gfp_t mask)
 216 {
 217         struct extent_state *state;
 218
 219         /*
 220          * The given mask might be not appropriate for the slab allocator,
 221          * drop the unsupported bits
 222          */
 223         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
 224         state = kmem_cache_alloc(extent_state_cache, mask);
 225         if (!state)
 226                 return state;
 227         state->state = 0;
 228         state->failrec = NULL;
 229         RB_CLEAR_NODE(&state->rb_node);
 230         btrfs_leak_debug_add(&state->leak_list, &states);
 231         refcount_set(&state->refs, 1);
 232         init_waitqueue_head(&state->wq);
 233         trace_alloc_extent_state(state, mask, _RET_IP_);
 234         return state;
 235 }
 236
 237 void free_extent_state(struct extent_state *state)
 238 {
 239         if (!state)
 240                 return;
 241         if (refcount_dec_and_test(&state->refs)) {
 242                 WARN_ON(extent_state_in_tree(state));
 243                 btrfs_leak_debug_del(&state->leak_list);
 244                 trace_free_extent_state(state, _RET_IP_);
 245                 kmem_cache_free(extent_state_cache, state);
 246         }
 247 }
 248
 249 static struct rb_node *tree_insert(struct rb_root *root,
 250                                    struct rb_node *search_start,
 251                                    u64 offset,
 252                                    struct rb_node *node,
 253                                    struct rb_node ***p_in,
 254                                    struct rb_node **parent_in)
 255 {
 256         struct rb_node **p;
 257         struct rb_node *parent = NULL;
 258         struct tree_entry *entry;
 259
 260         if (p_in && parent_in) {
 261                 p = *p_in;
 262                 parent = *parent_in;
 263                 goto do_insert;
 264         }
 265
 266         p = search_start ? &search_start : &root->rb_node;
 267         while (*p) {
 268                 parent = *p;
 269                 entry = rb_entry(parent, struct tree_entry, rb_node);
 270
 271                 if (offset < entry->start)
 272                         p = &(*p)->rb_left;
 273                 else if (offset > entry->end)
 274                         p = &(*p)->rb_right;
 275                 else
 276                         return parent;
 277         }
 278
 279 do_insert:
 280         rb_link_node(node, parent, p);
 281         rb_insert_color(node, root);
 282         return NULL;
 283 }
 284
 285 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 286                                       struct rb_node **prev_ret,
 287                                       struct rb_node **next_ret,
 288                                       struct rb_node ***p_ret,
 289                                       struct rb_node **parent_ret)
 290 {
 291         struct rb_root *root = &tree->state;
 292         struct rb_node **n = &root->rb_node;
 293         struct rb_node *prev = NULL;
 294         struct rb_node *orig_prev = NULL;
 295         struct tree_entry *entry;
 296         struct tree_entry *prev_entry = NULL;
 297
 298         while (*n) {
 299                 prev = *n;
 300                 entry = rb_entry(prev, struct tree_entry, rb_node);
 301                 prev_entry = entry;
 302
 303                 if (offset < entry->start)
 304                         n = &(*n)->rb_left;
 305                 else if (offset > entry->end)
 306                         n = &(*n)->rb_right;
 307                 else
 308                         return *n;
 309         }
 310
 311         if (p_ret)
 312                 *p_ret = n;
 313         if (parent_ret)
 314                 *parent_ret = prev;
 315
 316         if (prev_ret) {
 317                 orig_prev = prev;
 318                 while (prev && offset > prev_entry->end) {
 319                         prev = rb_next(prev);
 320                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 321                 }
 322                 *prev_ret = prev;
 323                 prev = orig_prev;
 324         }
 325
 326         if (next_ret) {
 327                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 328                 while (prev && offset < prev_entry->start) {
 329                         prev = rb_prev(prev);
 330                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 331                 }
 332                 *next_ret = prev;
 333         }
 334         return NULL;
 335 }
 336
 337 static inline struct rb_node *
 338 tree_search_for_insert(struct extent_io_tree *tree,
 339                        u64 offset,
 340                        struct rb_node ***p_ret,
 341                        struct rb_node **parent_ret)
 342 {
 343         struct rb_node *prev = NULL;
 344         struct rb_node *ret;
 345
 346         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 347         if (!ret)
 348                 return prev;
 349         return ret;
 350 }
 351
 352 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 353                                           u64 offset)
 354 {
 355         return tree_search_for_insert(tree, offset, NULL, NULL);
 356 }
 357
 358 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 359                      struct extent_state *other)
 360 {
 361         if (tree->ops && tree->ops->merge_extent_hook)
 362                 tree->ops->merge_extent_hook(tree->private_data, new, other);
 363 }
 364
 365 /*
 366  * utility function to look for merge candidates inside a given range.
 367  * Any extents with matching state are merged together into a single
 368  * extent in the tree.  Extents with EXTENT_IO in their state field
 369  * are not merged because the end_io handlers need to be able to do
 370  * operations on them without sleeping (or doing allocations/splits).
 371  *
 372  * This should be called with the tree lock held.
 373  */
 374 static void merge_state(struct extent_io_tree *tree,
 375                         struct extent_state *state)
 376 {
 377         struct extent_state *other;
 378         struct rb_node *other_node;
 379
 380         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 381                 return;
 382
 383         other_node = rb_prev(&state->rb_node);
 384         if (other_node) {
 385                 other = rb_entry(other_node, struct extent_state, rb_node);
 386                 if (other->end == state->start - 1 &&
 387                     other->state == state->state) {
 388                         merge_cb(tree, state, other);
 389                         state->start = other->start;
 390                         rb_erase(&other->rb_node, &tree->state);
 391                         RB_CLEAR_NODE(&other->rb_node);
 392                         free_extent_state(other);
 393                 }
 394         }
 395         other_node = rb_next(&state->rb_node);
 396         if (other_node) {
 397                 other = rb_entry(other_node, struct extent_state, rb_node);
 398                 if (other->start == state->end + 1 &&
 399                     other->state == state->state) {
 400                         merge_cb(tree, state, other);
 401                         state->end = other->end;
 402                         rb_erase(&other->rb_node, &tree->state);
 403                         RB_CLEAR_NODE(&other->rb_node);
 404                         free_extent_state(other);
 405                 }
 406         }
 407 }
 408
 409 static void set_state_cb(struct extent_io_tree *tree,
 410                          struct extent_state *state, unsigned *bits)
 411 {
 412         if (tree->ops && tree->ops->set_bit_hook)
 413                 tree->ops->set_bit_hook(tree->private_data, state, bits);
 414 }
 415
 416 static void clear_state_cb(struct extent_io_tree *tree,
 417                            struct extent_state *state, unsigned *bits)
 418 {
 419         if (tree->ops && tree->ops->clear_bit_hook)
 420                 tree->ops->clear_bit_hook(tree->private_data, state, bits);
 421 }
 422
 423 static void set_state_bits(struct extent_io_tree *tree,
 424                            struct extent_state *state, unsigned *bits,
 425                            struct extent_changeset *changeset);
 426
 427 /*
 428  * insert an extent_state struct into the tree.  'bits' are set on the
 429  * struct before it is inserted.
 430  *
 431  * This may return -EEXIST if the extent is already there, in which case the
 432  * state struct is freed.
 433  *
 434  * The tree lock is not taken internally.  This is a utility function and
 435  * probably isn't what you want to call (see set/clear_extent_bit).
 436  */
 437 static int insert_state(struct extent_io_tree *tree,
 438                         struct extent_state *state, u64 start, u64 end,
 439                         struct rb_node ***p,
 440                         struct rb_node **parent,
 441                         unsigned *bits, struct extent_changeset *changeset)
 442 {
 443         struct rb_node *node;
 444
 445         if (end < start)
 446                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 447                        end, start);
 448         state->start = start;
 449         state->end = end;
 450
 451         set_state_bits(tree, state, bits, changeset);
 452
 453         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 454         if (node) {
 455                 struct extent_state *found;
 456                 found = rb_entry(node, struct extent_state, rb_node);
 457                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
 458                        found->start, found->end, start, end);
 459                 return -EEXIST;
 460         }
 461         merge_state(tree, state);
 462         return 0;
 463 }
 464
 465 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 466                      u64 split)
 467 {
 468         if (tree->ops && tree->ops->split_extent_hook)
 469                 tree->ops->split_extent_hook(tree->private_data, orig, split);
 470 }
 471
 472 /*
 473  * split a given extent state struct in two, inserting the preallocated
 474  * struct 'prealloc' as the newly created second half.  'split' indicates an
 475  * offset inside 'orig' where it should be split.
 476  *
 477  * Before calling,
 478  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 479  * are two extent state structs in the tree:
 480  * prealloc: [orig->start, split - 1]
 481  * orig: [ split, orig->end ]
 482  *
 483  * The tree locks are not taken by this function. They need to be held
 484  * by the caller.
 485  */
 486 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 487                        struct extent_state *prealloc, u64 split)
 488 {
 489         struct rb_node *node;
 490
 491         split_cb(tree, orig, split);
 492
 493         prealloc->start = orig->start;
 494         prealloc->end = split - 1;
 495         prealloc->state = orig->state;
 496         orig->start = split;
 497
 498         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 499                            &prealloc->rb_node, NULL, NULL);
 500         if (node) {
 501                 free_extent_state(prealloc);
 502                 return -EEXIST;
 503         }
 504         return 0;
 505 }
 506
 507 static struct extent_state *next_state(struct extent_state *state)
 508 {
 509         struct rb_node *next = rb_next(&state->rb_node);
 510         if (next)
 511                 return rb_entry(next, struct extent_state, rb_node);
 512         else
 513                 return NULL;
 514 }
 515
 516 /*
 517  * utility function to clear some bits in an extent state struct.
 518  * it will optionally wake up any one waiting on this state (wake == 1).
 519  *
 520  * If no bits are set on the state struct after clearing things, the
 521  * struct is freed and removed from the tree
 522  */
 523 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 524                                             struct extent_state *state,
 525                                             unsigned *bits, int wake,
 526                                             struct extent_changeset *changeset)
 527 {
 528         struct extent_state *next;
 529         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 530
 531         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 532                 u64 range = state->end - state->start + 1;
 533                 WARN_ON(range > tree->dirty_bytes);
 534                 tree->dirty_bytes -= range;
 535         }
 536         clear_state_cb(tree, state, bits);
 537         add_extent_changeset(state, bits_to_clear, changeset, 0);
 538         state->state &= ~bits_to_clear;
 539         if (wake)
 540                 wake_up(&state->wq);
 541         if (state->state == 0) {
 542                 next = next_state(state);
 543                 if (extent_state_in_tree(state)) {
 544                         rb_erase(&state->rb_node, &tree->state);
 545                         RB_CLEAR_NODE(&state->rb_node);
 546                         free_extent_state(state);
 547                 } else {
 548                         WARN_ON(1);
 549                 }
 550         } else {
 551                 merge_state(tree, state);
 552                 next = next_state(state);
 553         }
 554         return next;
 555 }
 556
 557 static struct extent_state *
 558 alloc_extent_state_atomic(struct extent_state *prealloc)
 559 {
 560         if (!prealloc)
 561                 prealloc = alloc_extent_state(GFP_ATOMIC);
 562
 563         return prealloc;
 564 }
 565
 566 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 567 {
 568         btrfs_panic(tree_fs_info(tree), err,
 569                     "Locking error: Extent tree was modified by another thread while locked.");
 570 }
 571
 572 /*
 573  * clear some bits on a range in the tree.  This may require splitting
 574  * or inserting elements in the tree, so the gfp mask is used to
 575  * indicate which allocations or sleeping are allowed.
 576  *
 577  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 578  * the given range from the tree regardless of state (ie for truncate).
 579  *
 580  * the range [start, end] is inclusive.
 581  *
 582  * This takes the tree lock, and returns 0 on success and < 0 on error.
 583  */
 584 int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 585                               unsigned bits, int wake, int delete,
 586                               struct extent_state **cached_state,
 587                               gfp_t mask, struct extent_changeset *changeset)
 588 {
 589         struct extent_state *state;
 590         struct extent_state *cached;
 591         struct extent_state *prealloc = NULL;
 592         struct rb_node *node;
 593         u64 last_end;
 594         int err;
 595         int clear = 0;
 596
 597         btrfs_debug_check_extent_io_range(tree, start, end);
 598
 599         if (bits & EXTENT_DELALLOC)
 600                 bits |= EXTENT_NORESERVE;
 601
 602         if (delete)
 603                 bits |= ~EXTENT_CTLBITS;
 604         bits |= EXTENT_FIRST_DELALLOC;
 605
 606         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 607                 clear = 1;
 608 again:
 609         if (!prealloc && gfpflags_allow_blocking(mask)) {
 610                 /*
 611                  * Don't care for allocation failure here because we might end
 612                  * up not needing the pre-allocated extent state at all, which
 613                  * is the case if we only have in the tree extent states that
 614                  * cover our input range and don't cover too any other range.
 615                  * If we end up needing a new extent state we allocate it later.
 616                  */
 617                 prealloc = alloc_extent_state(mask);
 618         }
 619
 620         spin_lock(&tree->lock);
 621         if (cached_state) {
 622                 cached = *cached_state;
 623
 624                 if (clear) {
 625                         *cached_state = NULL;
 626                         cached_state = NULL;
 627                 }
 628
 629                 if (cached && extent_state_in_tree(cached) &&
 630                     cached->start <= start && cached->end > start) {
 631                         if (clear)
 632                                 refcount_dec(&cached->refs);
 633                         state = cached;
 634                         goto hit_next;
 635                 }
 636                 if (clear)
 637                         free_extent_state(cached);
 638         }
 639         /*
 640          * this search will find the extents that end after
 641          * our range starts
 642          */
 643         node = tree_search(tree, start);
 644         if (!node)
 645                 goto out;
 646         state = rb_entry(node, struct extent_state, rb_node);
 647 hit_next:
 648         if (state->start > end)
 649                 goto out;
 650         WARN_ON(state->end < start);
 651         last_end = state->end;
 652
 653         /* the state doesn't have the wanted bits, go ahead */
 654         if (!(state->state & bits)) {
 655                 state = next_state(state);
 656                 goto next;
 657         }
 658
 659         /*
 660          *     | ---- desired range ---- |
 661          *  | state | or
 662          *  | ------------- state -------------- |
 663          *
 664          * We need to split the extent we found, and may flip
 665          * bits on second half.
 666          *
 667          * If the extent we found extends past our range, we
 668          * just split and search again.  It'll get split again
 669          * the next time though.
 670          *
 671          * If the extent we found is inside our range, we clear
 672          * the desired bit on it.
 673          */
 674
 675         if (state->start < start) {
 676                 prealloc = alloc_extent_state_atomic(prealloc);
 677                 BUG_ON(!prealloc);
 678                 err = split_state(tree, state, prealloc, start);
 679                 if (err)
 680                         extent_io_tree_panic(tree, err);
 681
 682                 prealloc = NULL;
 683                 if (err)
 684                         goto out;
 685                 if (state->end <= end) {
 686                         state = clear_state_bit(tree, state, &bits, wake,
 687                                                 changeset);
 688                         goto next;
 689                 }
 690                 goto search_again;
 691         }
 692         /*
 693          * | ---- desired range ---- |
 694          *                        | state |
 695          * We need to split the extent, and clear the bit
 696          * on the first half
 697          */
 698         if (state->start <= end && state->end > end) {
 699                 prealloc = alloc_extent_state_atomic(prealloc);
 700                 BUG_ON(!prealloc);
 701                 err = split_state(tree, state, prealloc, end + 1);
 702                 if (err)
 703                         extent_io_tree_panic(tree, err);
 704
 705                 if (wake)
 706                         wake_up(&state->wq);
 707
 708                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 709
 710                 prealloc = NULL;
 711                 goto out;
 712         }
 713
 714         state = clear_state_bit(tree, state, &bits, wake, changeset);
 715 next:
 716         if (last_end == (u64)-1)
 717                 goto out;
 718         start = last_end + 1;
 719         if (start <= end && state && !need_resched())
 720                 goto hit_next;
 721
 722 search_again:
 723         if (start > end)
 724                 goto out;
 725         spin_unlock(&tree->lock);
 726         if (gfpflags_allow_blocking(mask))
 727                 cond_resched();
 728         goto again;
 729
 730 out:
 731         spin_unlock(&tree->lock);
 732         if (prealloc)
 733                 free_extent_state(prealloc);
 734
 735         return 0;
 736
 737 }
 738
 739 static void wait_on_state(struct extent_io_tree *tree,
 740                           struct extent_state *state)
 741                 __releases(tree->lock)
 742                 __acquires(tree->lock)
 743 {
 744         DEFINE_WAIT(wait);
 745         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 746         spin_unlock(&tree->lock);
 747         schedule();
 748         spin_lock(&tree->lock);
 749         finish_wait(&state->wq, &wait);
 750 }
 751
 752 /*
 753  * waits for one or more bits to clear on a range in the state tree.
 754  * The range [start, end] is inclusive.
 755  * The tree lock is taken by this function
 756  */
 757 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 758                             unsigned long bits)
 759 {
 760         struct extent_state *state;
 761         struct rb_node *node;
 762
 763         btrfs_debug_check_extent_io_range(tree, start, end);
 764
 765         spin_lock(&tree->lock);
 766 again:
 767         while (1) {
 768                 /*
 769                  * this search will find all the extents that end after
 770                  * our range starts
 771                  */
 772                 node = tree_search(tree, start);
 773 process_node:
 774                 if (!node)
 775                         break;
 776
 777                 state = rb_entry(node, struct extent_state, rb_node);
 778
 779                 if (state->start > end)
 780                         goto out;
 781
 782                 if (state->state & bits) {
 783                         start = state->start;
 784                         refcount_inc(&state->refs);
 785                         wait_on_state(tree, state);
 786                         free_extent_state(state);
 787                         goto again;
 788                 }
 789                 start = state->end + 1;
 790
 791                 if (start > end)
 792                         break;
 793
 794                 if (!cond_resched_lock(&tree->lock)) {
 795                         node = rb_next(node);
 796                         goto process_node;
 797                 }
 798         }
 799 out:
 800         spin_unlock(&tree->lock);
 801 }
 802
 803 static void set_state_bits(struct extent_io_tree *tree,
 804                            struct extent_state *state,
 805                            unsigned *bits, struct extent_changeset *changeset)
 806 {
 807         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 808
 809         set_state_cb(tree, state, bits);
 810         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 811                 u64 range = state->end - state->start + 1;
 812                 tree->dirty_bytes += range;
 813         }
 814         add_extent_changeset(state, bits_to_set, changeset, 1);
 815         state->state |= bits_to_set;
 816 }
 817
 818 static void cache_state_if_flags(struct extent_state *state,
 819                                  struct extent_state **cached_ptr,
 820                                  unsigned flags)
 821 {
 822         if (cached_ptr && !(*cached_ptr)) {
 823                 if (!flags || (state->state & flags)) {
 824                         *cached_ptr = state;
 825                         refcount_inc(&state->refs);
 826                 }
 827         }
 828 }
 829
 830 static void cache_state(struct extent_state *state,
 831                         struct extent_state **cached_ptr)
 832 {
 833         return cache_state_if_flags(state, cached_ptr,
 834                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 835 }
 836
 837 /*
 838  * set some bits on a range in the tree.  This may require allocations or
 839  * sleeping, so the gfp mask is used to indicate what is allowed.
 840  *
 841  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 842  * part of the range already has the desired bits set.  The start of the
 843  * existing range is returned in failed_start in this case.
 844  *
 845  * [start, end] is inclusive This takes the tree lock.
 846  */
 847
 848 static int __must_check
 849 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 850                  unsigned bits, unsigned exclusive_bits,
 851                  u64 *failed_start, struct extent_state **cached_state,
 852                  gfp_t mask, struct extent_changeset *changeset)
 853 {
 854         struct extent_state *state;
 855         struct extent_state *prealloc = NULL;
 856         struct rb_node *node;
 857         struct rb_node **p;
 858         struct rb_node *parent;
 859         int err = 0;
 860         u64 last_start;
 861         u64 last_end;
 862
 863         btrfs_debug_check_extent_io_range(tree, start, end);
 864
 865         bits |= EXTENT_FIRST_DELALLOC;
 866 again:
 867         if (!prealloc && gfpflags_allow_blocking(mask)) {
 868                 /*
 869                  * Don't care for allocation failure here because we might end
 870                  * up not needing the pre-allocated extent state at all, which
 871                  * is the case if we only have in the tree extent states that
 872                  * cover our input range and don't cover too any other range.
 873                  * If we end up needing a new extent state we allocate it later.
 874                  */
 875                 prealloc = alloc_extent_state(mask);
 876         }
 877
 878         spin_lock(&tree->lock);
 879         if (cached_state && *cached_state) {
 880                 state = *cached_state;
 881                 if (state->start <= start && state->end > start &&
 882                     extent_state_in_tree(state)) {
 883                         node = &state->rb_node;
 884                         goto hit_next;
 885                 }
 886         }
 887         /*
 888          * this search will find all the extents that end after
 889          * our range starts.
 890          */
 891         node = tree_search_for_insert(tree, start, &p, &parent);
 892         if (!node) {
 893                 prealloc = alloc_extent_state_atomic(prealloc);
 894                 BUG_ON(!prealloc);
 895                 err = insert_state(tree, prealloc, start, end,
 896                                    &p, &parent, &bits, changeset);
 897                 if (err)
 898                         extent_io_tree_panic(tree, err);
 899
 900                 cache_state(prealloc, cached_state);
 901                 prealloc = NULL;
 902                 goto out;
 903         }
 904         state = rb_entry(node, struct extent_state, rb_node);
 905 hit_next:
 906         last_start = state->start;
 907         last_end = state->end;
 908
 909         /*
 910          * | ---- desired range ---- |
 911          * | state |
 912          *
 913          * Just lock what we found and keep going
 914          */
 915         if (state->start == start && state->end <= end) {
 916                 if (state->state & exclusive_bits) {
 917                         *failed_start = state->start;
 918                         err = -EEXIST;
 919                         goto out;
 920                 }
 921
 922                 set_state_bits(tree, state, &bits, changeset);
 923                 cache_state(state, cached_state);
 924                 merge_state(tree, state);
 925                 if (last_end == (u64)-1)
 926                         goto out;
 927                 start = last_end + 1;
 928                 state = next_state(state);
 929                 if (start < end && state && state->start == start &&
 930                     !need_resched())
 931                         goto hit_next;
 932                 goto search_again;
 933         }
 934
 935         /*
 936          *     | ---- desired range ---- |
 937          * | state |
 938          *   or
 939          * | ------------- state -------------- |
 940          *
 941          * We need to split the extent we found, and may flip bits on
 942          * second half.
 943          *
 944          * If the extent we found extends past our
 945          * range, we just split and search again.  It'll get split
 946          * again the next time though.
 947          *
 948          * If the extent we found is inside our range, we set the
 949          * desired bit on it.
 950          */
 951         if (state->start < start) {
 952                 if (state->state & exclusive_bits) {
 953                         *failed_start = start;
 954                         err = -EEXIST;
 955                         goto out;
 956                 }
 957
 958                 prealloc = alloc_extent_state_atomic(prealloc);
 959                 BUG_ON(!prealloc);
 960                 err = split_state(tree, state, prealloc, start);
 961                 if (err)
 962                         extent_io_tree_panic(tree, err);
 963
 964                 prealloc = NULL;
 965                 if (err)
 966                         goto out;
 967                 if (state->end <= end) {
 968                         set_state_bits(tree, state, &bits, changeset);
 969                         cache_state(state, cached_state);
 970                         merge_state(tree, state);
 971                         if (last_end == (u64)-1)
 972                                 goto out;
 973                         start = last_end + 1;
 974                         state = next_state(state);
 975                         if (start < end && state && state->start == start &&
 976                             !need_resched())
 977                                 goto hit_next;
 978                 }
 979                 goto search_again;
 980         }
 981         /*
 982          * | ---- desired range ---- |
 983          *     | state | or               | state |
 984          *
 985          * There's a hole, we need to insert something in it and
 986          * ignore the extent we found.
 987          */
 988         if (state->start > start) {
 989                 u64 this_end;
 990                 if (end < last_start)
 991                         this_end = end;
 992                 else
 993                         this_end = last_start - 1;
 994
 995                 prealloc = alloc_extent_state_atomic(prealloc);
 996                 BUG_ON(!prealloc);
 997
 998                 /*
 999                  * Avoid to free 'prealloc' if it can be merged with
1000                  * the later extent.
1001                  */
1002                 err = insert_state(tree, prealloc, start, this_end,
1003                                    NULL, NULL, &bits, changeset);
1004                 if (err)
1005                         extent_io_tree_panic(tree, err);
1006
1007                 cache_state(prealloc, cached_state);
1008                 prealloc = NULL;
1009                 start = this_end + 1;
1010                 goto search_again;
1011         }
1012         /*
1013          * | ---- desired range ---- |
1014          *                        | state |
1015          * We need to split the extent, and set the bit
1016          * on the first half
1017          */
1018         if (state->start <= end && state->end > end) {
1019                 if (state->state & exclusive_bits) {
1020                         *failed_start = start;
1021                         err = -EEXIST;
1022                         goto out;
1023                 }
1024
1025                 prealloc = alloc_extent_state_atomic(prealloc);
1026                 BUG_ON(!prealloc);
1027                 err = split_state(tree, state, prealloc, end + 1);
1028                 if (err)
1029                         extent_io_tree_panic(tree, err);
1030
1031                 set_state_bits(tree, prealloc, &bits, changeset);
1032                 cache_state(prealloc, cached_state);
1033                 merge_state(tree, prealloc);
1034                 prealloc = NULL;
1035                 goto out;
1036         }
1037
1038 search_again:
1039         if (start > end)
1040                 goto out;
1041         spin_unlock(&tree->lock);
1042         if (gfpflags_allow_blocking(mask))
1043                 cond_resched();
1044         goto again;
1045
1046 out:
1047         spin_unlock(&tree->lock);
1048         if (prealloc)
1049                 free_extent_state(prealloc);
1050
1051         return err;
1052
1053 }
1054
1055 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1056                    unsigned bits, u64 * failed_start,
1057                    struct extent_state **cached_state, gfp_t mask)
1058 {
1059         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1060                                 cached_state, mask, NULL);
1061 }
1062
1063
1064 /**
1065  * convert_extent_bit - convert all bits in a given range from one bit to
1066  *                      another
1067  * @tree:       the io tree to search
1068  * @start:      the start offset in bytes
1069  * @end:        the end offset in bytes (inclusive)
1070  * @bits:       the bits to set in this range
1071  * @clear_bits: the bits to clear in this range
1072  * @cached_state:       state that we're going to cache
1073  *
1074  * This will go through and set bits for the given range.  If any states exist
1075  * already in this range they are set with the given bit and cleared of the
1076  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1077  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1078  * boundary bits like LOCK.
1079  *
1080  * All allocations are done with GFP_NOFS.
1081  */
1082 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1083                        unsigned bits, unsigned clear_bits,
1084                        struct extent_state **cached_state)
1085 {
1086         struct extent_state *state;
1087         struct extent_state *prealloc = NULL;
1088         struct rb_node *node;
1089         struct rb_node **p;
1090         struct rb_node *parent;
1091         int err = 0;
1092         u64 last_start;
1093         u64 last_end;
1094         bool first_iteration = true;
1095
1096         btrfs_debug_check_extent_io_range(tree, start, end);
1097
1098 again:
1099         if (!prealloc) {
1100                 /*
1101                  * Best effort, don't worry if extent state allocation fails
1102                  * here for the first iteration. We might have a cached state
1103                  * that matches exactly the target range, in which case no
1104                  * extent state allocations are needed. We'll only know this
1105                  * after locking the tree.
1106                  */
1107                 prealloc = alloc_extent_state(GFP_NOFS);
1108                 if (!prealloc && !first_iteration)
1109                         return -ENOMEM;
1110         }
1111
1112         spin_lock(&tree->lock);
1113         if (cached_state && *cached_state) {
1114                 state = *cached_state;
1115                 if (state->start <= start && state->end > start &&
1116                     extent_state_in_tree(state)) {
1117                         node = &state->rb_node;
1118                         goto hit_next;
1119                 }
1120         }
1121
1122         /*
1123          * this search will find all the extents that end after
1124          * our range starts.
1125          */
1126         node = tree_search_for_insert(tree, start, &p, &parent);
1127         if (!node) {
1128                 prealloc = alloc_extent_state_atomic(prealloc);
1129                 if (!prealloc) {
1130                         err = -ENOMEM;
1131                         goto out;
1132                 }
1133                 err = insert_state(tree, prealloc, start, end,
1134                                    &p, &parent, &bits, NULL);
1135                 if (err)
1136                         extent_io_tree_panic(tree, err);
1137                 cache_state(prealloc, cached_state);
1138                 prealloc = NULL;
1139                 goto out;
1140         }
1141         state = rb_entry(node, struct extent_state, rb_node);
1142 hit_next:
1143         last_start = state->start;
1144         last_end = state->end;
1145
1146         /*
1147          * | ---- desired range ---- |
1148          * | state |
1149          *
1150          * Just lock what we found and keep going
1151          */
1152         if (state->start == start && state->end <= end) {
1153                 set_state_bits(tree, state, &bits, NULL);
1154                 cache_state(state, cached_state);
1155                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1156                 if (last_end == (u64)-1)
1157                         goto out;
1158                 start = last_end + 1;
1159                 if (start < end && state && state->start == start &&
1160                     !need_resched())
1161                         goto hit_next;
1162                 goto search_again;
1163         }
1164
1165         /*
1166          *     | ---- desired range ---- |
1167          * | state |
1168          *   or
1169          * | ------------- state -------------- |
1170          *
1171          * We need to split the extent we found, and may flip bits on
1172          * second half.
1173          *
1174          * If the extent we found extends past our
1175          * range, we just split and search again.  It'll get split
1176          * again the next time though.
1177          *
1178          * If the extent we found is inside our range, we set the
1179          * desired bit on it.
1180          */
1181         if (state->start < start) {
1182                 prealloc = alloc_extent_state_atomic(prealloc);
1183                 if (!prealloc) {
1184                         err = -ENOMEM;
1185                         goto out;
1186                 }
1187                 err = split_state(tree, state, prealloc, start);
1188                 if (err)
1189                         extent_io_tree_panic(tree, err);
1190                 prealloc = NULL;
1191                 if (err)
1192                         goto out;
1193                 if (state->end <= end) {
1194                         set_state_bits(tree, state, &bits, NULL);
1195                         cache_state(state, cached_state);
1196                         state = clear_state_bit(tree, state, &clear_bits, 0,
1197                                                 NULL);
1198                         if (last_end == (u64)-1)
1199                                 goto out;
1200                         start = last_end + 1;
1201                         if (start < end && state && state->start == start &&
1202                             !need_resched())
1203                                 goto hit_next;
1204                 }
1205                 goto search_again;
1206         }
1207         /*
1208          * | ---- desired range ---- |
1209          *     | state | or               | state |
1210          *
1211          * There's a hole, we need to insert something in it and
1212          * ignore the extent we found.
1213          */
1214         if (state->start > start) {
1215                 u64 this_end;
1216                 if (end < last_start)
1217                         this_end = end;
1218                 else
1219                         this_end = last_start - 1;
1220
1221                 prealloc = alloc_extent_state_atomic(prealloc);
1222                 if (!prealloc) {
1223                         err = -ENOMEM;
1224                         goto out;
1225                 }
1226
1227                 /*
1228                  * Avoid to free 'prealloc' if it can be merged with
1229                  * the later extent.
1230                  */
1231                 err = insert_state(tree, prealloc, start, this_end,
1232                                    NULL, NULL, &bits, NULL);
1233                 if (err)
1234                         extent_io_tree_panic(tree, err);
1235                 cache_state(prealloc, cached_state);
1236                 prealloc = NULL;
1237                 start = this_end + 1;
1238                 goto search_again;
1239         }
1240         /*
1241          * | ---- desired range ---- |
1242          *                        | state |
1243          * We need to split the extent, and set the bit
1244          * on the first half
1245          */
1246         if (state->start <= end && state->end > end) {
1247                 prealloc = alloc_extent_state_atomic(prealloc);
1248                 if (!prealloc) {
1249                         err = -ENOMEM;
1250                         goto out;
1251                 }
1252
1253                 err = split_state(tree, state, prealloc, end + 1);
1254                 if (err)
1255                         extent_io_tree_panic(tree, err);
1256
1257                 set_state_bits(tree, prealloc, &bits, NULL);
1258                 cache_state(prealloc, cached_state);
1259                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1260                 prealloc = NULL;
1261                 goto out;
1262         }
1263
1264 search_again:
1265         if (start > end)
1266                 goto out;
1267         spin_unlock(&tree->lock);
1268         cond_resched();
1269         first_iteration = false;
1270         goto again;
1271
1272 out:
1273         spin_unlock(&tree->lock);
1274         if (prealloc)
1275                 free_extent_state(prealloc);
1276
1277         return err;
1278 }
1279
1280 /* wrappers around set/clear extent bit */
1281 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1282                            unsigned bits, struct extent_changeset *changeset)
1283 {
1284         /*
1285          * We don't support EXTENT_LOCKED yet, as current changeset will
1286          * record any bits changed, so for EXTENT_LOCKED case, it will
1287          * either fail with -EEXIST or changeset will record the whole
1288          * range.
1289          */
1290         BUG_ON(bits & EXTENT_LOCKED);
1291
1292         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1293                                 changeset);
1294 }
1295
1296 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1297                      unsigned bits, int wake, int delete,
1298                      struct extent_state **cached)
1299 {
1300         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1301                                   cached, GFP_NOFS, NULL);
1302 }
1303
1304 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1305                 unsigned bits, struct extent_changeset *changeset)
1306 {
1307         /*
1308          * Don't support EXTENT_LOCKED case, same reason as
1309          * set_record_extent_bits().
1310          */
1311         BUG_ON(bits & EXTENT_LOCKED);
1312
1313         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1314                                   changeset);
1315 }
1316
1317 /*
1318  * either insert or lock state struct between start and end use mask to tell
1319  * us if waiting is desired.
1320  */
1321 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1322                      struct extent_state **cached_state)
1323 {
1324         int err;
1325         u64 failed_start;
1326
1327         while (1) {
1328                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1329                                        EXTENT_LOCKED, &failed_start,
1330                                        cached_state, GFP_NOFS, NULL);
1331                 if (err == -EEXIST) {
1332                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1333                         start = failed_start;
1334                 } else
1335                         break;
1336                 WARN_ON(start > end);
1337         }
1338         return err;
1339 }
1340
1341 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1342 {
1343         int err;
1344         u64 failed_start;
1345
1346         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1347                                &failed_start, NULL, GFP_NOFS, NULL);
1348         if (err == -EEXIST) {
1349                 if (failed_start > start)
1350                         clear_extent_bit(tree, start, failed_start - 1,
1351                                          EXTENT_LOCKED, 1, 0, NULL);
1352                 return 0;
1353         }
1354         return 1;
1355 }
1356
1357 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1358 {
1359         unsigned long index = start >> PAGE_SHIFT;
1360         unsigned long end_index = end >> PAGE_SHIFT;
1361         struct page *page;
1362
1363         while (index <= end_index) {
1364                 page = find_get_page(inode->i_mapping, index);
1365                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1366                 clear_page_dirty_for_io(page);
1367                 put_page(page);
1368                 index++;
1369         }
1370 }
1371
1372 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1373 {
1374         unsigned long index = start >> PAGE_SHIFT;
1375         unsigned long end_index = end >> PAGE_SHIFT;
1376         struct page *page;
1377
1378         while (index <= end_index) {
1379                 page = find_get_page(inode->i_mapping, index);
1380                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1381                 __set_page_dirty_nobuffers(page);
1382                 account_page_redirty(page);
1383                 put_page(page);
1384                 index++;
1385         }
1386 }
1387
1388 /*
1389  * helper function to set both pages and extents in the tree writeback
1390  */
1391 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1392 {
1393         tree->ops->set_range_writeback(tree->private_data, start, end);
1394 }
1395
1396 /* find the first state struct with 'bits' set after 'start', and
1397  * return it.  tree->lock must be held.  NULL will returned if
1398  * nothing was found after 'start'
1399  */
1400 static struct extent_state *
1401 find_first_extent_bit_state(struct extent_io_tree *tree,
1402                             u64 start, unsigned bits)
1403 {
1404         struct rb_node *node;
1405         struct extent_state *state;
1406
1407         /*
1408          * this search will find all the extents that end after
1409          * our range starts.
1410          */
1411         node = tree_search(tree, start);
1412         if (!node)
1413                 goto out;
1414
1415         while (1) {
1416                 state = rb_entry(node, struct extent_state, rb_node);
1417                 if (state->end >= start && (state->state & bits))
1418                         return state;
1419
1420                 node = rb_next(node);
1421                 if (!node)
1422                         break;
1423         }
1424 out:
1425         return NULL;
1426 }
1427
1428 /*
1429  * find the first offset in the io tree with 'bits' set. zero is
1430  * returned if we find something, and *start_ret and *end_ret are
1431  * set to reflect the state struct that was found.
1432  *
1433  * If nothing was found, 1 is returned. If found something, return 0.
1434  */
1435 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1436                           u64 *start_ret, u64 *end_ret, unsigned bits,
1437                           struct extent_state **cached_state)
1438 {
1439         struct extent_state *state;
1440         struct rb_node *n;
1441         int ret = 1;
1442
1443         spin_lock(&tree->lock);
1444         if (cached_state && *cached_state) {
1445                 state = *cached_state;
1446                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1447                         n = rb_next(&state->rb_node);
1448                         while (n) {
1449                                 state = rb_entry(n, struct extent_state,
1450                                                  rb_node);
1451                                 if (state->state & bits)
1452                                         goto got_it;
1453                                 n = rb_next(n);
1454                         }
1455                         free_extent_state(*cached_state);
1456                         *cached_state = NULL;
1457                         goto out;
1458                 }
1459                 free_extent_state(*cached_state);
1460                 *cached_state = NULL;
1461         }
1462
1463         state = find_first_extent_bit_state(tree, start, bits);
1464 got_it:
1465         if (state) {
1466                 cache_state_if_flags(state, cached_state, 0);
1467                 *start_ret = state->start;
1468                 *end_ret = state->end;
1469                 ret = 0;
1470         }
1471 out:
1472         spin_unlock(&tree->lock);
1473         return ret;
1474 }
1475
1476 /*
1477  * find a contiguous range of bytes in the file marked as delalloc, not
1478  * more than 'max_bytes'.  start and end are used to return the range,
1479  *
1480  * 1 is returned if we find something, 0 if nothing was in the tree
1481  */
1482 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1483                                         u64 *start, u64 *end, u64 max_bytes,
1484                                         struct extent_state **cached_state)
1485 {
1486         struct rb_node *node;
1487         struct extent_state *state;
1488         u64 cur_start = *start;
1489         u64 found = 0;
1490         u64 total_bytes = 0;
1491
1492         spin_lock(&tree->lock);
1493
1494         /*
1495          * this search will find all the extents that end after
1496          * our range starts.
1497          */
1498         node = tree_search(tree, cur_start);
1499         if (!node) {
1500                 if (!found)
1501                         *end = (u64)-1;
1502                 goto out;
1503         }
1504
1505         while (1) {
1506                 state = rb_entry(node, struct extent_state, rb_node);
1507                 if (found && (state->start != cur_start ||
1508                               (state->state & EXTENT_BOUNDARY))) {
1509                         goto out;
1510                 }
1511                 if (!(state->state & EXTENT_DELALLOC)) {
1512                         if (!found)
1513                                 *end = state->end;
1514                         goto out;
1515                 }
1516                 if (!found) {
1517                         *start = state->start;
1518                         *cached_state = state;
1519                         refcount_inc(&state->refs);
1520                 }
1521                 found++;
1522                 *end = state->end;
1523                 cur_start = state->end + 1;
1524                 node = rb_next(node);
1525                 total_bytes += state->end - state->start + 1;
1526                 if (total_bytes >= max_bytes)
1527                         break;
1528                 if (!node)
1529                         break;
1530         }
1531 out:
1532         spin_unlock(&tree->lock);
1533         return found;
1534 }
1535
1536 static int __process_pages_contig(struct address_space *mapping,
1537                                   struct page *locked_page,
1538                                   pgoff_t start_index, pgoff_t end_index,
1539                                   unsigned long page_ops, pgoff_t *index_ret);
1540
1541 static noinline void __unlock_for_delalloc(struct inode *inode,
1542                                            struct page *locked_page,
1543                                            u64 start, u64 end)
1544 {
1545         unsigned long index = start >> PAGE_SHIFT;
1546         unsigned long end_index = end >> PAGE_SHIFT;
1547
1548         ASSERT(locked_page);
1549         if (index == locked_page->index && end_index == index)
1550                 return;
1551
1552         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1553                                PAGE_UNLOCK, NULL);
1554 }
1555
1556 static noinline int lock_delalloc_pages(struct inode *inode,
1557                                         struct page *locked_page,
1558                                         u64 delalloc_start,
1559                                         u64 delalloc_end)
1560 {
1561         unsigned long index = delalloc_start >> PAGE_SHIFT;
1562         unsigned long index_ret = index;
1563         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1564         int ret;
1565
1566         ASSERT(locked_page);
1567         if (index == locked_page->index && index == end_index)
1568                 return 0;
1569
1570         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1571                                      end_index, PAGE_LOCK, &index_ret);
1572         if (ret == -EAGAIN)
1573                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1574                                       (u64)index_ret << PAGE_SHIFT);
1575         return ret;
1576 }
1577
1578 /*
1579  * find a contiguous range of bytes in the file marked as delalloc, not
1580  * more than 'max_bytes'.  start and end are used to return the range,
1581  *
1582  * 1 is returned if we find something, 0 if nothing was in the tree
1583  */
1584 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1585                                     struct extent_io_tree *tree,
1586                                     struct page *locked_page, u64 *start,
1587                                     u64 *end, u64 max_bytes)
1588 {
1589         u64 delalloc_start;
1590         u64 delalloc_end;
1591         u64 found;
1592         struct extent_state *cached_state = NULL;
1593         int ret;
1594         int loops = 0;
1595
1596 again:
1597         /* step one, find a bunch of delalloc bytes starting at start */
1598         delalloc_start = *start;
1599         delalloc_end = 0;
1600         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1601                                     max_bytes, &cached_state);
1602         if (!found || delalloc_end <= *start) {
1603                 *start = delalloc_start;
1604                 *end = delalloc_end;
1605                 free_extent_state(cached_state);
1606                 return 0;
1607         }
1608
1609         /*
1610          * start comes from the offset of locked_page.  We have to lock
1611          * pages in order, so we can't process delalloc bytes before
1612          * locked_page
1613          */
1614         if (delalloc_start < *start)
1615                 delalloc_start = *start;
1616
1617         /*
1618          * make sure to limit the number of pages we try to lock down
1619          */
1620         if (delalloc_end + 1 - delalloc_start > max_bytes)
1621                 delalloc_end = delalloc_start + max_bytes - 1;
1622
1623         /* step two, lock all the pages after the page that has start */
1624         ret = lock_delalloc_pages(inode, locked_page,
1625                                   delalloc_start, delalloc_end);
1626         if (ret == -EAGAIN) {
1627                 /* some of the pages are gone, lets avoid looping by
1628                  * shortening the size of the delalloc range we're searching
1629                  */
1630                 free_extent_state(cached_state);
1631                 cached_state = NULL;
1632                 if (!loops) {
1633                         max_bytes = PAGE_SIZE;
1634                         loops = 1;
1635                         goto again;
1636                 } else {
1637                         found = 0;
1638                         goto out_failed;
1639                 }
1640         }
1641         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1642
1643         /* step three, lock the state bits for the whole range */
1644         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1645
1646         /* then test to make sure it is all still delalloc */
1647         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1648                              EXTENT_DELALLOC, 1, cached_state);
1649         if (!ret) {
1650                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1651                                      &cached_state, GFP_NOFS);
1652                 __unlock_for_delalloc(inode, locked_page,
1653                               delalloc_start, delalloc_end);
1654                 cond_resched();
1655                 goto again;
1656         }
1657         free_extent_state(cached_state);
1658         *start = delalloc_start;
1659         *end = delalloc_end;
1660 out_failed:
1661         return found;
1662 }
1663
1664 static int __process_pages_contig(struct address_space *mapping,
1665                                   struct page *locked_page,
1666                                   pgoff_t start_index, pgoff_t end_index,
1667                                   unsigned long page_ops, pgoff_t *index_ret)
1668 {
1669         unsigned long nr_pages = end_index - start_index + 1;
1670         unsigned long pages_locked = 0;
1671         pgoff_t index = start_index;
1672         struct page *pages[16];
1673         unsigned ret;
1674         int err = 0;
1675         int i;
1676
1677         if (page_ops & PAGE_LOCK) {
1678                 ASSERT(page_ops == PAGE_LOCK);
1679                 ASSERT(index_ret && *index_ret == start_index);
1680         }
1681
1682         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1683                 mapping_set_error(mapping, -EIO);
1684
1685         while (nr_pages > 0) {
1686                 ret = find_get_pages_contig(mapping, index,
1687                                      min_t(unsigned long,
1688                                      nr_pages, ARRAY_SIZE(pages)), pages);
1689                 if (ret == 0) {
1690                         /*
1691                          * Only if we're going to lock these pages,
1692                          * can we find nothing at @index.
1693                          */
1694                         ASSERT(page_ops & PAGE_LOCK);
1695                         err = -EAGAIN;
1696                         goto out;
1697                 }
1698
1699                 for (i = 0; i < ret; i++) {
1700                         if (page_ops & PAGE_SET_PRIVATE2)
1701                                 SetPagePrivate2(pages[i]);
1702
1703                         if (pages[i] == locked_page) {
1704                                 put_page(pages[i]);
1705                                 pages_locked++;
1706                                 continue;
1707                         }
1708                         if (page_ops & PAGE_CLEAR_DIRTY)
1709                                 clear_page_dirty_for_io(pages[i]);
1710                         if (page_ops & PAGE_SET_WRITEBACK)
1711                                 set_page_writeback(pages[i]);
1712                         if (page_ops & PAGE_SET_ERROR)
1713                                 SetPageError(pages[i]);
1714                         if (page_ops & PAGE_END_WRITEBACK)
1715                                 end_page_writeback(pages[i]);
1716                         if (page_ops & PAGE_UNLOCK)
1717                                 unlock_page(pages[i]);
1718                         if (page_ops & PAGE_LOCK) {
1719                                 lock_page(pages[i]);
1720                                 if (!PageDirty(pages[i]) ||
1721                                     pages[i]->mapping != mapping) {
1722                                         unlock_page(pages[i]);
1723                                         put_page(pages[i]);
1724                                         err = -EAGAIN;
1725                                         goto out;
1726                                 }
1727                         }
1728                         put_page(pages[i]);
1729                         pages_locked++;
1730                 }
1731                 nr_pages -= ret;
1732                 index += ret;
1733                 cond_resched();
1734         }
1735 out:
1736         if (err && index_ret)
1737                 *index_ret = start_index + pages_locked - 1;
1738         return err;
1739 }
1740
1741 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1742                                  u64 delalloc_end, struct page *locked_page,
1743                                  unsigned clear_bits,
1744                                  unsigned long page_ops)
1745 {
1746         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1747                          NULL);
1748
1749         __process_pages_contig(inode->i_mapping, locked_page,
1750                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1751                                page_ops, NULL);
1752 }
1753
1754 /*
1755  * count the number of bytes in the tree that have a given bit(s)
1756  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1757  * cached.  The total number found is returned.
1758  */
1759 u64 count_range_bits(struct extent_io_tree *tree,
1760                      u64 *start, u64 search_end, u64 max_bytes,
1761                      unsigned bits, int contig)
1762 {
1763         struct rb_node *node;
1764         struct extent_state *state;
1765         u64 cur_start = *start;
1766         u64 total_bytes = 0;
1767         u64 last = 0;
1768         int found = 0;
1769
1770         if (WARN_ON(search_end <= cur_start))
1771                 return 0;
1772
1773         spin_lock(&tree->lock);
1774         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1775                 total_bytes = tree->dirty_bytes;
1776                 goto out;
1777         }
1778         /*
1779          * this search will find all the extents that end after
1780          * our range starts.
1781          */
1782         node = tree_search(tree, cur_start);
1783         if (!node)
1784                 goto out;
1785
1786         while (1) {
1787                 state = rb_entry(node, struct extent_state, rb_node);
1788                 if (state->start > search_end)
1789                         break;
1790                 if (contig && found && state->start > last + 1)
1791                         break;
1792                 if (state->end >= cur_start && (state->state & bits) == bits) {
1793                         total_bytes += min(search_end, state->end) + 1 -
1794                                        max(cur_start, state->start);
1795                         if (total_bytes >= max_bytes)
1796                                 break;
1797                         if (!found) {
1798                                 *start = max(cur_start, state->start);
1799                                 found = 1;
1800                         }
1801                         last = state->end;
1802                 } else if (contig && found) {
1803                         break;
1804                 }
1805                 node = rb_next(node);
1806                 if (!node)
1807                         break;
1808         }
1809 out:
1810         spin_unlock(&tree->lock);
1811         return total_bytes;
1812 }
1813
1814 /*
1815  * set the private field for a given byte offset in the tree.  If there isn't
1816  * an extent_state there already, this does nothing.
1817  */
1818 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1819                 struct io_failure_record *failrec)
1820 {
1821         struct rb_node *node;
1822         struct extent_state *state;
1823         int ret = 0;
1824
1825         spin_lock(&tree->lock);
1826         /*
1827          * this search will find all the extents that end after
1828          * our range starts.
1829          */
1830         node = tree_search(tree, start);
1831         if (!node) {
1832                 ret = -ENOENT;
1833                 goto out;
1834         }
1835         state = rb_entry(node, struct extent_state, rb_node);
1836         if (state->start != start) {
1837                 ret = -ENOENT;
1838                 goto out;
1839         }
1840         state->failrec = failrec;
1841 out:
1842         spin_unlock(&tree->lock);
1843         return ret;
1844 }
1845
1846 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1847                 struct io_failure_record **failrec)
1848 {
1849         struct rb_node *node;
1850         struct extent_state *state;
1851         int ret = 0;
1852
1853         spin_lock(&tree->lock);
1854         /*
1855          * this search will find all the extents that end after
1856          * our range starts.
1857          */
1858         node = tree_search(tree, start);
1859         if (!node) {
1860                 ret = -ENOENT;
1861                 goto out;
1862         }
1863         state = rb_entry(node, struct extent_state, rb_node);
1864         if (state->start != start) {
1865                 ret = -ENOENT;
1866                 goto out;
1867         }
1868         *failrec = state->failrec;
1869 out:
1870         spin_unlock(&tree->lock);
1871         return ret;
1872 }
1873
1874 /*
1875  * searches a range in the state tree for a given mask.
1876  * If 'filled' == 1, this returns 1 only if every extent in the tree
1877  * has the bits set.  Otherwise, 1 is returned if any bit in the
1878  * range is found set.
1879  */
1880 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1881                    unsigned bits, int filled, struct extent_state *cached)
1882 {
1883         struct extent_state *state = NULL;
1884         struct rb_node *node;
1885         int bitset = 0;
1886
1887         spin_lock(&tree->lock);
1888         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1889             cached->end > start)
1890                 node = &cached->rb_node;
1891         else
1892                 node = tree_search(tree, start);
1893         while (node && start <= end) {
1894                 state = rb_entry(node, struct extent_state, rb_node);
1895
1896                 if (filled && state->start > start) {
1897                         bitset = 0;
1898                         break;
1899                 }
1900
1901                 if (state->start > end)
1902                         break;
1903
1904                 if (state->state & bits) {
1905                         bitset = 1;
1906                         if (!filled)
1907                                 break;
1908                 } else if (filled) {
1909                         bitset = 0;
1910                         break;
1911                 }
1912
1913                 if (state->end == (u64)-1)
1914                         break;
1915
1916                 start = state->end + 1;
1917                 if (start > end)
1918                         break;
1919                 node = rb_next(node);
1920                 if (!node) {
1921                         if (filled)
1922                                 bitset = 0;
1923                         break;
1924                 }
1925         }
1926         spin_unlock(&tree->lock);
1927         return bitset;
1928 }
1929
1930 /*
1931  * helper function to set a given page up to date if all the
1932  * extents in the tree for that page are up to date
1933  */
1934 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1935 {
1936         u64 start = page_offset(page);
1937         u64 end = start + PAGE_SIZE - 1;
1938         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1939                 SetPageUptodate(page);
1940 }
1941
1942 int free_io_failure(struct extent_io_tree *failure_tree,
1943                     struct extent_io_tree *io_tree,
1944                     struct io_failure_record *rec)
1945 {
1946         int ret;
1947         int err = 0;
1948
1949         set_state_failrec(failure_tree, rec->start, NULL);
1950         ret = clear_extent_bits(failure_tree, rec->start,
1951                                 rec->start + rec->len - 1,
1952                                 EXTENT_LOCKED | EXTENT_DIRTY);
1953         if (ret)
1954                 err = ret;
1955
1956         ret = clear_extent_bits(io_tree, rec->start,
1957                                 rec->start + rec->len - 1,
1958                                 EXTENT_DAMAGED);
1959         if (ret && !err)
1960                 err = ret;
1961
1962         kfree(rec);
1963         return err;
1964 }
1965
1966 /*
1967  * this bypasses the standard btrfs submit functions deliberately, as
1968  * the standard behavior is to write all copies in a raid setup. here we only
1969  * want to write the one bad copy. so we do the mapping for ourselves and issue
1970  * submit_bio directly.
1971  * to avoid any synchronization issues, wait for the data after writing, which
1972  * actually prevents the read that triggered the error from finishing.
1973  * currently, there can be no more than two copies of every data bit. thus,
1974  * exactly one rewrite is required.
1975  */
1976 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
1977                       u64 length, u64 logical, struct page *page,
1978                       unsigned int pg_offset, int mirror_num)
1979 {
1980         struct bio *bio;
1981         struct btrfs_device *dev;
1982         u64 map_length = 0;
1983         u64 sector;
1984         struct btrfs_bio *bbio = NULL;
1985         int ret;
1986
1987         ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
1988         BUG_ON(!mirror_num);
1989
1990         bio = btrfs_io_bio_alloc(1);
1991         bio->bi_iter.bi_size = 0;
1992         map_length = length;
1993
1994         /*
1995          * Avoid races with device replace and make sure our bbio has devices
1996          * associated to its stripes that don't go away while we are doing the
1997          * read repair operation.
1998          */
1999         btrfs_bio_counter_inc_blocked(fs_info);
2000         if (btrfs_is_parity_mirror(fs_info, logical, length)) {
2001                 /*
2002                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2003                  * to update all raid stripes, but here we just want to correct
2004                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2005                  * stripe's dev and sector.
2006                  */
2007                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2008                                       &map_length, &bbio, 0);
2009                 if (ret) {
2010                         btrfs_bio_counter_dec(fs_info);
2011                         bio_put(bio);
2012                         return -EIO;
2013                 }
2014                 ASSERT(bbio->mirror_num == 1);
2015         } else {
2016                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2017                                       &map_length, &bbio, mirror_num);
2018                 if (ret) {
2019                         btrfs_bio_counter_dec(fs_info);
2020                         bio_put(bio);
2021                         return -EIO;
2022                 }
2023                 BUG_ON(mirror_num != bbio->mirror_num);
2024         }
2025
2026         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2027         bio->bi_iter.bi_sector = sector;
2028         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2029         btrfs_put_bbio(bbio);
2030         if (!dev || !dev->bdev ||
2031             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2032                 btrfs_bio_counter_dec(fs_info);
2033                 bio_put(bio);
2034                 return -EIO;
2035         }
2036         bio_set_dev(bio, dev->bdev);
2037         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2038         bio_add_page(bio, page, length, pg_offset);
2039
2040         if (btrfsic_submit_bio_wait(bio)) {
2041                 /* try to remap that extent elsewhere? */
2042                 btrfs_bio_counter_dec(fs_info);
2043                 bio_put(bio);
2044                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2045                 return -EIO;
2046         }
2047
2048         btrfs_info_rl_in_rcu(fs_info,
2049                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2050                                   ino, start,
2051                                   rcu_str_deref(dev->name), sector);
2052         btrfs_bio_counter_dec(fs_info);
2053         bio_put(bio);
2054         return 0;
2055 }
2056
2057 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2058                          struct extent_buffer *eb, int mirror_num)
2059 {
2060         u64 start = eb->start;
2061         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2062         int ret = 0;
2063
2064         if (sb_rdonly(fs_info->sb))
2065                 return -EROFS;
2066
2067         for (i = 0; i < num_pages; i++) {
2068                 struct page *p = eb->pages[i];
2069
2070                 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2071                                         start - page_offset(p), mirror_num);
2072                 if (ret)
2073                         break;
2074                 start += PAGE_SIZE;
2075         }
2076
2077         return ret;
2078 }
2079
2080 /*
2081  * each time an IO finishes, we do a fast check in the IO failure tree
2082  * to see if we need to process or clean up an io_failure_record
2083  */
2084 int clean_io_failure(struct btrfs_fs_info *fs_info,
2085                      struct extent_io_tree *failure_tree,
2086                      struct extent_io_tree *io_tree, u64 start,
2087                      struct page *page, u64 ino, unsigned int pg_offset)
2088 {
2089         u64 private;
2090         struct io_failure_record *failrec;
2091         struct extent_state *state;
2092         int num_copies;
2093         int ret;
2094
2095         private = 0;
2096         ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2097                                EXTENT_DIRTY, 0);
2098         if (!ret)
2099                 return 0;
2100
2101         ret = get_state_failrec(failure_tree, start, &failrec);
2102         if (ret)
2103                 return 0;
2104
2105         BUG_ON(!failrec->this_mirror);
2106
2107         if (failrec->in_validation) {
2108                 /* there was no real error, just free the record */
2109                 btrfs_debug(fs_info,
2110                         "clean_io_failure: freeing dummy error at %llu",
2111                         failrec->start);
2112                 goto out;
2113         }
2114         if (sb_rdonly(fs_info->sb))
2115                 goto out;
2116
2117         spin_lock(&io_tree->lock);
2118         state = find_first_extent_bit_state(io_tree,
2119                                             failrec->start,
2120                                             EXTENT_LOCKED);
2121         spin_unlock(&io_tree->lock);
2122
2123         if (state && state->start <= failrec->start &&
2124             state->end >= failrec->start + failrec->len - 1) {
2125                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2126                                               failrec->len);
2127                 if (num_copies > 1)  {
2128                         repair_io_failure(fs_info, ino, start, failrec->len,
2129                                           failrec->logical, page, pg_offset,
2130                                           failrec->failed_mirror);
2131                 }
2132         }
2133
2134 out:
2135         free_io_failure(failure_tree, io_tree, failrec);
2136
2137         return 0;
2138 }
2139
2140 /*
2141  * Can be called when
2142  * - hold extent lock
2143  * - under ordered extent
2144  * - the inode is freeing
2145  */
2146 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2147 {
2148         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2149         struct io_failure_record *failrec;
2150         struct extent_state *state, *next;
2151
2152         if (RB_EMPTY_ROOT(&failure_tree->state))
2153                 return;
2154
2155         spin_lock(&failure_tree->lock);
2156         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2157         while (state) {
2158                 if (state->start > end)
2159                         break;
2160
2161                 ASSERT(state->end <= end);
2162
2163                 next = next_state(state);
2164
2165                 failrec = state->failrec;
2166                 free_extent_state(state);
2167                 kfree(failrec);
2168
2169                 state = next;
2170         }
2171         spin_unlock(&failure_tree->lock);
2172 }
2173
2174 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2175                 struct io_failure_record **failrec_ret)
2176 {
2177         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2178         struct io_failure_record *failrec;
2179         struct extent_map *em;
2180         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2181         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2182         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2183         int ret;
2184         u64 logical;
2185
2186         ret = get_state_failrec(failure_tree, start, &failrec);
2187         if (ret) {
2188                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2189                 if (!failrec)
2190                         return -ENOMEM;
2191
2192                 failrec->start = start;
2193                 failrec->len = end - start + 1;
2194                 failrec->this_mirror = 0;
2195                 failrec->bio_flags = 0;
2196                 failrec->in_validation = 0;
2197
2198                 read_lock(&em_tree->lock);
2199                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2200                 if (!em) {
2201                         read_unlock(&em_tree->lock);
2202                         kfree(failrec);
2203                         return -EIO;
2204                 }
2205
2206                 if (em->start > start || em->start + em->len <= start) {
2207                         free_extent_map(em);
2208                         em = NULL;
2209                 }
2210                 read_unlock(&em_tree->lock);
2211                 if (!em) {
2212                         kfree(failrec);
2213                         return -EIO;
2214                 }
2215
2216                 logical = start - em->start;
2217                 logical = em->block_start + logical;
2218                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2219                         logical = em->block_start;
2220                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2221                         extent_set_compress_type(&failrec->bio_flags,
2222                                                  em->compress_type);
2223                 }
2224
2225                 btrfs_debug(fs_info,
2226                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2227                         logical, start, failrec->len);
2228
2229                 failrec->logical = logical;
2230                 free_extent_map(em);
2231
2232                 /* set the bits in the private failure tree */
2233                 ret = set_extent_bits(failure_tree, start, end,
2234                                         EXTENT_LOCKED | EXTENT_DIRTY);
2235                 if (ret >= 0)
2236                         ret = set_state_failrec(failure_tree, start, failrec);
2237                 /* set the bits in the inode's tree */
2238                 if (ret >= 0)
2239                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2240                 if (ret < 0) {
2241                         kfree(failrec);
2242                         return ret;
2243                 }
2244         } else {
2245                 btrfs_debug(fs_info,
2246                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2247                         failrec->logical, failrec->start, failrec->len,
2248                         failrec->in_validation);
2249                 /*
2250                  * when data can be on disk more than twice, add to failrec here
2251                  * (e.g. with a list for failed_mirror) to make
2252                  * clean_io_failure() clean all those errors at once.
2253                  */
2254         }
2255
2256         *failrec_ret = failrec;
2257
2258         return 0;
2259 }
2260
2261 bool btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2262                            struct io_failure_record *failrec, int failed_mirror)
2263 {
2264         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2265         int num_copies;
2266
2267         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2268         if (num_copies == 1) {
2269                 /*
2270                  * we only have a single copy of the data, so don't bother with
2271                  * all the retry and error correction code that follows. no
2272                  * matter what the error is, it is very likely to persist.
2273                  */
2274                 btrfs_debug(fs_info,
2275                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2276                         num_copies, failrec->this_mirror, failed_mirror);
2277                 return false;
2278         }
2279
2280         /*
2281          * there are two premises:
2282          *      a) deliver good data to the caller
2283          *      b) correct the bad sectors on disk
2284          */
2285         if (failed_bio->bi_vcnt > 1) {
2286                 /*
2287                  * to fulfill b), we need to know the exact failing sectors, as
2288                  * we don't want to rewrite any more than the failed ones. thus,
2289                  * we need separate read requests for the failed bio
2290                  *
2291                  * if the following BUG_ON triggers, our validation request got
2292                  * merged. we need separate requests for our algorithm to work.
2293                  */
2294                 BUG_ON(failrec->in_validation);
2295                 failrec->in_validation = 1;
2296                 failrec->this_mirror = failed_mirror;
2297         } else {
2298                 /*
2299                  * we're ready to fulfill a) and b) alongside. get a good copy
2300                  * of the failed sector and if we succeed, we have setup
2301                  * everything for repair_io_failure to do the rest for us.
2302                  */
2303                 if (failrec->in_validation) {
2304                         BUG_ON(failrec->this_mirror != failed_mirror);
2305                         failrec->in_validation = 0;
2306                         failrec->this_mirror = 0;
2307                 }
2308                 failrec->failed_mirror = failed_mirror;
2309                 failrec->this_mirror++;
2310                 if (failrec->this_mirror == failed_mirror)
2311                         failrec->this_mirror++;
2312         }
2313
2314         if (failrec->this_mirror > num_copies) {
2315                 btrfs_debug(fs_info,
2316                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2317                         num_copies, failrec->this_mirror, failed_mirror);
2318                 return false;
2319         }
2320
2321         return true;
2322 }
2323
2324
2325 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2326                                     struct io_failure_record *failrec,
2327                                     struct page *page, int pg_offset, int icsum,
2328                                     bio_end_io_t *endio_func, void *data)
2329 {
2330         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2331         struct bio *bio;
2332         struct btrfs_io_bio *btrfs_failed_bio;
2333         struct btrfs_io_bio *btrfs_bio;
2334
2335         bio = btrfs_io_bio_alloc(1);
2336         bio->bi_end_io = endio_func;
2337         bio->bi_iter.bi_sector = failrec->logical >> 9;
2338         bio_set_dev(bio, fs_info->fs_devices->latest_bdev);
2339         bio->bi_iter.bi_size = 0;
2340         bio->bi_private = data;
2341
2342         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2343         if (btrfs_failed_bio->csum) {
2344                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2345
2346                 btrfs_bio = btrfs_io_bio(bio);
2347                 btrfs_bio->csum = btrfs_bio->csum_inline;
2348                 icsum *= csum_size;
2349                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2350                        csum_size);
2351         }
2352
2353         bio_add_page(bio, page, failrec->len, pg_offset);
2354
2355         return bio;
2356 }
2357
2358 /*
2359  * this is a generic handler for readpage errors (default
2360  * readpage_io_failed_hook). if other copies exist, read those and write back
2361  * good data to the failed position. does not investigate in remapping the
2362  * failed extent elsewhere, hoping the device will be smart enough to do this as
2363  * needed
2364  */
2365
2366 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2367                               struct page *page, u64 start, u64 end,
2368                               int failed_mirror)
2369 {
2370         struct io_failure_record *failrec;
2371         struct inode *inode = page->mapping->host;
2372         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2373         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2374         struct bio *bio;
2375         int read_mode = 0;
2376         blk_status_t status;
2377         int ret;
2378
2379         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2380
2381         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2382         if (ret)
2383                 return ret;
2384
2385         if (!btrfs_check_repairable(inode, failed_bio, failrec,
2386                                     failed_mirror)) {
2387                 free_io_failure(failure_tree, tree, failrec);
2388                 return -EIO;
2389         }
2390
2391         if (failed_bio->bi_vcnt > 1)
2392                 read_mode |= REQ_FAILFAST_DEV;
2393
2394         phy_offset >>= inode->i_sb->s_blocksize_bits;
2395         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2396                                       start - page_offset(page),
2397                                       (int)phy_offset, failed_bio->bi_end_io,
2398                                       NULL);
2399         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2400
2401         btrfs_debug(btrfs_sb(inode->i_sb),
2402                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2403                 read_mode, failrec->this_mirror, failrec->in_validation);
2404
2405         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2406                                          failrec->bio_flags, 0);
2407         if (status) {
2408                 free_io_failure(failure_tree, tree, failrec);
2409                 bio_put(bio);
2410                 ret = blk_status_to_errno(status);
2411         }
2412
2413         return ret;
2414 }
2415
2416 /* lots and lots of room for performance fixes in the end_bio funcs */
2417
2418 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2419 {
2420         int uptodate = (err == 0);
2421         struct extent_io_tree *tree;
2422         int ret = 0;
2423
2424         tree = &BTRFS_I(page->mapping->host)->io_tree;
2425
2426         if (tree->ops && tree->ops->writepage_end_io_hook)
2427                 tree->ops->writepage_end_io_hook(page, start, end, NULL,
2428                                 uptodate);
2429
2430         if (!uptodate) {
2431                 ClearPageUptodate(page);
2432                 SetPageError(page);
2433                 ret = err < 0 ? err : -EIO;
2434                 mapping_set_error(page->mapping, ret);
2435         }
2436 }
2437
2438 /*
2439  * after a writepage IO is done, we need to:
2440  * clear the uptodate bits on error
2441  * clear the writeback bits in the extent tree for this IO
2442  * end_page_writeback if the page has no more pending IO
2443  *
2444  * Scheduling is not allowed, so the extent state tree is expected
2445  * to have one and only one object corresponding to this IO.
2446  */
2447 static void end_bio_extent_writepage(struct bio *bio)
2448 {
2449         int error = blk_status_to_errno(bio->bi_status);
2450         struct bio_vec *bvec;
2451         u64 start;
2452         u64 end;
2453         int i;
2454
2455         ASSERT(!bio_flagged(bio, BIO_CLONED));
2456         bio_for_each_segment_all(bvec, bio, i) {
2457                 struct page *page = bvec->bv_page;
2458                 struct inode *inode = page->mapping->host;
2459                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2460
2461                 /* We always issue full-page reads, but if some block
2462                  * in a page fails to read, blk_update_request() will
2463                  * advance bv_offset and adjust bv_len to compensate.
2464                  * Print a warning for nonzero offsets, and an error
2465                  * if they don't add up to a full page.  */
2466                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2467                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2468                                 btrfs_err(fs_info,
2469                                    "partial page write in btrfs with offset %u and length %u",
2470                                         bvec->bv_offset, bvec->bv_len);
2471                         else
2472                                 btrfs_info(fs_info,
2473                                    "incomplete page write in btrfs with offset %u and length %u",
2474                                         bvec->bv_offset, bvec->bv_len);
2475                 }
2476
2477                 start = page_offset(page);
2478                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2479
2480                 end_extent_writepage(page, error, start, end);
2481                 end_page_writeback(page);
2482         }
2483
2484         bio_put(bio);
2485 }
2486
2487 static void
2488 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2489                               int uptodate)
2490 {
2491         struct extent_state *cached = NULL;
2492         u64 end = start + len - 1;
2493
2494         if (uptodate && tree->track_uptodate)
2495                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2496         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2497 }
2498
2499 /*
2500  * after a readpage IO is done, we need to:
2501  * clear the uptodate bits on error
2502  * set the uptodate bits if things worked
2503  * set the page up to date if all extents in the tree are uptodate
2504  * clear the lock bit in the extent tree
2505  * unlock the page if there are no other extents locked for it
2506  *
2507  * Scheduling is not allowed, so the extent state tree is expected
2508  * to have one and only one object corresponding to this IO.
2509  */
2510 static void end_bio_extent_readpage(struct bio *bio)
2511 {
2512         struct bio_vec *bvec;
2513         int uptodate = !bio->bi_status;
2514         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2515         struct extent_io_tree *tree, *failure_tree;
2516         u64 offset = 0;
2517         u64 start;
2518         u64 end;
2519         u64 len;
2520         u64 extent_start = 0;
2521         u64 extent_len = 0;
2522         int mirror;
2523         int ret;
2524         int i;
2525
2526         ASSERT(!bio_flagged(bio, BIO_CLONED));
2527         bio_for_each_segment_all(bvec, bio, i) {
2528                 struct page *page = bvec->bv_page;
2529                 struct inode *inode = page->mapping->host;
2530                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2531
2532                 btrfs_debug(fs_info,
2533                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2534                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2535                         io_bio->mirror_num);
2536                 tree = &BTRFS_I(inode)->io_tree;
2537                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2538
2539                 /* We always issue full-page reads, but if some block
2540                  * in a page fails to read, blk_update_request() will
2541                  * advance bv_offset and adjust bv_len to compensate.
2542                  * Print a warning for nonzero offsets, and an error
2543                  * if they don't add up to a full page.  */
2544                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2545                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2546                                 btrfs_err(fs_info,
2547                                         "partial page read in btrfs with offset %u and length %u",
2548                                         bvec->bv_offset, bvec->bv_len);
2549                         else
2550                                 btrfs_info(fs_info,
2551                                         "incomplete page read in btrfs with offset %u and length %u",
2552                                         bvec->bv_offset, bvec->bv_len);
2553                 }
2554
2555                 start = page_offset(page);
2556                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2557                 len = bvec->bv_len;
2558
2559                 mirror = io_bio->mirror_num;
2560                 if (likely(uptodate && tree->ops)) {
2561                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2562                                                               page, start, end,
2563                                                               mirror);
2564                         if (ret)
2565                                 uptodate = 0;
2566                         else
2567                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2568                                                  failure_tree, tree, start,
2569                                                  page,
2570                                                  btrfs_ino(BTRFS_I(inode)), 0);
2571                 }
2572
2573                 if (likely(uptodate))
2574                         goto readpage_ok;
2575
2576                 if (tree->ops) {
2577                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2578                         if (ret == -EAGAIN) {
2579                                 /*
2580                                  * Data inode's readpage_io_failed_hook() always
2581                                  * returns -EAGAIN.
2582                                  *
2583                                  * The generic bio_readpage_error handles errors
2584                                  * the following way: If possible, new read
2585                                  * requests are created and submitted and will
2586                                  * end up in end_bio_extent_readpage as well (if
2587                                  * we're lucky, not in the !uptodate case). In
2588                                  * that case it returns 0 and we just go on with
2589                                  * the next page in our bio. If it can't handle
2590                                  * the error it will return -EIO and we remain
2591                                  * responsible for that page.
2592                                  */
2593                                 ret = bio_readpage_error(bio, offset, page,
2594                                                          start, end, mirror);
2595                                 if (ret == 0) {
2596                                         uptodate = !bio->bi_status;
2597                                         offset += len;
2598                                         continue;
2599                                 }
2600                         }
2601
2602                         /*
2603                          * metadata's readpage_io_failed_hook() always returns
2604                          * -EIO and fixes nothing.  -EIO is also returned if
2605                          * data inode error could not be fixed.
2606                          */
2607                         ASSERT(ret == -EIO);
2608                 }
2609 readpage_ok:
2610                 if (likely(uptodate)) {
2611                         loff_t i_size = i_size_read(inode);
2612                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2613                         unsigned off;
2614
2615                         /* Zero out the end if this page straddles i_size */
2616                         off = i_size & (PAGE_SIZE-1);
2617                         if (page->index == end_index && off)
2618                                 zero_user_segment(page, off, PAGE_SIZE);
2619                         SetPageUptodate(page);
2620                 } else {
2621                         ClearPageUptodate(page);
2622                         SetPageError(page);
2623                 }
2624                 unlock_page(page);
2625                 offset += len;
2626
2627                 if (unlikely(!uptodate)) {
2628                         if (extent_len) {
2629                                 endio_readpage_release_extent(tree,
2630                                                               extent_start,
2631                                                               extent_len, 1);
2632                                 extent_start = 0;
2633                                 extent_len = 0;
2634                         }
2635                         endio_readpage_release_extent(tree, start,
2636                                                       end - start + 1, 0);
2637                 } else if (!extent_len) {
2638                         extent_start = start;
2639                         extent_len = end + 1 - start;
2640                 } else if (extent_start + extent_len == start) {
2641                         extent_len += end + 1 - start;
2642                 } else {
2643                         endio_readpage_release_extent(tree, extent_start,
2644                                                       extent_len, uptodate);
2645                         extent_start = start;
2646                         extent_len = end + 1 - start;
2647                 }
2648         }
2649
2650         if (extent_len)
2651                 endio_readpage_release_extent(tree, extent_start, extent_len,
2652                                               uptodate);
2653         if (io_bio->end_io)
2654                 io_bio->end_io(io_bio, blk_status_to_errno(bio->bi_status));
2655         bio_put(bio);
2656 }
2657
2658 /*
2659  * Initialize the members up to but not including 'bio'. Use after allocating a
2660  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2661  * 'bio' because use of __GFP_ZERO is not supported.
2662  */
2663 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2664 {
2665         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2666 }
2667
2668 /*
2669  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2670  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2671  * for the appropriate container_of magic
2672  */
2673 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte)
2674 {
2675         struct bio *bio;
2676
2677         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, btrfs_bioset);
2678         bio_set_dev(bio, bdev);
2679         bio->bi_iter.bi_sector = first_byte >> 9;
2680         btrfs_io_bio_init(btrfs_io_bio(bio));
2681         return bio;
2682 }
2683
2684 struct bio *btrfs_bio_clone(struct bio *bio)
2685 {
2686         struct btrfs_io_bio *btrfs_bio;
2687         struct bio *new;
2688
2689         /* Bio allocation backed by a bioset does not fail */
2690         new = bio_clone_fast(bio, GFP_NOFS, btrfs_bioset);
2691         btrfs_bio = btrfs_io_bio(new);
2692         btrfs_io_bio_init(btrfs_bio);
2693         btrfs_bio->iter = bio->bi_iter;
2694         return new;
2695 }
2696
2697 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2698 {
2699         struct bio *bio;
2700
2701         /* Bio allocation backed by a bioset does not fail */
2702         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, btrfs_bioset);
2703         btrfs_io_bio_init(btrfs_io_bio(bio));
2704         return bio;
2705 }
2706
2707 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2708 {
2709         struct bio *bio;
2710         struct btrfs_io_bio *btrfs_bio;
2711
2712         /* this will never fail when it's backed by a bioset */
2713         bio = bio_clone_fast(orig, GFP_NOFS, btrfs_bioset);
2714         ASSERT(bio);
2715
2716         btrfs_bio = btrfs_io_bio(bio);
2717         btrfs_io_bio_init(btrfs_bio);
2718
2719         bio_trim(bio, offset >> 9, size >> 9);
2720         btrfs_bio->iter = bio->bi_iter;
2721         return bio;
2722 }
2723
2724 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2725                                        unsigned long bio_flags)
2726 {
2727         blk_status_t ret = 0;
2728         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2729         struct page *page = bvec->bv_page;
2730         struct extent_io_tree *tree = bio->bi_private;
2731         u64 start;
2732
2733         start = page_offset(page) + bvec->bv_offset;
2734
2735         bio->bi_private = NULL;
2736
2737         if (tree->ops)
2738                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
2739                                            mirror_num, bio_flags, start);
2740         else
2741                 btrfsic_submit_bio(bio);
2742
2743         return blk_status_to_errno(ret);
2744 }
2745
2746 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2747                      unsigned long offset, size_t size, struct bio *bio,
2748                      unsigned long bio_flags)
2749 {
2750         int ret = 0;
2751         if (tree->ops)
2752                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2753                                                 bio_flags);
2754         return ret;
2755
2756 }
2757
2758 /*
2759  * @opf:        bio REQ_OP_* and REQ_* flags as one value
2760  */
2761 static int submit_extent_page(unsigned int opf, struct extent_io_tree *tree,
2762                               struct writeback_control *wbc,
2763                               struct page *page, u64 offset,
2764                               size_t size, unsigned long pg_offset,
2765                               struct block_device *bdev,
2766                               struct bio **bio_ret,
2767                               bio_end_io_t end_io_func,
2768                               int mirror_num,
2769                               unsigned long prev_bio_flags,
2770                               unsigned long bio_flags,
2771                               bool force_bio_submit)
2772 {
2773         int ret = 0;
2774         struct bio *bio;
2775         int contig = 0;
2776         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2777         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2778         sector_t sector = offset >> 9;
2779
2780         if (bio_ret && *bio_ret) {
2781                 bio = *bio_ret;
2782                 if (old_compressed)
2783                         contig = bio->bi_iter.bi_sector == sector;
2784                 else
2785                         contig = bio_end_sector(bio) == sector;
2786
2787                 if (prev_bio_flags != bio_flags || !contig ||
2788                     force_bio_submit ||
2789                     merge_bio(tree, page, pg_offset, page_size, bio, bio_flags) ||
2790                     bio_add_page(bio, page, page_size, pg_offset) < page_size) {
2791                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2792                         if (ret < 0) {
2793                                 *bio_ret = NULL;
2794                                 return ret;
2795                         }
2796                         bio = NULL;
2797                 } else {
2798                         if (wbc)
2799                                 wbc_account_io(wbc, page, page_size);
2800                         return 0;
2801                 }
2802         }
2803
2804         bio = btrfs_bio_alloc(bdev, offset);
2805         bio_add_page(bio, page, page_size, pg_offset);
2806         bio->bi_end_io = end_io_func;
2807         bio->bi_private = tree;
2808         bio->bi_write_hint = page->mapping->host->i_write_hint;
2809         bio->bi_opf = opf;
2810         if (wbc) {
2811                 wbc_init_bio(wbc, bio);
2812                 wbc_account_io(wbc, page, page_size);
2813         }
2814
2815         if (bio_ret)
2816                 *bio_ret = bio;
2817         else
2818                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2819
2820         return ret;
2821 }
2822
2823 static void attach_extent_buffer_page(struct extent_buffer *eb,
2824                                       struct page *page)
2825 {
2826         if (!PagePrivate(page)) {
2827                 SetPagePrivate(page);
2828                 get_page(page);
2829                 set_page_private(page, (unsigned long)eb);
2830         } else {
2831                 WARN_ON(page->private != (unsigned long)eb);
2832         }
2833 }
2834
2835 void set_page_extent_mapped(struct page *page)
2836 {
2837         if (!PagePrivate(page)) {
2838                 SetPagePrivate(page);
2839                 get_page(page);
2840                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2841         }
2842 }
2843
2844 static struct extent_map *
2845 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2846                  u64 start, u64 len, get_extent_t *get_extent,
2847                  struct extent_map **em_cached)
2848 {
2849         struct extent_map *em;
2850
2851         if (em_cached && *em_cached) {
2852                 em = *em_cached;
2853                 if (extent_map_in_tree(em) && start >= em->start &&
2854                     start < extent_map_end(em)) {
2855                         refcount_inc(&em->refs);
2856                         return em;
2857                 }
2858
2859                 free_extent_map(em);
2860                 *em_cached = NULL;
2861         }
2862
2863         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2864         if (em_cached && !IS_ERR_OR_NULL(em)) {
2865                 BUG_ON(*em_cached);
2866                 refcount_inc(&em->refs);
2867                 *em_cached = em;
2868         }
2869         return em;
2870 }
2871 /*
2872  * basic readpage implementation.  Locked extent state structs are inserted
2873  * into the tree that are removed when the IO is done (by the end_io
2874  * handlers)
2875  * XXX JDM: This needs looking at to ensure proper page locking
2876  * return 0 on success, otherwise return error
2877  */
2878 static int __do_readpage(struct extent_io_tree *tree,
2879                          struct page *page,
2880                          get_extent_t *get_extent,
2881                          struct extent_map **em_cached,
2882                          struct bio **bio, int mirror_num,
2883                          unsigned long *bio_flags, unsigned int read_flags,
2884                          u64 *prev_em_start)
2885 {
2886         struct inode *inode = page->mapping->host;
2887         u64 start = page_offset(page);
2888         u64 page_end = start + PAGE_SIZE - 1;
2889         u64 end;
2890         u64 cur = start;
2891         u64 extent_offset;
2892         u64 last_byte = i_size_read(inode);
2893         u64 block_start;
2894         u64 cur_end;
2895         struct extent_map *em;
2896         struct block_device *bdev;
2897         int ret = 0;
2898         int nr = 0;
2899         size_t pg_offset = 0;
2900         size_t iosize;
2901         size_t disk_io_size;
2902         size_t blocksize = inode->i_sb->s_blocksize;
2903         unsigned long this_bio_flag = 0;
2904
2905         set_page_extent_mapped(page);
2906
2907         end = page_end;
2908         if (!PageUptodate(page)) {
2909                 if (cleancache_get_page(page) == 0) {
2910                         BUG_ON(blocksize != PAGE_SIZE);
2911                         unlock_extent(tree, start, end);
2912                         goto out;
2913                 }
2914         }
2915
2916         if (page->index == last_byte >> PAGE_SHIFT) {
2917                 char *userpage;
2918                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2919
2920                 if (zero_offset) {
2921                         iosize = PAGE_SIZE - zero_offset;
2922                         userpage = kmap_atomic(page);
2923                         memset(userpage + zero_offset, 0, iosize);
2924                         flush_dcache_page(page);
2925                         kunmap_atomic(userpage);
2926                 }
2927         }
2928         while (cur <= end) {
2929                 bool force_bio_submit = false;
2930                 u64 offset;
2931
2932                 if (cur >= last_byte) {
2933                         char *userpage;
2934                         struct extent_state *cached = NULL;
2935
2936                         iosize = PAGE_SIZE - pg_offset;
2937                         userpage = kmap_atomic(page);
2938                         memset(userpage + pg_offset, 0, iosize);
2939                         flush_dcache_page(page);
2940                         kunmap_atomic(userpage);
2941                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2942                                             &cached, GFP_NOFS);
2943                         unlock_extent_cached(tree, cur,
2944                                              cur + iosize - 1,
2945                                              &cached, GFP_NOFS);
2946                         break;
2947                 }
2948                 em = __get_extent_map(inode, page, pg_offset, cur,
2949                                       end - cur + 1, get_extent, em_cached);
2950                 if (IS_ERR_OR_NULL(em)) {
2951                         SetPageError(page);
2952                         unlock_extent(tree, cur, end);
2953                         break;
2954                 }
2955                 extent_offset = cur - em->start;
2956                 BUG_ON(extent_map_end(em) <= cur);
2957                 BUG_ON(end < cur);
2958
2959                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2960                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2961                         extent_set_compress_type(&this_bio_flag,
2962                                                  em->compress_type);
2963                 }
2964
2965                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2966                 cur_end = min(extent_map_end(em) - 1, end);
2967                 iosize = ALIGN(iosize, blocksize);
2968                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2969                         disk_io_size = em->block_len;
2970                         offset = em->block_start;
2971                 } else {
2972                         offset = em->block_start + extent_offset;
2973                         disk_io_size = iosize;
2974                 }
2975                 bdev = em->bdev;
2976                 block_start = em->block_start;
2977                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2978                         block_start = EXTENT_MAP_HOLE;
2979
2980                 /*
2981                  * If we have a file range that points to a compressed extent
2982                  * and it's followed by a consecutive file range that points to
2983                  * to the same compressed extent (possibly with a different
2984                  * offset and/or length, so it either points to the whole extent
2985                  * or only part of it), we must make sure we do not submit a
2986                  * single bio to populate the pages for the 2 ranges because
2987                  * this makes the compressed extent read zero out the pages
2988                  * belonging to the 2nd range. Imagine the following scenario:
2989                  *
2990                  *  File layout
2991                  *  [0 - 8K]                     [8K - 24K]
2992                  *    |                               |
2993                  *    |                               |
2994                  * points to extent X,         points to extent X,
2995                  * offset 4K, length of 8K     offset 0, length 16K
2996                  *
2997                  * [extent X, compressed length = 4K uncompressed length = 16K]
2998                  *
2999                  * If the bio to read the compressed extent covers both ranges,
3000                  * it will decompress extent X into the pages belonging to the
3001                  * first range and then it will stop, zeroing out the remaining
3002                  * pages that belong to the other range that points to extent X.
3003                  * So here we make sure we submit 2 bios, one for the first
3004                  * range and another one for the third range. Both will target
3005                  * the same physical extent from disk, but we can't currently
3006                  * make the compressed bio endio callback populate the pages
3007                  * for both ranges because each compressed bio is tightly
3008                  * coupled with a single extent map, and each range can have
3009                  * an extent map with a different offset value relative to the
3010                  * uncompressed data of our extent and different lengths. This
3011                  * is a corner case so we prioritize correctness over
3012                  * non-optimal behavior (submitting 2 bios for the same extent).
3013                  */
3014                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3015                     prev_em_start && *prev_em_start != (u64)-1 &&
3016                     *prev_em_start != em->orig_start)
3017                         force_bio_submit = true;
3018
3019                 if (prev_em_start)
3020                         *prev_em_start = em->orig_start;
3021
3022                 free_extent_map(em);
3023                 em = NULL;
3024
3025                 /* we've found a hole, just zero and go on */
3026                 if (block_start == EXTENT_MAP_HOLE) {
3027                         char *userpage;
3028                         struct extent_state *cached = NULL;
3029
3030                         userpage = kmap_atomic(page);
3031                         memset(userpage + pg_offset, 0, iosize);
3032                         flush_dcache_page(page);
3033                         kunmap_atomic(userpage);
3034
3035                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3036                                             &cached, GFP_NOFS);
3037                         unlock_extent_cached(tree, cur,
3038                                              cur + iosize - 1,
3039                                              &cached, GFP_NOFS);
3040                         cur = cur + iosize;
3041                         pg_offset += iosize;
3042                         continue;
3043                 }
3044                 /* the get_extent function already copied into the page */
3045                 if (test_range_bit(tree, cur, cur_end,
3046                                    EXTENT_UPTODATE, 1, NULL)) {
3047                         check_page_uptodate(tree, page);
3048                         unlock_extent(tree, cur, cur + iosize - 1);
3049                         cur = cur + iosize;
3050                         pg_offset += iosize;
3051                         continue;
3052                 }
3053                 /* we have an inline extent but it didn't get marked up
3054                  * to date.  Error out
3055                  */
3056                 if (block_start == EXTENT_MAP_INLINE) {
3057                         SetPageError(page);
3058                         unlock_extent(tree, cur, cur + iosize - 1);
3059                         cur = cur + iosize;
3060                         pg_offset += iosize;
3061                         continue;
3062                 }
3063
3064                 ret = submit_extent_page(REQ_OP_READ | read_flags, tree, NULL,
3065                                          page, offset, disk_io_size,
3066                                          pg_offset, bdev, bio,
3067                                          end_bio_extent_readpage, mirror_num,
3068                                          *bio_flags,
3069                                          this_bio_flag,
3070                                          force_bio_submit);
3071                 if (!ret) {
3072                         nr++;
3073                         *bio_flags = this_bio_flag;
3074                 } else {
3075                         SetPageError(page);
3076                         unlock_extent(tree, cur, cur + iosize - 1);
3077                         goto out;
3078                 }
3079                 cur = cur + iosize;
3080                 pg_offset += iosize;
3081         }
3082 out:
3083         if (!nr) {
3084                 if (!PageError(page))
3085                         SetPageUptodate(page);
3086                 unlock_page(page);
3087         }
3088         return ret;
3089 }
3090
3091 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3092                                              struct page *pages[], int nr_pages,
3093                                              u64 start, u64 end,
3094                                              struct extent_map **em_cached,
3095                                              struct bio **bio,
3096                                              unsigned long *bio_flags,
3097                                              u64 *prev_em_start)
3098 {
3099         struct inode *inode;
3100         struct btrfs_ordered_extent *ordered;
3101         int index;
3102
3103         inode = pages[0]->mapping->host;
3104         while (1) {
3105                 lock_extent(tree, start, end);
3106                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3107                                                      end - start + 1);
3108                 if (!ordered)
3109                         break;
3110                 unlock_extent(tree, start, end);
3111                 btrfs_start_ordered_extent(inode, ordered, 1);
3112                 btrfs_put_ordered_extent(ordered);
3113         }
3114
3115         for (index = 0; index < nr_pages; index++) {
3116                 __do_readpage(tree, pages[index], btrfs_get_extent, em_cached,
3117                                 bio, 0, bio_flags, 0, prev_em_start);
3118                 put_page(pages[index]);
3119         }
3120 }
3121
3122 static void __extent_readpages(struct extent_io_tree *tree,
3123                                struct page *pages[],
3124                                int nr_pages,
3125                                struct extent_map **em_cached,
3126                                struct bio **bio, unsigned long *bio_flags,
3127                                u64 *prev_em_start)
3128 {
3129         u64 start = 0;
3130         u64 end = 0;
3131         u64 page_start;
3132         int index;
3133         int first_index = 0;
3134
3135         for (index = 0; index < nr_pages; index++) {
3136                 page_start = page_offset(pages[index]);
3137                 if (!end) {
3138                         start = page_start;
3139                         end = start + PAGE_SIZE - 1;
3140                         first_index = index;
3141                 } else if (end + 1 == page_start) {
3142                         end += PAGE_SIZE;
3143                 } else {
3144                         __do_contiguous_readpages(tree, &pages[first_index],
3145                                                   index - first_index, start,
3146                                                   end, em_cached,
3147                                                   bio, bio_flags,
3148                                                   prev_em_start);
3149                         start = page_start;
3150                         end = start + PAGE_SIZE - 1;
3151                         first_index = index;
3152                 }
3153         }
3154
3155         if (end)
3156                 __do_contiguous_readpages(tree, &pages[first_index],
3157                                           index - first_index, start,
3158                                           end, em_cached, bio,
3159                                           bio_flags, prev_em_start);
3160 }
3161
3162 static int __extent_read_full_page(struct extent_io_tree *tree,
3163                                    struct page *page,
3164                                    get_extent_t *get_extent,
3165                                    struct bio **bio, int mirror_num,
3166                                    unsigned long *bio_flags,
3167                                    unsigned int read_flags)
3168 {
3169         struct inode *inode = page->mapping->host;
3170         struct btrfs_ordered_extent *ordered;
3171         u64 start = page_offset(page);
3172         u64 end = start + PAGE_SIZE - 1;
3173         int ret;
3174
3175         while (1) {
3176                 lock_extent(tree, start, end);
3177                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3178                                                 PAGE_SIZE);
3179                 if (!ordered)
3180                         break;
3181                 unlock_extent(tree, start, end);
3182                 btrfs_start_ordered_extent(inode, ordered, 1);
3183                 btrfs_put_ordered_extent(ordered);
3184         }
3185
3186         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3187                             bio_flags, read_flags, NULL);
3188         return ret;
3189 }
3190
3191 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3192                             get_extent_t *get_extent, int mirror_num)
3193 {
3194         struct bio *bio = NULL;
3195         unsigned long bio_flags = 0;
3196         int ret;
3197
3198         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3199                                       &bio_flags, 0);
3200         if (bio)
3201                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3202         return ret;
3203 }
3204
3205 static void update_nr_written(struct writeback_control *wbc,
3206                               unsigned long nr_written)
3207 {
3208         wbc->nr_to_write -= nr_written;
3209 }
3210
3211 /*
3212  * helper for __extent_writepage, doing all of the delayed allocation setup.
3213  *
3214  * This returns 1 if our fill_delalloc function did all the work required
3215  * to write the page (copy into inline extent).  In this case the IO has
3216  * been started and the page is already unlocked.
3217  *
3218  * This returns 0 if all went well (page still locked)
3219  * This returns < 0 if there were errors (page still locked)
3220  */
3221 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3222                               struct page *page, struct writeback_control *wbc,
3223                               struct extent_page_data *epd,
3224                               u64 delalloc_start,
3225                               unsigned long *nr_written)
3226 {
3227         struct extent_io_tree *tree = epd->tree;
3228         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3229         u64 nr_delalloc;
3230         u64 delalloc_to_write = 0;
3231         u64 delalloc_end = 0;
3232         int ret;
3233         int page_started = 0;
3234
3235         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3236                 return 0;
3237
3238         while (delalloc_end < page_end) {
3239                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3240                                                page,
3241                                                &delalloc_start,
3242                                                &delalloc_end,
3243                                                BTRFS_MAX_EXTENT_SIZE);
3244                 if (nr_delalloc == 0) {
3245                         delalloc_start = delalloc_end + 1;
3246                         continue;
3247                 }
3248                 ret = tree->ops->fill_delalloc(inode, page,
3249                                                delalloc_start,
3250                                                delalloc_end,
3251                                                &page_started,
3252                                                nr_written, wbc);
3253                 /* File system has been set read-only */
3254                 if (ret) {
3255                         SetPageError(page);
3256                         /* fill_delalloc should be return < 0 for error
3257                          * but just in case, we use > 0 here meaning the
3258                          * IO is started, so we don't want to return > 0
3259                          * unless things are going well.
3260                          */
3261                         ret = ret < 0 ? ret : -EIO;
3262                         goto done;
3263                 }
3264                 /*
3265                  * delalloc_end is already one less than the total length, so
3266                  * we don't subtract one from PAGE_SIZE
3267                  */
3268                 delalloc_to_write += (delalloc_end - delalloc_start +
3269                                       PAGE_SIZE) >> PAGE_SHIFT;
3270                 delalloc_start = delalloc_end + 1;
3271         }
3272         if (wbc->nr_to_write < delalloc_to_write) {
3273                 int thresh = 8192;
3274
3275                 if (delalloc_to_write < thresh * 2)
3276                         thresh = delalloc_to_write;
3277                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3278                                          thresh);
3279         }
3280
3281         /* did the fill delalloc function already unlock and start
3282          * the IO?
3283          */
3284         if (page_started) {
3285                 /*
3286                  * we've unlocked the page, so we can't update
3287                  * the mapping's writeback index, just update
3288                  * nr_to_write.
3289                  */
3290                 wbc->nr_to_write -= *nr_written;
3291                 return 1;
3292         }
3293
3294         ret = 0;
3295
3296 done:
3297         return ret;
3298 }
3299
3300 /*
3301  * helper for __extent_writepage.  This calls the writepage start hooks,
3302  * and does the loop to map the page into extents and bios.
3303  *
3304  * We return 1 if the IO is started and the page is unlocked,
3305  * 0 if all went well (page still locked)
3306  * < 0 if there were errors (page still locked)
3307  */
3308 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3309                                  struct page *page,
3310                                  struct writeback_control *wbc,
3311                                  struct extent_page_data *epd,
3312                                  loff_t i_size,
3313                                  unsigned long nr_written,
3314                                  unsigned int write_flags, int *nr_ret)
3315 {
3316         struct extent_io_tree *tree = epd->tree;
3317         u64 start = page_offset(page);
3318         u64 page_end = start + PAGE_SIZE - 1;
3319         u64 end;
3320         u64 cur = start;
3321         u64 extent_offset;
3322         u64 block_start;
3323         u64 iosize;
3324         struct extent_map *em;
3325         struct block_device *bdev;
3326         size_t pg_offset = 0;
3327         size_t blocksize;
3328         int ret = 0;
3329         int nr = 0;
3330         bool compressed;
3331
3332         if (tree->ops && tree->ops->writepage_start_hook) {
3333                 ret = tree->ops->writepage_start_hook(page, start,
3334                                                       page_end);
3335                 if (ret) {
3336                         /* Fixup worker will requeue */
3337                         if (ret == -EBUSY)
3338                                 wbc->pages_skipped++;
3339                         else
3340                                 redirty_page_for_writepage(wbc, page);
3341
3342                         update_nr_written(wbc, nr_written);
3343                         unlock_page(page);
3344                         return 1;
3345                 }
3346         }
3347
3348         /*
3349          * we don't want to touch the inode after unlocking the page,
3350          * so we update the mapping writeback index now
3351          */
3352         update_nr_written(wbc, nr_written + 1);
3353
3354         end = page_end;
3355         if (i_size <= start) {
3356                 if (tree->ops && tree->ops->writepage_end_io_hook)
3357                         tree->ops->writepage_end_io_hook(page, start,
3358                                                          page_end, NULL, 1);
3359                 goto done;
3360         }
3361
3362         blocksize = inode->i_sb->s_blocksize;
3363
3364         while (cur <= end) {
3365                 u64 em_end;
3366                 u64 offset;
3367
3368                 if (cur >= i_size) {
3369                         if (tree->ops && tree->ops->writepage_end_io_hook)
3370                                 tree->ops->writepage_end_io_hook(page, cur,
3371                                                          page_end, NULL, 1);
3372                         break;
3373                 }
3374                 em = btrfs_get_extent(BTRFS_I(inode), page, pg_offset, cur,
3375                                      end - cur + 1, 1);
3376                 if (IS_ERR_OR_NULL(em)) {
3377                         SetPageError(page);
3378                         ret = PTR_ERR_OR_ZERO(em);
3379                         break;
3380                 }
3381
3382                 extent_offset = cur - em->start;
3383                 em_end = extent_map_end(em);
3384                 BUG_ON(em_end <= cur);
3385                 BUG_ON(end < cur);
3386                 iosize = min(em_end - cur, end - cur + 1);
3387                 iosize = ALIGN(iosize, blocksize);
3388                 offset = em->block_start + extent_offset;
3389                 bdev = em->bdev;
3390                 block_start = em->block_start;
3391                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3392                 free_extent_map(em);
3393                 em = NULL;
3394
3395                 /*
3396                  * compressed and inline extents are written through other
3397                  * paths in the FS
3398                  */
3399                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3400                     block_start == EXTENT_MAP_INLINE) {
3401                         /*
3402                          * end_io notification does not happen here for
3403                          * compressed extents
3404                          */
3405                         if (!compressed && tree->ops &&
3406                             tree->ops->writepage_end_io_hook)
3407                                 tree->ops->writepage_end_io_hook(page, cur,
3408                                                          cur + iosize - 1,
3409                                                          NULL, 1);
3410                         else if (compressed) {
3411                                 /* we don't want to end_page_writeback on
3412                                  * a compressed extent.  this happens
3413                                  * elsewhere
3414                                  */
3415                                 nr++;
3416                         }
3417
3418                         cur += iosize;
3419                         pg_offset += iosize;
3420                         continue;
3421                 }
3422
3423                 set_range_writeback(tree, cur, cur + iosize - 1);
3424                 if (!PageWriteback(page)) {
3425                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3426                                    "page %lu not writeback, cur %llu end %llu",
3427                                page->index, cur, end);
3428                 }
3429
3430                 ret = submit_extent_page(REQ_OP_WRITE | write_flags, tree, wbc,
3431                                          page, offset, iosize, pg_offset,
3432                                          bdev, &epd->bio,
3433                                          end_bio_extent_writepage,
3434                                          0, 0, 0, false);
3435                 if (ret) {
3436                         SetPageError(page);
3437                         if (PageWriteback(page))
3438                                 end_page_writeback(page);
3439                 }
3440
3441                 cur = cur + iosize;
3442                 pg_offset += iosize;
3443                 nr++;
3444         }
3445 done:
3446         *nr_ret = nr;
3447         return ret;
3448 }
3449
3450 /*
3451  * the writepage semantics are similar to regular writepage.  extent
3452  * records are inserted to lock ranges in the tree, and as dirty areas
3453  * are found, they are marked writeback.  Then the lock bits are removed
3454  * and the end_io handler clears the writeback ranges
3455  */
3456 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3457                               struct extent_page_data *epd)
3458 {
3459         struct inode *inode = page->mapping->host;
3460         u64 start = page_offset(page);
3461         u64 page_end = start + PAGE_SIZE - 1;
3462         int ret;
3463         int nr = 0;
3464         size_t pg_offset = 0;
3465         loff_t i_size = i_size_read(inode);
3466         unsigned long end_index = i_size >> PAGE_SHIFT;
3467         unsigned int write_flags = 0;
3468         unsigned long nr_written = 0;
3469
3470         write_flags = wbc_to_write_flags(wbc);
3471
3472         trace___extent_writepage(page, inode, wbc);
3473
3474         WARN_ON(!PageLocked(page));
3475
3476         ClearPageError(page);
3477
3478         pg_offset = i_size & (PAGE_SIZE - 1);
3479         if (page->index > end_index ||
3480            (page->index == end_index && !pg_offset)) {
3481                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3482                 unlock_page(page);
3483                 return 0;
3484         }
3485
3486         if (page->index == end_index) {
3487                 char *userpage;
3488
3489                 userpage = kmap_atomic(page);
3490                 memset(userpage + pg_offset, 0,
3491                        PAGE_SIZE - pg_offset);
3492                 kunmap_atomic(userpage);
3493                 flush_dcache_page(page);
3494         }
3495
3496         pg_offset = 0;
3497
3498         set_page_extent_mapped(page);
3499
3500         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3501         if (ret == 1)
3502                 goto done_unlocked;
3503         if (ret)
3504                 goto done;
3505
3506         ret = __extent_writepage_io(inode, page, wbc, epd,
3507                                     i_size, nr_written, write_flags, &nr);
3508         if (ret == 1)
3509                 goto done_unlocked;
3510
3511 done:
3512         if (nr == 0) {
3513                 /* make sure the mapping tag for page dirty gets cleared */
3514                 set_page_writeback(page);
3515                 end_page_writeback(page);
3516         }
3517         if (PageError(page)) {
3518                 ret = ret < 0 ? ret : -EIO;
3519                 end_extent_writepage(page, ret, start, page_end);
3520         }
3521         unlock_page(page);
3522         return ret;
3523
3524 done_unlocked:
3525         return 0;
3526 }
3527
3528 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3529 {
3530         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3531                        TASK_UNINTERRUPTIBLE);
3532 }
3533
3534 static noinline_for_stack int
3535 lock_extent_buffer_for_io(struct extent_buffer *eb,
3536                           struct btrfs_fs_info *fs_info,
3537                           struct extent_page_data *epd)
3538 {
3539         unsigned long i, num_pages;
3540         int flush = 0;
3541         int ret = 0;
3542
3543         if (!btrfs_try_tree_write_lock(eb)) {
3544                 flush = 1;
3545                 flush_write_bio(epd);
3546                 btrfs_tree_lock(eb);
3547         }
3548
3549         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3550                 btrfs_tree_unlock(eb);
3551                 if (!epd->sync_io)
3552                         return 0;
3553                 if (!flush) {
3554                         flush_write_bio(epd);
3555                         flush = 1;
3556                 }
3557                 while (1) {
3558                         wait_on_extent_buffer_writeback(eb);
3559                         btrfs_tree_lock(eb);
3560                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3561                                 break;
3562                         btrfs_tree_unlock(eb);
3563                 }
3564         }
3565
3566         /*
3567          * We need to do this to prevent races in people who check if the eb is
3568          * under IO since we can end up having no IO bits set for a short period
3569          * of time.
3570          */
3571         spin_lock(&eb->refs_lock);
3572         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3573                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3574                 spin_unlock(&eb->refs_lock);
3575                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3576                 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
3577                                          -eb->len,
3578                                          fs_info->dirty_metadata_batch);
3579                 ret = 1;
3580         } else {
3581                 spin_unlock(&eb->refs_lock);
3582         }
3583
3584         btrfs_tree_unlock(eb);
3585
3586         if (!ret)
3587                 return ret;
3588
3589         num_pages = num_extent_pages(eb->start, eb->len);
3590         for (i = 0; i < num_pages; i++) {
3591                 struct page *p = eb->pages[i];
3592
3593                 if (!trylock_page(p)) {
3594                         if (!flush) {
3595                                 flush_write_bio(epd);
3596                                 flush = 1;
3597                         }
3598                         lock_page(p);
3599                 }
3600         }
3601
3602         return ret;
3603 }
3604
3605 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3606 {
3607         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3608         smp_mb__after_atomic();
3609         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3610 }
3611
3612 static void set_btree_ioerr(struct page *page)
3613 {
3614         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3615
3616         SetPageError(page);
3617         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3618                 return;
3619
3620         /*
3621          * If writeback for a btree extent that doesn't belong to a log tree
3622          * failed, increment the counter transaction->eb_write_errors.
3623          * We do this because while the transaction is running and before it's
3624          * committing (when we call filemap_fdata[write|wait]_range against
3625          * the btree inode), we might have
3626          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3627          * returns an error or an error happens during writeback, when we're
3628          * committing the transaction we wouldn't know about it, since the pages
3629          * can be no longer dirty nor marked anymore for writeback (if a
3630          * subsequent modification to the extent buffer didn't happen before the
3631          * transaction commit), which makes filemap_fdata[write|wait]_range not
3632          * able to find the pages tagged with SetPageError at transaction
3633          * commit time. So if this happens we must abort the transaction,
3634          * otherwise we commit a super block with btree roots that point to
3635          * btree nodes/leafs whose content on disk is invalid - either garbage
3636          * or the content of some node/leaf from a past generation that got
3637          * cowed or deleted and is no longer valid.
3638          *
3639          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3640          * not be enough - we need to distinguish between log tree extents vs
3641          * non-log tree extents, and the next filemap_fdatawait_range() call
3642          * will catch and clear such errors in the mapping - and that call might
3643          * be from a log sync and not from a transaction commit. Also, checking
3644          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3645          * not done and would not be reliable - the eb might have been released
3646          * from memory and reading it back again means that flag would not be
3647          * set (since it's a runtime flag, not persisted on disk).
3648          *
3649          * Using the flags below in the btree inode also makes us achieve the
3650          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3651          * writeback for all dirty pages and before filemap_fdatawait_range()
3652          * is called, the writeback for all dirty pages had already finished
3653          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3654          * filemap_fdatawait_range() would return success, as it could not know
3655          * that writeback errors happened (the pages were no longer tagged for
3656          * writeback).
3657          */
3658         switch (eb->log_index) {
3659         case -1:
3660                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3661                 break;
3662         case 0:
3663                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3664                 break;
3665         case 1:
3666                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3667                 break;
3668         default:
3669                 BUG(); /* unexpected, logic error */
3670         }
3671 }
3672
3673 static void end_bio_extent_buffer_writepage(struct bio *bio)
3674 {
3675         struct bio_vec *bvec;
3676         struct extent_buffer *eb;
3677         int i, done;
3678
3679         ASSERT(!bio_flagged(bio, BIO_CLONED));
3680         bio_for_each_segment_all(bvec, bio, i) {
3681                 struct page *page = bvec->bv_page;
3682
3683                 eb = (struct extent_buffer *)page->private;
3684                 BUG_ON(!eb);
3685                 done = atomic_dec_and_test(&eb->io_pages);
3686
3687                 if (bio->bi_status ||
3688                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3689                         ClearPageUptodate(page);
3690                         set_btree_ioerr(page);
3691                 }
3692
3693                 end_page_writeback(page);
3694
3695                 if (!done)
3696                         continue;
3697
3698                 end_extent_buffer_writeback(eb);
3699         }
3700
3701         bio_put(bio);
3702 }
3703
3704 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3705                         struct btrfs_fs_info *fs_info,
3706                         struct writeback_control *wbc,
3707                         struct extent_page_data *epd)
3708 {
3709         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3710         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3711         u64 offset = eb->start;
3712         u32 nritems;
3713         unsigned long i, num_pages;
3714         unsigned long start, end;
3715         unsigned int write_flags = wbc_to_write_flags(wbc) | REQ_META;
3716         int ret = 0;
3717
3718         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3719         num_pages = num_extent_pages(eb->start, eb->len);
3720         atomic_set(&eb->io_pages, num_pages);
3721
3722         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3723         nritems = btrfs_header_nritems(eb);
3724         if (btrfs_header_level(eb) > 0) {
3725                 end = btrfs_node_key_ptr_offset(nritems);
3726
3727                 memzero_extent_buffer(eb, end, eb->len - end);
3728         } else {
3729                 /*
3730                  * leaf:
3731                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3732                  */
3733                 start = btrfs_item_nr_offset(nritems);
3734                 end = BTRFS_LEAF_DATA_OFFSET + leaf_data_end(fs_info, eb);
3735                 memzero_extent_buffer(eb, start, end - start);
3736         }
3737
3738         for (i = 0; i < num_pages; i++) {
3739                 struct page *p = eb->pages[i];
3740
3741                 clear_page_dirty_for_io(p);
3742                 set_page_writeback(p);
3743                 ret = submit_extent_page(REQ_OP_WRITE | write_flags, tree, wbc,
3744                                          p, offset, PAGE_SIZE, 0, bdev,
3745                                          &epd->bio,
3746                                          end_bio_extent_buffer_writepage,
3747                                          0, 0, 0, false);
3748                 if (ret) {
3749                         set_btree_ioerr(p);
3750                         if (PageWriteback(p))
3751                                 end_page_writeback(p);
3752                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3753                                 end_extent_buffer_writeback(eb);
3754                         ret = -EIO;
3755                         break;
3756                 }
3757                 offset += PAGE_SIZE;
3758                 update_nr_written(wbc, 1);
3759                 unlock_page(p);
3760         }
3761
3762         if (unlikely(ret)) {
3763                 for (; i < num_pages; i++) {
3764                         struct page *p = eb->pages[i];
3765                         clear_page_dirty_for_io(p);
3766                         unlock_page(p);
3767                 }
3768         }
3769
3770         return ret;
3771 }
3772
3773 int btree_write_cache_pages(struct address_space *mapping,
3774                                    struct writeback_control *wbc)
3775 {
3776         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3777         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3778         struct extent_buffer *eb, *prev_eb = NULL;
3779         struct extent_page_data epd = {
3780                 .bio = NULL,
3781                 .tree = tree,
3782                 .extent_locked = 0,
3783                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3784         };
3785         int ret = 0;
3786         int done = 0;
3787         int nr_to_write_done = 0;
3788         struct pagevec pvec;
3789         int nr_pages;
3790         pgoff_t index;
3791         pgoff_t end;            /* Inclusive */
3792         int scanned = 0;
3793         int tag;
3794
3795         pagevec_init(&pvec);
3796         if (wbc->range_cyclic) {
3797                 index = mapping->writeback_index; /* Start from prev offset */
3798                 end = -1;
3799         } else {
3800                 index = wbc->range_start >> PAGE_SHIFT;
3801                 end = wbc->range_end >> PAGE_SHIFT;
3802                 scanned = 1;
3803         }
3804         if (wbc->sync_mode == WB_SYNC_ALL)
3805                 tag = PAGECACHE_TAG_TOWRITE;
3806         else
3807                 tag = PAGECACHE_TAG_DIRTY;
3808 retry:
3809         if (wbc->sync_mode == WB_SYNC_ALL)
3810                 tag_pages_for_writeback(mapping, index, end);
3811         while (!done && !nr_to_write_done && (index <= end) &&
3812                (nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
3813                         tag))) {
3814                 unsigned i;
3815
3816                 scanned = 1;
3817                 for (i = 0; i < nr_pages; i++) {
3818                         struct page *page = pvec.pages[i];
3819
3820                         if (!PagePrivate(page))
3821                                 continue;
3822
3823                         spin_lock(&mapping->private_lock);
3824                         if (!PagePrivate(page)) {
3825                                 spin_unlock(&mapping->private_lock);
3826                                 continue;
3827                         }
3828
3829                         eb = (struct extent_buffer *)page->private;
3830
3831                         /*
3832                          * Shouldn't happen and normally this would be a BUG_ON
3833                          * but no sense in crashing the users box for something
3834                          * we can survive anyway.
3835                          */
3836                         if (WARN_ON(!eb)) {
3837                                 spin_unlock(&mapping->private_lock);
3838                                 continue;
3839                         }
3840
3841                         if (eb == prev_eb) {
3842                                 spin_unlock(&mapping->private_lock);
3843                                 continue;
3844                         }
3845
3846                         ret = atomic_inc_not_zero(&eb->refs);
3847                         spin_unlock(&mapping->private_lock);
3848                         if (!ret)
3849                                 continue;
3850
3851                         prev_eb = eb;
3852                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3853                         if (!ret) {
3854                                 free_extent_buffer(eb);
3855                                 continue;
3856                         }
3857
3858                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3859                         if (ret) {
3860                                 done = 1;
3861                                 free_extent_buffer(eb);
3862                                 break;
3863                         }
3864                         free_extent_buffer(eb);
3865
3866                         /*
3867                          * the filesystem may choose to bump up nr_to_write.
3868                          * We have to make sure to honor the new nr_to_write
3869                          * at any time
3870                          */
3871                         nr_to_write_done = wbc->nr_to_write <= 0;
3872                 }
3873                 pagevec_release(&pvec);
3874                 cond_resched();
3875         }
3876         if (!scanned && !done) {
3877                 /*
3878                  * We hit the last page and there is more work to be done: wrap
3879                  * back to the start of the file
3880                  */
3881                 scanned = 1;
3882                 index = 0;
3883                 goto retry;
3884         }
3885         flush_write_bio(&epd);
3886         return ret;
3887 }
3888
3889 /**
3890  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3891  * @mapping: address space structure to write
3892  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3893  * @data: data passed to __extent_writepage function
3894  *
3895  * If a page is already under I/O, write_cache_pages() skips it, even
3896  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3897  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3898  * and msync() need to guarantee that all the data which was dirty at the time
3899  * the call was made get new I/O started against them.  If wbc->sync_mode is
3900  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3901  * existing IO to complete.
3902  */
3903 static int extent_write_cache_pages(struct address_space *mapping,
3904                              struct writeback_control *wbc,
3905                              struct extent_page_data *epd)
3906 {
3907         struct inode *inode = mapping->host;
3908         int ret = 0;
3909         int done = 0;
3910         int nr_to_write_done = 0;
3911         struct pagevec pvec;
3912         int nr_pages;
3913         pgoff_t index;
3914         pgoff_t end;            /* Inclusive */
3915         pgoff_t done_index;
3916         int range_whole = 0;
3917         int scanned = 0;
3918         int tag;
3919
3920         /*
3921          * We have to hold onto the inode so that ordered extents can do their
3922          * work when the IO finishes.  The alternative to this is failing to add
3923          * an ordered extent if the igrab() fails there and that is a huge pain
3924          * to deal with, so instead just hold onto the inode throughout the
3925          * writepages operation.  If it fails here we are freeing up the inode
3926          * anyway and we'd rather not waste our time writing out stuff that is
3927          * going to be truncated anyway.
3928          */
3929         if (!igrab(inode))
3930                 return 0;
3931
3932         pagevec_init(&pvec);
3933         if (wbc->range_cyclic) {
3934                 index = mapping->writeback_index; /* Start from prev offset */
3935                 end = -1;
3936         } else {
3937                 index = wbc->range_start >> PAGE_SHIFT;
3938                 end = wbc->range_end >> PAGE_SHIFT;
3939                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3940                         range_whole = 1;
3941                 scanned = 1;
3942         }
3943         if (wbc->sync_mode == WB_SYNC_ALL)
3944                 tag = PAGECACHE_TAG_TOWRITE;
3945         else
3946                 tag = PAGECACHE_TAG_DIRTY;
3947 retry:
3948         if (wbc->sync_mode == WB_SYNC_ALL)
3949                 tag_pages_for_writeback(mapping, index, end);
3950         done_index = index;
3951         while (!done && !nr_to_write_done && (index <= end) &&
3952                         (nr_pages = pagevec_lookup_range_tag(&pvec, mapping,
3953                                                 &index, end, tag))) {
3954                 unsigned i;
3955
3956                 scanned = 1;
3957                 for (i = 0; i < nr_pages; i++) {
3958                         struct page *page = pvec.pages[i];
3959
3960                         done_index = page->index;
3961                         /*
3962                          * At this point we hold neither mapping->tree_lock nor
3963                          * lock on the page itself: the page may be truncated or
3964                          * invalidated (changing page->mapping to NULL), or even
3965                          * swizzled back from swapper_space to tmpfs file
3966                          * mapping
3967                          */
3968                         if (!trylock_page(page)) {
3969                                 flush_write_bio(epd);
3970                                 lock_page(page);
3971                         }
3972
3973                         if (unlikely(page->mapping != mapping)) {
3974                                 unlock_page(page);
3975                                 continue;
3976                         }
3977
3978                         if (wbc->sync_mode != WB_SYNC_NONE) {
3979                                 if (PageWriteback(page))
3980                                         flush_write_bio(epd);
3981                                 wait_on_page_writeback(page);
3982                         }
3983
3984                         if (PageWriteback(page) ||
3985                             !clear_page_dirty_for_io(page)) {
3986                                 unlock_page(page);
3987                                 continue;
3988                         }
3989
3990                         ret = __extent_writepage(page, wbc, epd);
3991
3992                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3993                                 unlock_page(page);
3994                                 ret = 0;
3995                         }
3996                         if (ret < 0) {
3997                                 /*
3998                                  * done_index is set past this page,
3999                                  * so media errors will not choke
4000                                  * background writeout for the entire
4001                                  * file. This has consequences for
4002                                  * range_cyclic semantics (ie. it may
4003                                  * not be suitable for data integrity
4004                                  * writeout).
4005                                  */
4006                                 done_index = page->index + 1;
4007                                 done = 1;
4008                                 break;
4009                         }
4010
4011                         /*
4012                          * the filesystem may choose to bump up nr_to_write.
4013                          * We have to make sure to honor the new nr_to_write
4014                          * at any time
4015                          */
4016                         nr_to_write_done = wbc->nr_to_write <= 0;
4017                 }
4018                 pagevec_release(&pvec);
4019                 cond_resched();
4020         }
4021         if (!scanned && !done) {
4022                 /*
4023                  * We hit the last page and there is more work to be done: wrap
4024                  * back to the start of the file
4025                  */
4026                 scanned = 1;
4027                 index = 0;
4028                 goto retry;
4029         }
4030
4031         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4032                 mapping->writeback_index = done_index;
4033
4034         btrfs_add_delayed_iput(inode);
4035         return ret;
4036 }
4037
4038 static void flush_write_bio(struct extent_page_data *epd)
4039 {
4040         if (epd->bio) {
4041                 int ret;
4042
4043                 ret = submit_one_bio(epd->bio, 0, 0);
4044                 BUG_ON(ret < 0); /* -ENOMEM */
4045                 epd->bio = NULL;
4046         }
4047 }
4048
4049 int extent_write_full_page(struct page *page, struct writeback_control *wbc)
4050 {
4051         int ret;
4052         struct extent_page_data epd = {
4053                 .bio = NULL,
4054                 .tree = &BTRFS_I(page->mapping->host)->io_tree,
4055                 .extent_locked = 0,
4056                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4057         };
4058
4059         ret = __extent_writepage(page, wbc, &epd);
4060
4061         flush_write_bio(&epd);
4062         return ret;
4063 }
4064
4065 int extent_write_locked_range(struct inode *inode, u64 start, u64 end,
4066                               int mode)
4067 {
4068         int ret = 0;
4069         struct address_space *mapping = inode->i_mapping;
4070         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
4071         struct page *page;
4072         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4073                 PAGE_SHIFT;
4074
4075         struct extent_page_data epd = {
4076                 .bio = NULL,
4077                 .tree = tree,
4078                 .extent_locked = 1,
4079                 .sync_io = mode == WB_SYNC_ALL,
4080         };
4081         struct writeback_control wbc_writepages = {
4082                 .sync_mode      = mode,
4083                 .nr_to_write    = nr_pages * 2,
4084                 .range_start    = start,
4085                 .range_end      = end + 1,
4086         };
4087
4088         while (start <= end) {
4089                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4090                 if (clear_page_dirty_for_io(page))
4091                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4092                 else {
4093                         if (tree->ops && tree->ops->writepage_end_io_hook)
4094                                 tree->ops->writepage_end_io_hook(page, start,
4095                                                  start + PAGE_SIZE - 1,
4096                                                  NULL, 1);
4097                         unlock_page(page);
4098                 }
4099                 put_page(page);
4100                 start += PAGE_SIZE;
4101         }
4102
4103         flush_write_bio(&epd);
4104         return ret;
4105 }
4106
4107 int extent_writepages(struct extent_io_tree *tree,
4108                       struct address_space *mapping,
4109                       struct writeback_control *wbc)
4110 {
4111         int ret = 0;
4112         struct extent_page_data epd = {
4113                 .bio = NULL,
4114                 .tree = tree,
4115                 .extent_locked = 0,
4116                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4117         };
4118
4119         ret = extent_write_cache_pages(mapping, wbc, &epd);
4120         flush_write_bio(&epd);
4121         return ret;
4122 }
4123
4124 int extent_readpages(struct extent_io_tree *tree,
4125                      struct address_space *mapping,
4126                      struct list_head *pages, unsigned nr_pages)
4127 {
4128         struct bio *bio = NULL;
4129         unsigned page_idx;
4130         unsigned long bio_flags = 0;
4131         struct page *pagepool[16];
4132         struct page *page;
4133         struct extent_map *em_cached = NULL;
4134         int nr = 0;
4135         u64 prev_em_start = (u64)-1;
4136
4137         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4138                 page = list_entry(pages->prev, struct page, lru);
4139
4140                 prefetchw(&page->flags);
4141                 list_del(&page->lru);
4142                 if (add_to_page_cache_lru(page, mapping,
4143                                         page->index,
4144                                         readahead_gfp_mask(mapping))) {
4145                         put_page(page);
4146                         continue;
4147                 }
4148
4149                 pagepool[nr++] = page;
4150                 if (nr < ARRAY_SIZE(pagepool))
4151                         continue;
4152                 __extent_readpages(tree, pagepool, nr, &em_cached, &bio,
4153                                 &bio_flags, &prev_em_start);
4154                 nr = 0;
4155         }
4156         if (nr)
4157                 __extent_readpages(tree, pagepool, nr, &em_cached, &bio,
4158                                 &bio_flags, &prev_em_start);
4159
4160         if (em_cached)
4161                 free_extent_map(em_cached);
4162
4163         BUG_ON(!list_empty(pages));
4164         if (bio)
4165                 return submit_one_bio(bio, 0, bio_flags);
4166         return 0;
4167 }
4168
4169 /*
4170  * basic invalidatepage code, this waits on any locked or writeback
4171  * ranges corresponding to the page, and then deletes any extent state
4172  * records from the tree
4173  */
4174 int extent_invalidatepage(struct extent_io_tree *tree,
4175                           struct page *page, unsigned long offset)
4176 {
4177         struct extent_state *cached_state = NULL;
4178         u64 start = page_offset(page);
4179         u64 end = start + PAGE_SIZE - 1;
4180         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4181
4182         start += ALIGN(offset, blocksize);
4183         if (start > end)
4184                 return 0;
4185
4186         lock_extent_bits(tree, start, end, &cached_state);
4187         wait_on_page_writeback(page);
4188         clear_extent_bit(tree, start, end,
4189                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4190                          EXTENT_DO_ACCOUNTING,
4191                          1, 1, &cached_state);
4192         return 0;
4193 }
4194
4195 /*
4196  * a helper for releasepage, this tests for areas of the page that
4197  * are locked or under IO and drops the related state bits if it is safe
4198  * to drop the page.
4199  */
4200 static int try_release_extent_state(struct extent_map_tree *map,
4201                                     struct extent_io_tree *tree,
4202                                     struct page *page, gfp_t mask)
4203 {
4204         u64 start = page_offset(page);
4205         u64 end = start + PAGE_SIZE - 1;
4206         int ret = 1;
4207
4208         if (test_range_bit(tree, start, end,
4209                            EXTENT_IOBITS, 0, NULL))
4210                 ret = 0;
4211         else {
4212                 /*
4213                  * at this point we can safely clear everything except the
4214                  * locked bit and the nodatasum bit
4215                  */
4216                 ret = __clear_extent_bit(tree, start, end,
4217                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4218                                  0, 0, NULL, mask, NULL);
4219
4220                 /* if clear_extent_bit failed for enomem reasons,
4221                  * we can't allow the release to continue.
4222                  */
4223                 if (ret < 0)
4224                         ret = 0;
4225                 else
4226                         ret = 1;
4227         }
4228         return ret;
4229 }
4230
4231 /*
4232  * a helper for releasepage.  As long as there are no locked extents
4233  * in the range corresponding to the page, both state records and extent
4234  * map records are removed
4235  */
4236 int try_release_extent_mapping(struct extent_map_tree *map,
4237                                struct extent_io_tree *tree, struct page *page,
4238                                gfp_t mask)
4239 {
4240         struct extent_map *em;
4241         u64 start = page_offset(page);
4242         u64 end = start + PAGE_SIZE - 1;
4243
4244         if (gfpflags_allow_blocking(mask) &&
4245             page->mapping->host->i_size > SZ_16M) {
4246                 u64 len;
4247                 while (start <= end) {
4248                         len = end - start + 1;
4249                         write_lock(&map->lock);
4250                         em = lookup_extent_mapping(map, start, len);
4251                         if (!em) {
4252                                 write_unlock(&map->lock);
4253                                 break;
4254                         }
4255                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4256                             em->start != start) {
4257                                 write_unlock(&map->lock);
4258                                 free_extent_map(em);
4259                                 break;
4260                         }
4261                         if (!test_range_bit(tree, em->start,
4262                                             extent_map_end(em) - 1,
4263                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4264                                             0, NULL)) {
4265                                 remove_extent_mapping(map, em);
4266                                 /* once for the rb tree */
4267                                 free_extent_map(em);
4268                         }
4269                         start = extent_map_end(em);
4270                         write_unlock(&map->lock);
4271
4272                         /* once for us */
4273                         free_extent_map(em);
4274                 }
4275         }
4276         return try_release_extent_state(map, tree, page, mask);
4277 }
4278
4279 /*
4280  * helper function for fiemap, which doesn't want to see any holes.
4281  * This maps until we find something past 'last'
4282  */
4283 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4284                                                 u64 offset, u64 last)
4285 {
4286         u64 sectorsize = btrfs_inode_sectorsize(inode);
4287         struct extent_map *em;
4288         u64 len;
4289
4290         if (offset >= last)
4291                 return NULL;
4292
4293         while (1) {
4294                 len = last - offset;
4295                 if (len == 0)
4296                         break;
4297                 len = ALIGN(len, sectorsize);
4298                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0, offset,
4299                                 len, 0);
4300                 if (IS_ERR_OR_NULL(em))
4301                         return em;
4302
4303                 /* if this isn't a hole return it */
4304                 if (em->block_start != EXTENT_MAP_HOLE)
4305                         return em;
4306
4307                 /* this is a hole, advance to the next extent */
4308                 offset = extent_map_end(em);
4309                 free_extent_map(em);
4310                 if (offset >= last)
4311                         break;
4312         }
4313         return NULL;
4314 }
4315
4316 /*
4317  * To cache previous fiemap extent
4318  *
4319  * Will be used for merging fiemap extent
4320  */
4321 struct fiemap_cache {
4322         u64 offset;
4323         u64 phys;
4324         u64 len;
4325         u32 flags;
4326         bool cached;
4327 };
4328
4329 /*
4330  * Helper to submit fiemap extent.
4331  *
4332  * Will try to merge current fiemap extent specified by @offset, @phys,
4333  * @len and @flags with cached one.
4334  * And only when we fails to merge, cached one will be submitted as
4335  * fiemap extent.
4336  *
4337  * Return value is the same as fiemap_fill_next_extent().
4338  */
4339 static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
4340                                 struct fiemap_cache *cache,
4341                                 u64 offset, u64 phys, u64 len, u32 flags)
4342 {
4343         int ret = 0;
4344
4345         if (!cache->cached)
4346                 goto assign;
4347
4348         /*
4349          * Sanity check, extent_fiemap() should have ensured that new
4350          * fiemap extent won't overlap with cahced one.
4351          * Not recoverable.
4352          *
4353          * NOTE: Physical address can overlap, due to compression
4354          */
4355         if (cache->offset + cache->len > offset) {
4356                 WARN_ON(1);
4357                 return -EINVAL;
4358         }
4359
4360         /*
4361          * Only merges fiemap extents if
4362          * 1) Their logical addresses are continuous
4363          *
4364          * 2) Their physical addresses are continuous
4365          *    So truly compressed (physical size smaller than logical size)
4366          *    extents won't get merged with each other
4367          *
4368          * 3) Share same flags except FIEMAP_EXTENT_LAST
4369          *    So regular extent won't get merged with prealloc extent
4370          */
4371         if (cache->offset + cache->len  == offset &&
4372             cache->phys + cache->len == phys  &&
4373             (cache->flags & ~FIEMAP_EXTENT_LAST) ==
4374                         (flags & ~FIEMAP_EXTENT_LAST)) {
4375                 cache->len += len;
4376                 cache->flags |= flags;
4377                 goto try_submit_last;
4378         }
4379
4380         /* Not mergeable, need to submit cached one */
4381         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4382                                       cache->len, cache->flags);
4383         cache->cached = false;
4384         if (ret)
4385                 return ret;
4386 assign:
4387         cache->cached = true;
4388         cache->offset = offset;
4389         cache->phys = phys;
4390         cache->len = len;
4391         cache->flags = flags;
4392 try_submit_last:
4393         if (cache->flags & FIEMAP_EXTENT_LAST) {
4394                 ret = fiemap_fill_next_extent(fieinfo, cache->offset,
4395                                 cache->phys, cache->len, cache->flags);
4396                 cache->cached = false;
4397         }
4398         return ret;
4399 }
4400
4401 /*
4402  * Emit last fiemap cache
4403  *
4404  * The last fiemap cache may still be cached in the following case:
4405  * 0                  4k                    8k
4406  * |<- Fiemap range ->|
4407  * |<------------  First extent ----------->|
4408  *
4409  * In this case, the first extent range will be cached but not emitted.
4410  * So we must emit it before ending extent_fiemap().
4411  */
4412 static int emit_last_fiemap_cache(struct btrfs_fs_info *fs_info,
4413                                   struct fiemap_extent_info *fieinfo,
4414                                   struct fiemap_cache *cache)
4415 {
4416         int ret;
4417
4418         if (!cache->cached)
4419                 return 0;
4420
4421         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4422                                       cache->len, cache->flags);
4423         cache->cached = false;
4424         if (ret > 0)
4425                 ret = 0;
4426         return ret;
4427 }
4428
4429 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4430                 __u64 start, __u64 len)
4431 {
4432         int ret = 0;
4433         u64 off = start;
4434         u64 max = start + len;
4435         u32 flags = 0;
4436         u32 found_type;
4437         u64 last;
4438         u64 last_for_get_extent = 0;
4439         u64 disko = 0;
4440         u64 isize = i_size_read(inode);
4441         struct btrfs_key found_key;
4442         struct extent_map *em = NULL;
4443         struct extent_state *cached_state = NULL;
4444         struct btrfs_path *path;
4445         struct btrfs_root *root = BTRFS_I(inode)->root;
4446         struct fiemap_cache cache = { 0 };
4447         int end = 0;
4448         u64 em_start = 0;
4449         u64 em_len = 0;
4450         u64 em_end = 0;
4451
4452         if (len == 0)
4453                 return -EINVAL;
4454
4455         path = btrfs_alloc_path();
4456         if (!path)
4457                 return -ENOMEM;
4458         path->leave_spinning = 1;
4459
4460         start = round_down(start, btrfs_inode_sectorsize(inode));
4461         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4462
4463         /*
4464          * lookup the last file extent.  We're not using i_size here
4465          * because there might be preallocation past i_size
4466          */
4467         ret = btrfs_lookup_file_extent(NULL, root, path,
4468                         btrfs_ino(BTRFS_I(inode)), -1, 0);
4469         if (ret < 0) {
4470                 btrfs_free_path(path);
4471                 return ret;
4472         } else {
4473                 WARN_ON(!ret);
4474                 if (ret == 1)
4475                         ret = 0;
4476         }
4477
4478         path->slots[0]--;
4479         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4480         found_type = found_key.type;
4481
4482         /* No extents, but there might be delalloc bits */
4483         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4484             found_type != BTRFS_EXTENT_DATA_KEY) {
4485                 /* have to trust i_size as the end */
4486                 last = (u64)-1;
4487                 last_for_get_extent = isize;
4488         } else {
4489                 /*
4490                  * remember the start of the last extent.  There are a
4491                  * bunch of different factors that go into the length of the
4492                  * extent, so its much less complex to remember where it started
4493                  */
4494                 last = found_key.offset;
4495                 last_for_get_extent = last + 1;
4496         }
4497         btrfs_release_path(path);
4498
4499         /*
4500          * we might have some extents allocated but more delalloc past those
4501          * extents.  so, we trust isize unless the start of the last extent is
4502          * beyond isize
4503          */
4504         if (last < isize) {
4505                 last = (u64)-1;
4506                 last_for_get_extent = isize;
4507         }
4508
4509         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4510                          &cached_state);
4511
4512         em = get_extent_skip_holes(inode, start, last_for_get_extent);
4513         if (!em)
4514                 goto out;
4515         if (IS_ERR(em)) {
4516                 ret = PTR_ERR(em);
4517                 goto out;
4518         }
4519
4520         while (!end) {
4521                 u64 offset_in_extent = 0;
4522
4523                 /* break if the extent we found is outside the range */
4524                 if (em->start >= max || extent_map_end(em) < off)
4525                         break;
4526
4527                 /*
4528                  * get_extent may return an extent that starts before our
4529                  * requested range.  We have to make sure the ranges
4530                  * we return to fiemap always move forward and don't
4531                  * overlap, so adjust the offsets here
4532                  */
4533                 em_start = max(em->start, off);
4534
4535                 /*
4536                  * record the offset from the start of the extent
4537                  * for adjusting the disk offset below.  Only do this if the
4538                  * extent isn't compressed since our in ram offset may be past
4539                  * what we have actually allocated on disk.
4540                  */
4541                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4542                         offset_in_extent = em_start - em->start;
4543                 em_end = extent_map_end(em);
4544                 em_len = em_end - em_start;
4545                 disko = 0;
4546                 flags = 0;
4547
4548                 /*
4549                  * bump off for our next call to get_extent
4550                  */
4551                 off = extent_map_end(em);
4552                 if (off >= max)
4553                         end = 1;
4554
4555                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4556                         end = 1;
4557                         flags |= FIEMAP_EXTENT_LAST;
4558                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4559                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4560                                   FIEMAP_EXTENT_NOT_ALIGNED);
4561                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4562                         flags |= (FIEMAP_EXTENT_DELALLOC |
4563                                   FIEMAP_EXTENT_UNKNOWN);
4564                 } else if (fieinfo->fi_extents_max) {
4565                         u64 bytenr = em->block_start -
4566                                 (em->start - em->orig_start);
4567
4568                         disko = em->block_start + offset_in_extent;
4569
4570                         /*
4571                          * As btrfs supports shared space, this information
4572                          * can be exported to userspace tools via
4573                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4574                          * then we're just getting a count and we can skip the
4575                          * lookup stuff.
4576                          */
4577                         ret = btrfs_check_shared(root,
4578                                                  btrfs_ino(BTRFS_I(inode)),
4579                                                  bytenr);
4580                         if (ret < 0)
4581                                 goto out_free;
4582                         if (ret)
4583                                 flags |= FIEMAP_EXTENT_SHARED;
4584                         ret = 0;
4585                 }
4586                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4587                         flags |= FIEMAP_EXTENT_ENCODED;
4588                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4589                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4590
4591                 free_extent_map(em);
4592                 em = NULL;
4593                 if ((em_start >= last) || em_len == (u64)-1 ||
4594                    (last == (u64)-1 && isize <= em_end)) {
4595                         flags |= FIEMAP_EXTENT_LAST;
4596                         end = 1;
4597                 }
4598
4599                 /* now scan forward to see if this is really the last extent. */
4600                 em = get_extent_skip_holes(inode, off, last_for_get_extent);
4601                 if (IS_ERR(em)) {
4602                         ret = PTR_ERR(em);
4603                         goto out;
4604                 }
4605                 if (!em) {
4606                         flags |= FIEMAP_EXTENT_LAST;
4607                         end = 1;
4608                 }
4609                 ret = emit_fiemap_extent(fieinfo, &cache, em_start, disko,
4610                                            em_len, flags);
4611                 if (ret) {
4612                         if (ret == 1)
4613                                 ret = 0;
4614                         goto out_free;
4615                 }
4616         }
4617 out_free:
4618         if (!ret)
4619                 ret = emit_last_fiemap_cache(root->fs_info, fieinfo, &cache);
4620         free_extent_map(em);
4621 out:
4622         btrfs_free_path(path);
4623         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4624                              &cached_state, GFP_NOFS);
4625         return ret;
4626 }
4627
4628 static void __free_extent_buffer(struct extent_buffer *eb)
4629 {
4630         btrfs_leak_debug_del(&eb->leak_list);
4631         kmem_cache_free(extent_buffer_cache, eb);
4632 }
4633
4634 int extent_buffer_under_io(struct extent_buffer *eb)
4635 {
4636         return (atomic_read(&eb->io_pages) ||
4637                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4638                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4639 }
4640
4641 /*
4642  * Helper for releasing extent buffer page.
4643  */
4644 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4645 {
4646         unsigned long index;
4647         struct page *page;
4648         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4649
4650         BUG_ON(extent_buffer_under_io(eb));
4651
4652         index = num_extent_pages(eb->start, eb->len);
4653         if (index == 0)
4654                 return;
4655
4656         do {
4657                 index--;
4658                 page = eb->pages[index];
4659                 if (!page)
4660                         continue;
4661                 if (mapped)
4662                         spin_lock(&page->mapping->private_lock);
4663                 /*
4664                  * We do this since we'll remove the pages after we've
4665                  * removed the eb from the radix tree, so we could race
4666                  * and have this page now attached to the new eb.  So
4667                  * only clear page_private if it's still connected to
4668                  * this eb.
4669                  */
4670                 if (PagePrivate(page) &&
4671                     page->private == (unsigned long)eb) {
4672                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4673                         BUG_ON(PageDirty(page));
4674                         BUG_ON(PageWriteback(page));
4675                         /*
4676                          * We need to make sure we haven't be attached
4677                          * to a new eb.
4678                          */
4679                         ClearPagePrivate(page);
4680                         set_page_private(page, 0);
4681                         /* One for the page private */
4682                         put_page(page);
4683                 }
4684
4685                 if (mapped)
4686                         spin_unlock(&page->mapping->private_lock);
4687
4688                 /* One for when we allocated the page */
4689                 put_page(page);
4690         } while (index != 0);
4691 }
4692
4693 /*
4694  * Helper for releasing the extent buffer.
4695  */
4696 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4697 {
4698         btrfs_release_extent_buffer_page(eb);
4699         __free_extent_buffer(eb);
4700 }
4701
4702 static struct extent_buffer *
4703 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4704                       unsigned long len)
4705 {
4706         struct extent_buffer *eb = NULL;
4707
4708         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4709         eb->start = start;
4710         eb->len = len;
4711         eb->fs_info = fs_info;
4712         eb->bflags = 0;
4713         rwlock_init(&eb->lock);
4714         atomic_set(&eb->write_locks, 0);
4715         atomic_set(&eb->read_locks, 0);
4716         atomic_set(&eb->blocking_readers, 0);
4717         atomic_set(&eb->blocking_writers, 0);
4718         atomic_set(&eb->spinning_readers, 0);
4719         atomic_set(&eb->spinning_writers, 0);
4720         eb->lock_nested = 0;
4721         init_waitqueue_head(&eb->write_lock_wq);
4722         init_waitqueue_head(&eb->read_lock_wq);
4723
4724         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4725
4726         spin_lock_init(&eb->refs_lock);
4727         atomic_set(&eb->refs, 1);
4728         atomic_set(&eb->io_pages, 0);
4729
4730         /*
4731          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4732          */
4733         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4734                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4735         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4736
4737         return eb;
4738 }
4739
4740 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4741 {
4742         unsigned long i;
4743         struct page *p;
4744         struct extent_buffer *new;
4745         unsigned long num_pages = num_extent_pages(src->start, src->len);
4746
4747         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4748         if (new == NULL)
4749                 return NULL;
4750
4751         for (i = 0; i < num_pages; i++) {
4752                 p = alloc_page(GFP_NOFS);
4753                 if (!p) {
4754                         btrfs_release_extent_buffer(new);
4755                         return NULL;
4756                 }
4757                 attach_extent_buffer_page(new, p);
4758                 WARN_ON(PageDirty(p));
4759                 SetPageUptodate(p);
4760                 new->pages[i] = p;
4761                 copy_page(page_address(p), page_address(src->pages[i]));
4762         }
4763
4764         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4765         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4766
4767         return new;
4768 }
4769
4770 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4771                                                   u64 start, unsigned long len)
4772 {
4773         struct extent_buffer *eb;
4774         unsigned long num_pages;
4775         unsigned long i;
4776
4777         num_pages = num_extent_pages(start, len);
4778
4779         eb = __alloc_extent_buffer(fs_info, start, len);
4780         if (!eb)
4781                 return NULL;
4782
4783         for (i = 0; i < num_pages; i++) {
4784                 eb->pages[i] = alloc_page(GFP_NOFS);
4785                 if (!eb->pages[i])
4786                         goto err;
4787         }
4788         set_extent_buffer_uptodate(eb);
4789         btrfs_set_header_nritems(eb, 0);
4790         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4791
4792         return eb;
4793 err:
4794         for (; i > 0; i--)
4795                 __free_page(eb->pages[i - 1]);
4796         __free_extent_buffer(eb);
4797         return NULL;
4798 }
4799
4800 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4801                                                 u64 start)
4802 {
4803         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4804 }
4805
4806 static void check_buffer_tree_ref(struct extent_buffer *eb)
4807 {
4808         int refs;
4809         /* the ref bit is tricky.  We have to make sure it is set
4810          * if we have the buffer dirty.   Otherwise the
4811          * code to free a buffer can end up dropping a dirty
4812          * page
4813          *
4814          * Once the ref bit is set, it won't go away while the
4815          * buffer is dirty or in writeback, and it also won't
4816          * go away while we have the reference count on the
4817          * eb bumped.
4818          *
4819          * We can't just set the ref bit without bumping the
4820          * ref on the eb because free_extent_buffer might
4821          * see the ref bit and try to clear it.  If this happens
4822          * free_extent_buffer might end up dropping our original
4823          * ref by mistake and freeing the page before we are able
4824          * to add one more ref.
4825          *
4826          * So bump the ref count first, then set the bit.  If someone
4827          * beat us to it, drop the ref we added.
4828          */
4829         refs = atomic_read(&eb->refs);
4830         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4831                 return;
4832
4833         spin_lock(&eb->refs_lock);
4834         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4835                 atomic_inc(&eb->refs);
4836         spin_unlock(&eb->refs_lock);
4837 }
4838
4839 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4840                 struct page *accessed)
4841 {
4842         unsigned long num_pages, i;
4843
4844         check_buffer_tree_ref(eb);
4845
4846         num_pages = num_extent_pages(eb->start, eb->len);
4847         for (i = 0; i < num_pages; i++) {
4848                 struct page *p = eb->pages[i];
4849
4850                 if (p != accessed)
4851                         mark_page_accessed(p);
4852         }
4853 }
4854
4855 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4856                                          u64 start)
4857 {
4858         struct extent_buffer *eb;
4859
4860         rcu_read_lock();
4861         eb = radix_tree_lookup(&fs_info->buffer_radix,
4862                                start >> PAGE_SHIFT);
4863         if (eb && atomic_inc_not_zero(&eb->refs)) {
4864                 rcu_read_unlock();
4865                 /*
4866                  * Lock our eb's refs_lock to avoid races with
4867                  * free_extent_buffer. When we get our eb it might be flagged
4868                  * with EXTENT_BUFFER_STALE and another task running
4869                  * free_extent_buffer might have seen that flag set,
4870                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4871                  * writeback flags not set) and it's still in the tree (flag
4872                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4873                  * of decrementing the extent buffer's reference count twice.
4874                  * So here we could race and increment the eb's reference count,
4875                  * clear its stale flag, mark it as dirty and drop our reference
4876                  * before the other task finishes executing free_extent_buffer,
4877                  * which would later result in an attempt to free an extent
4878                  * buffer that is dirty.
4879                  */
4880                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4881                         spin_lock(&eb->refs_lock);
4882                         spin_unlock(&eb->refs_lock);
4883                 }
4884                 mark_extent_buffer_accessed(eb, NULL);
4885                 return eb;
4886         }
4887         rcu_read_unlock();
4888
4889         return NULL;
4890 }
4891
4892 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4893 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4894                                         u64 start)
4895 {
4896         struct extent_buffer *eb, *exists = NULL;
4897         int ret;
4898
4899         eb = find_extent_buffer(fs_info, start);
4900         if (eb)
4901                 return eb;
4902         eb = alloc_dummy_extent_buffer(fs_info, start);
4903         if (!eb)
4904                 return NULL;
4905         eb->fs_info = fs_info;
4906 again:
4907         ret = radix_tree_preload(GFP_NOFS);
4908         if (ret)
4909                 goto free_eb;
4910         spin_lock(&fs_info->buffer_lock);
4911         ret = radix_tree_insert(&fs_info->buffer_radix,
4912                                 start >> PAGE_SHIFT, eb);
4913         spin_unlock(&fs_info->buffer_lock);
4914         radix_tree_preload_end();
4915         if (ret == -EEXIST) {
4916                 exists = find_extent_buffer(fs_info, start);
4917                 if (exists)
4918                         goto free_eb;
4919                 else
4920                         goto again;
4921         }
4922         check_buffer_tree_ref(eb);
4923         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4924
4925         /*
4926          * We will free dummy extent buffer's if they come into
4927          * free_extent_buffer with a ref count of 2, but if we are using this we
4928          * want the buffers to stay in memory until we're done with them, so
4929          * bump the ref count again.
4930          */
4931         atomic_inc(&eb->refs);
4932         return eb;
4933 free_eb:
4934         btrfs_release_extent_buffer(eb);
4935         return exists;
4936 }
4937 #endif
4938
4939 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4940                                           u64 start)
4941 {
4942         unsigned long len = fs_info->nodesize;
4943         unsigned long num_pages = num_extent_pages(start, len);
4944         unsigned long i;
4945         unsigned long index = start >> PAGE_SHIFT;
4946         struct extent_buffer *eb;
4947         struct extent_buffer *exists = NULL;
4948         struct page *p;
4949         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4950         int uptodate = 1;
4951         int ret;
4952
4953         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
4954                 btrfs_err(fs_info, "bad tree block start %llu", start);
4955                 return ERR_PTR(-EINVAL);
4956         }
4957
4958         eb = find_extent_buffer(fs_info, start);
4959         if (eb)
4960                 return eb;
4961
4962         eb = __alloc_extent_buffer(fs_info, start, len);
4963         if (!eb)
4964                 return ERR_PTR(-ENOMEM);
4965
4966         for (i = 0; i < num_pages; i++, index++) {
4967                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4968                 if (!p) {
4969                         exists = ERR_PTR(-ENOMEM);
4970                         goto free_eb;
4971                 }
4972
4973                 spin_lock(&mapping->private_lock);
4974                 if (PagePrivate(p)) {
4975                         /*
4976                          * We could have already allocated an eb for this page
4977                          * and attached one so lets see if we can get a ref on
4978                          * the existing eb, and if we can we know it's good and
4979                          * we can just return that one, else we know we can just
4980                          * overwrite page->private.
4981                          */
4982                         exists = (struct extent_buffer *)p->private;
4983                         if (atomic_inc_not_zero(&exists->refs)) {
4984                                 spin_unlock(&mapping->private_lock);
4985                                 unlock_page(p);
4986                                 put_page(p);
4987                                 mark_extent_buffer_accessed(exists, p);
4988                                 goto free_eb;
4989                         }
4990                         exists = NULL;
4991
4992                         /*
4993                          * Do this so attach doesn't complain and we need to
4994                          * drop the ref the old guy had.
4995                          */
4996                         ClearPagePrivate(p);
4997                         WARN_ON(PageDirty(p));
4998                         put_page(p);
4999                 }
5000                 attach_extent_buffer_page(eb, p);
5001                 spin_unlock(&mapping->private_lock);
5002                 WARN_ON(PageDirty(p));
5003                 eb->pages[i] = p;
5004                 if (!PageUptodate(p))
5005                         uptodate = 0;
5006
5007                 /*
5008                  * see below about how we avoid a nasty race with release page
5009                  * and why we unlock later
5010                  */
5011         }
5012         if (uptodate)
5013                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5014 again:
5015         ret = radix_tree_preload(GFP_NOFS);
5016         if (ret) {
5017                 exists = ERR_PTR(ret);
5018                 goto free_eb;
5019         }
5020
5021         spin_lock(&fs_info->buffer_lock);
5022         ret = radix_tree_insert(&fs_info->buffer_radix,
5023                                 start >> PAGE_SHIFT, eb);
5024         spin_unlock(&fs_info->buffer_lock);
5025         radix_tree_preload_end();
5026         if (ret == -EEXIST) {
5027                 exists = find_extent_buffer(fs_info, start);
5028                 if (exists)
5029                         goto free_eb;
5030                 else
5031                         goto again;
5032         }
5033         /* add one reference for the tree */
5034         check_buffer_tree_ref(eb);
5035         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5036
5037         /*
5038          * there is a race where release page may have
5039          * tried to find this extent buffer in the radix
5040          * but failed.  It will tell the VM it is safe to
5041          * reclaim the, and it will clear the page private bit.
5042          * We must make sure to set the page private bit properly
5043          * after the extent buffer is in the radix tree so
5044          * it doesn't get lost
5045          */
5046         SetPageChecked(eb->pages[0]);
5047         for (i = 1; i < num_pages; i++) {
5048                 p = eb->pages[i];
5049                 ClearPageChecked(p);
5050                 unlock_page(p);
5051         }
5052         unlock_page(eb->pages[0]);
5053         return eb;
5054
5055 free_eb:
5056         WARN_ON(!atomic_dec_and_test(&eb->refs));
5057         for (i = 0; i < num_pages; i++) {
5058                 if (eb->pages[i])
5059                         unlock_page(eb->pages[i]);
5060         }
5061
5062         btrfs_release_extent_buffer(eb);
5063         return exists;
5064 }
5065
5066 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5067 {
5068         struct extent_buffer *eb =
5069                         container_of(head, struct extent_buffer, rcu_head);
5070
5071         __free_extent_buffer(eb);
5072 }
5073
5074 /* Expects to have eb->eb_lock already held */
5075 static int release_extent_buffer(struct extent_buffer *eb)
5076 {
5077         WARN_ON(atomic_read(&eb->refs) == 0);
5078         if (atomic_dec_and_test(&eb->refs)) {
5079                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5080                         struct btrfs_fs_info *fs_info = eb->fs_info;
5081
5082                         spin_unlock(&eb->refs_lock);
5083
5084                         spin_lock(&fs_info->buffer_lock);
5085                         radix_tree_delete(&fs_info->buffer_radix,
5086                                           eb->start >> PAGE_SHIFT);
5087                         spin_unlock(&fs_info->buffer_lock);
5088                 } else {
5089                         spin_unlock(&eb->refs_lock);
5090                 }
5091
5092                 /* Should be safe to release our pages at this point */
5093                 btrfs_release_extent_buffer_page(eb);
5094 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5095                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5096                         __free_extent_buffer(eb);
5097                         return 1;
5098                 }
5099 #endif
5100                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5101                 return 1;
5102         }
5103         spin_unlock(&eb->refs_lock);
5104
5105         return 0;
5106 }
5107
5108 void free_extent_buffer(struct extent_buffer *eb)
5109 {
5110         int refs;
5111         int old;
5112         if (!eb)
5113                 return;
5114
5115         while (1) {
5116                 refs = atomic_read(&eb->refs);
5117                 if (refs <= 3)
5118                         break;
5119                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5120                 if (old == refs)
5121                         return;
5122         }
5123
5124         spin_lock(&eb->refs_lock);
5125         if (atomic_read(&eb->refs) == 2 &&
5126             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5127                 atomic_dec(&eb->refs);
5128
5129         if (atomic_read(&eb->refs) == 2 &&
5130             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5131             !extent_buffer_under_io(eb) &&
5132             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5133                 atomic_dec(&eb->refs);
5134
5135         /*
5136          * I know this is terrible, but it's temporary until we stop tracking
5137          * the uptodate bits and such for the extent buffers.
5138          */
5139         release_extent_buffer(eb);
5140 }
5141
5142 void free_extent_buffer_stale(struct extent_buffer *eb)
5143 {
5144         if (!eb)
5145                 return;
5146
5147         spin_lock(&eb->refs_lock);
5148         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5149
5150         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5151             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5152                 atomic_dec(&eb->refs);
5153         release_extent_buffer(eb);
5154 }
5155
5156 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5157 {
5158         unsigned long i;
5159         unsigned long num_pages;
5160         struct page *page;
5161
5162         num_pages = num_extent_pages(eb->start, eb->len);
5163
5164         for (i = 0; i < num_pages; i++) {
5165                 page = eb->pages[i];
5166                 if (!PageDirty(page))
5167                         continue;
5168
5169                 lock_page(page);
5170                 WARN_ON(!PagePrivate(page));
5171
5172                 clear_page_dirty_for_io(page);
5173                 spin_lock_irq(&page->mapping->tree_lock);
5174                 if (!PageDirty(page)) {
5175                         radix_tree_tag_clear(&page->mapping->page_tree,
5176                                                 page_index(page),
5177                                                 PAGECACHE_TAG_DIRTY);
5178                 }
5179                 spin_unlock_irq(&page->mapping->tree_lock);
5180                 ClearPageError(page);
5181                 unlock_page(page);
5182         }
5183         WARN_ON(atomic_read(&eb->refs) == 0);
5184 }
5185
5186 int set_extent_buffer_dirty(struct extent_buffer *eb)
5187 {
5188         unsigned long i;
5189         unsigned long num_pages;
5190         int was_dirty = 0;
5191
5192         check_buffer_tree_ref(eb);
5193
5194         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5195
5196         num_pages = num_extent_pages(eb->start, eb->len);
5197         WARN_ON(atomic_read(&eb->refs) == 0);
5198         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5199
5200         for (i = 0; i < num_pages; i++)
5201                 set_page_dirty(eb->pages[i]);
5202         return was_dirty;
5203 }
5204
5205 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5206 {
5207         unsigned long i;
5208         struct page *page;
5209         unsigned long num_pages;
5210
5211         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5212         num_pages = num_extent_pages(eb->start, eb->len);
5213         for (i = 0; i < num_pages; i++) {
5214                 page = eb->pages[i];
5215                 if (page)
5216                         ClearPageUptodate(page);
5217         }
5218 }
5219
5220 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5221 {
5222         unsigned long i;
5223         struct page *page;
5224         unsigned long num_pages;
5225
5226         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5227         num_pages = num_extent_pages(eb->start, eb->len);
5228         for (i = 0; i < num_pages; i++) {
5229                 page = eb->pages[i];
5230                 SetPageUptodate(page);
5231         }
5232 }
5233
5234 int extent_buffer_uptodate(struct extent_buffer *eb)
5235 {
5236         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5237 }
5238
5239 int read_extent_buffer_pages(struct extent_io_tree *tree,
5240                              struct extent_buffer *eb, int wait, int mirror_num)
5241 {
5242         unsigned long i;
5243         struct page *page;
5244         int err;
5245         int ret = 0;
5246         int locked_pages = 0;
5247         int all_uptodate = 1;
5248         unsigned long num_pages;
5249         unsigned long num_reads = 0;
5250         struct bio *bio = NULL;
5251         unsigned long bio_flags = 0;
5252
5253         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5254                 return 0;
5255
5256         num_pages = num_extent_pages(eb->start, eb->len);
5257         for (i = 0; i < num_pages; i++) {
5258                 page = eb->pages[i];
5259                 if (wait == WAIT_NONE) {
5260                         if (!trylock_page(page))
5261                                 goto unlock_exit;
5262                 } else {
5263                         lock_page(page);
5264                 }
5265                 locked_pages++;
5266         }
5267         /*
5268          * We need to firstly lock all pages to make sure that
5269          * the uptodate bit of our pages won't be affected by
5270          * clear_extent_buffer_uptodate().
5271          */
5272         for (i = 0; i < num_pages; i++) {
5273                 page = eb->pages[i];
5274                 if (!PageUptodate(page)) {
5275                         num_reads++;
5276                         all_uptodate = 0;
5277                 }
5278         }
5279
5280         if (all_uptodate) {
5281                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5282                 goto unlock_exit;
5283         }
5284
5285         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5286         eb->read_mirror = 0;
5287         atomic_set(&eb->io_pages, num_reads);
5288         for (i = 0; i < num_pages; i++) {
5289                 page = eb->pages[i];
5290
5291                 if (!PageUptodate(page)) {
5292                         if (ret) {
5293                                 atomic_dec(&eb->io_pages);
5294                                 unlock_page(page);
5295                                 continue;
5296                         }
5297
5298                         ClearPageError(page);
5299                         err = __extent_read_full_page(tree, page,
5300                                                       btree_get_extent, &bio,
5301                                                       mirror_num, &bio_flags,
5302                                                       REQ_META);
5303                         if (err) {
5304                                 ret = err;
5305                                 /*
5306                                  * We use &bio in above __extent_read_full_page,
5307                                  * so we ensure that if it returns error, the
5308                                  * current page fails to add itself to bio and
5309                                  * it's been unlocked.
5310                                  *
5311                                  * We must dec io_pages by ourselves.
5312                                  */
5313                                 atomic_dec(&eb->io_pages);
5314                         }
5315                 } else {
5316                         unlock_page(page);
5317                 }
5318         }
5319
5320         if (bio) {
5321                 err = submit_one_bio(bio, mirror_num, bio_flags);
5322                 if (err)
5323                         return err;
5324         }
5325
5326         if (ret || wait != WAIT_COMPLETE)
5327                 return ret;
5328
5329         for (i = 0; i < num_pages; i++) {
5330                 page = eb->pages[i];
5331                 wait_on_page_locked(page);
5332                 if (!PageUptodate(page))
5333                         ret = -EIO;
5334         }
5335
5336         return ret;
5337
5338 unlock_exit:
5339         while (locked_pages > 0) {
5340                 locked_pages--;
5341                 page = eb->pages[locked_pages];
5342                 unlock_page(page);
5343         }
5344         return ret;
5345 }
5346
5347 void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
5348                         unsigned long start, unsigned long len)
5349 {
5350         size_t cur;
5351         size_t offset;
5352         struct page *page;
5353         char *kaddr;
5354         char *dst = (char *)dstv;
5355         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5356         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5357
5358         if (start + len > eb->len) {
5359                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5360                      eb->start, eb->len, start, len);
5361                 memset(dst, 0, len);
5362                 return;
5363         }
5364
5365         offset = (start_offset + start) & (PAGE_SIZE - 1);
5366
5367         while (len > 0) {
5368                 page = eb->pages[i];
5369
5370                 cur = min(len, (PAGE_SIZE - offset));
5371                 kaddr = page_address(page);
5372                 memcpy(dst, kaddr + offset, cur);
5373
5374                 dst += cur;
5375                 len -= cur;
5376                 offset = 0;
5377                 i++;
5378         }
5379 }
5380
5381 int read_extent_buffer_to_user(const struct extent_buffer *eb,
5382                                void __user *dstv,
5383                                unsigned long start, unsigned long len)
5384 {
5385         size_t cur;
5386         size_t offset;
5387         struct page *page;
5388         char *kaddr;
5389         char __user *dst = (char __user *)dstv;
5390         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5391         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5392         int ret = 0;
5393
5394         WARN_ON(start > eb->len);
5395         WARN_ON(start + len > eb->start + eb->len);
5396
5397         offset = (start_offset + start) & (PAGE_SIZE - 1);
5398
5399         while (len > 0) {
5400                 page = eb->pages[i];
5401
5402                 cur = min(len, (PAGE_SIZE - offset));
5403                 kaddr = page_address(page);
5404                 if (copy_to_user(dst, kaddr + offset, cur)) {
5405                         ret = -EFAULT;
5406                         break;
5407                 }
5408
5409                 dst += cur;
5410                 len -= cur;
5411                 offset = 0;
5412                 i++;
5413         }
5414
5415         return ret;
5416 }
5417
5418 /*
5419  * return 0 if the item is found within a page.
5420  * return 1 if the item spans two pages.
5421  * return -EINVAL otherwise.
5422  */
5423 int map_private_extent_buffer(const struct extent_buffer *eb,
5424                               unsigned long start, unsigned long min_len,
5425                               char **map, unsigned long *map_start,
5426                               unsigned long *map_len)
5427 {
5428         size_t offset = start & (PAGE_SIZE - 1);
5429         char *kaddr;
5430         struct page *p;
5431         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5432         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5433         unsigned long end_i = (start_offset + start + min_len - 1) >>
5434                 PAGE_SHIFT;
5435
5436         if (start + min_len > eb->len) {
5437                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5438                        eb->start, eb->len, start, min_len);
5439                 return -EINVAL;
5440         }
5441
5442         if (i != end_i)
5443                 return 1;
5444
5445         if (i == 0) {
5446                 offset = start_offset;
5447                 *map_start = 0;
5448         } else {
5449                 offset = 0;
5450                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5451         }
5452
5453         p = eb->pages[i];
5454         kaddr = page_address(p);
5455         *map = kaddr + offset;
5456         *map_len = PAGE_SIZE - offset;
5457         return 0;
5458 }
5459
5460 int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
5461                          unsigned long start, unsigned long len)
5462 {
5463         size_t cur;
5464         size_t offset;
5465         struct page *page;
5466         char *kaddr;
5467         char *ptr = (char *)ptrv;
5468         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5469         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5470         int ret = 0;
5471
5472         WARN_ON(start > eb->len);
5473         WARN_ON(start + len > eb->start + eb->len);
5474
5475         offset = (start_offset + start) & (PAGE_SIZE - 1);
5476
5477         while (len > 0) {
5478                 page = eb->pages[i];
5479
5480                 cur = min(len, (PAGE_SIZE - offset));
5481
5482                 kaddr = page_address(page);
5483                 ret = memcmp(ptr, kaddr + offset, cur);
5484                 if (ret)
5485                         break;
5486
5487                 ptr += cur;
5488                 len -= cur;
5489                 offset = 0;
5490                 i++;
5491         }
5492         return ret;
5493 }
5494
5495 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5496                 const void *srcv)
5497 {
5498         char *kaddr;
5499
5500         WARN_ON(!PageUptodate(eb->pages[0]));
5501         kaddr = page_address(eb->pages[0]);
5502         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5503                         BTRFS_FSID_SIZE);
5504 }
5505
5506 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv)
5507 {
5508         char *kaddr;
5509
5510         WARN_ON(!PageUptodate(eb->pages[0]));
5511         kaddr = page_address(eb->pages[0]);
5512         memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5513                         BTRFS_FSID_SIZE);
5514 }
5515
5516 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5517                          unsigned long start, unsigned long len)
5518 {
5519         size_t cur;
5520         size_t offset;
5521         struct page *page;
5522         char *kaddr;
5523         char *src = (char *)srcv;
5524         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5525         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5526
5527         WARN_ON(start > eb->len);
5528         WARN_ON(start + len > eb->start + eb->len);
5529
5530         offset = (start_offset + start) & (PAGE_SIZE - 1);
5531
5532         while (len > 0) {
5533                 page = eb->pages[i];
5534                 WARN_ON(!PageUptodate(page));
5535
5536                 cur = min(len, PAGE_SIZE - offset);
5537                 kaddr = page_address(page);
5538                 memcpy(kaddr + offset, src, cur);
5539
5540                 src += cur;
5541                 len -= cur;
5542                 offset = 0;
5543                 i++;
5544         }
5545 }
5546
5547 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5548                 unsigned long len)
5549 {
5550         size_t cur;
5551         size_t offset;
5552         struct page *page;
5553         char *kaddr;
5554         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5555         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5556
5557         WARN_ON(start > eb->len);
5558         WARN_ON(start + len > eb->start + eb->len);
5559
5560         offset = (start_offset + start) & (PAGE_SIZE - 1);
5561
5562         while (len > 0) {
5563                 page = eb->pages[i];
5564                 WARN_ON(!PageUptodate(page));
5565
5566                 cur = min(len, PAGE_SIZE - offset);
5567                 kaddr = page_address(page);
5568                 memset(kaddr + offset, 0, cur);
5569
5570                 len -= cur;
5571                 offset = 0;
5572                 i++;
5573         }
5574 }
5575
5576 void copy_extent_buffer_full(struct extent_buffer *dst,
5577                              struct extent_buffer *src)
5578 {
5579         int i;
5580         unsigned num_pages;
5581
5582         ASSERT(dst->len == src->len);
5583
5584         num_pages = num_extent_pages(dst->start, dst->len);
5585         for (i = 0; i < num_pages; i++)
5586                 copy_page(page_address(dst->pages[i]),
5587                                 page_address(src->pages[i]));
5588 }
5589
5590 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5591                         unsigned long dst_offset, unsigned long src_offset,
5592                         unsigned long len)
5593 {
5594         u64 dst_len = dst->len;
5595         size_t cur;
5596         size_t offset;
5597         struct page *page;
5598         char *kaddr;
5599         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5600         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5601
5602         WARN_ON(src->len != dst_len);
5603
5604         offset = (start_offset + dst_offset) &
5605                 (PAGE_SIZE - 1);
5606
5607         while (len > 0) {
5608                 page = dst->pages[i];
5609                 WARN_ON(!PageUptodate(page));
5610
5611                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5612
5613                 kaddr = page_address(page);
5614                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5615
5616                 src_offset += cur;
5617                 len -= cur;
5618                 offset = 0;
5619                 i++;
5620         }
5621 }
5622
5623 void le_bitmap_set(u8 *map, unsigned int start, int len)
5624 {
5625         u8 *p = map + BIT_BYTE(start);
5626         const unsigned int size = start + len;
5627         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5628         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5629
5630         while (len - bits_to_set >= 0) {
5631                 *p |= mask_to_set;
5632                 len -= bits_to_set;
5633                 bits_to_set = BITS_PER_BYTE;
5634                 mask_to_set = ~0;
5635                 p++;
5636         }
5637         if (len) {
5638                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5639                 *p |= mask_to_set;
5640         }
5641 }
5642
5643 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5644 {
5645         u8 *p = map + BIT_BYTE(start);
5646         const unsigned int size = start + len;
5647         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5648         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5649
5650         while (len - bits_to_clear >= 0) {
5651                 *p &= ~mask_to_clear;
5652                 len -= bits_to_clear;
5653                 bits_to_clear = BITS_PER_BYTE;
5654                 mask_to_clear = ~0;
5655                 p++;
5656         }
5657         if (len) {
5658                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5659                 *p &= ~mask_to_clear;
5660         }
5661 }
5662
5663 /*
5664  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5665  * given bit number
5666  * @eb: the extent buffer
5667  * @start: offset of the bitmap item in the extent buffer
5668  * @nr: bit number
5669  * @page_index: return index of the page in the extent buffer that contains the
5670  * given bit number
5671  * @page_offset: return offset into the page given by page_index
5672  *
5673  * This helper hides the ugliness of finding the byte in an extent buffer which
5674  * contains a given bit.
5675  */
5676 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5677                                     unsigned long start, unsigned long nr,
5678                                     unsigned long *page_index,
5679                                     size_t *page_offset)
5680 {
5681         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5682         size_t byte_offset = BIT_BYTE(nr);
5683         size_t offset;
5684
5685         /*
5686          * The byte we want is the offset of the extent buffer + the offset of
5687          * the bitmap item in the extent buffer + the offset of the byte in the
5688          * bitmap item.
5689          */
5690         offset = start_offset + start + byte_offset;
5691
5692         *page_index = offset >> PAGE_SHIFT;
5693         *page_offset = offset & (PAGE_SIZE - 1);
5694 }
5695
5696 /**
5697  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5698  * @eb: the extent buffer
5699  * @start: offset of the bitmap item in the extent buffer
5700  * @nr: bit number to test
5701  */
5702 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5703                            unsigned long nr)
5704 {
5705         u8 *kaddr;
5706         struct page *page;
5707         unsigned long i;
5708         size_t offset;
5709
5710         eb_bitmap_offset(eb, start, nr, &i, &offset);
5711         page = eb->pages[i];
5712         WARN_ON(!PageUptodate(page));
5713         kaddr = page_address(page);
5714         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5715 }
5716
5717 /**
5718  * extent_buffer_bitmap_set - set an area of a bitmap
5719  * @eb: the extent buffer
5720  * @start: offset of the bitmap item in the extent buffer
5721  * @pos: bit number of the first bit
5722  * @len: number of bits to set
5723  */
5724 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5725                               unsigned long pos, unsigned long len)
5726 {
5727         u8 *kaddr;
5728         struct page *page;
5729         unsigned long i;
5730         size_t offset;
5731         const unsigned int size = pos + len;
5732         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5733         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5734
5735         eb_bitmap_offset(eb, start, pos, &i, &offset);
5736         page = eb->pages[i];
5737         WARN_ON(!PageUptodate(page));
5738         kaddr = page_address(page);
5739
5740         while (len >= bits_to_set) {
5741                 kaddr[offset] |= mask_to_set;
5742                 len -= bits_to_set;
5743                 bits_to_set = BITS_PER_BYTE;
5744                 mask_to_set = ~0;
5745                 if (++offset >= PAGE_SIZE && len > 0) {
5746                         offset = 0;
5747                         page = eb->pages[++i];
5748                         WARN_ON(!PageUptodate(page));
5749                         kaddr = page_address(page);
5750                 }
5751         }
5752         if (len) {
5753                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5754                 kaddr[offset] |= mask_to_set;
5755         }
5756 }
5757
5758
5759 /**
5760  * extent_buffer_bitmap_clear - clear an area of a bitmap
5761  * @eb: the extent buffer
5762  * @start: offset of the bitmap item in the extent buffer
5763  * @pos: bit number of the first bit
5764  * @len: number of bits to clear
5765  */
5766 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5767                                 unsigned long pos, unsigned long len)
5768 {
5769         u8 *kaddr;
5770         struct page *page;
5771         unsigned long i;
5772         size_t offset;
5773         const unsigned int size = pos + len;
5774         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5775         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5776
5777         eb_bitmap_offset(eb, start, pos, &i, &offset);
5778         page = eb->pages[i];
5779         WARN_ON(!PageUptodate(page));
5780         kaddr = page_address(page);
5781
5782         while (len >= bits_to_clear) {
5783                 kaddr[offset] &= ~mask_to_clear;
5784                 len -= bits_to_clear;
5785                 bits_to_clear = BITS_PER_BYTE;
5786                 mask_to_clear = ~0;
5787                 if (++offset >= PAGE_SIZE && len > 0) {
5788                         offset = 0;
5789                         page = eb->pages[++i];
5790                         WARN_ON(!PageUptodate(page));
5791                         kaddr = page_address(page);
5792                 }
5793         }
5794         if (len) {
5795                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5796                 kaddr[offset] &= ~mask_to_clear;
5797         }
5798 }
5799
5800 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5801 {
5802         unsigned long distance = (src > dst) ? src - dst : dst - src;
5803         return distance < len;
5804 }
5805
5806 static void copy_pages(struct page *dst_page, struct page *src_page,
5807                        unsigned long dst_off, unsigned long src_off,
5808                        unsigned long len)
5809 {
5810         char *dst_kaddr = page_address(dst_page);
5811         char *src_kaddr;
5812         int must_memmove = 0;
5813
5814         if (dst_page != src_page) {
5815                 src_kaddr = page_address(src_page);
5816         } else {
5817                 src_kaddr = dst_kaddr;
5818                 if (areas_overlap(src_off, dst_off, len))
5819                         must_memmove = 1;
5820         }
5821
5822         if (must_memmove)
5823                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5824         else
5825                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5826 }
5827
5828 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5829                            unsigned long src_offset, unsigned long len)
5830 {
5831         struct btrfs_fs_info *fs_info = dst->fs_info;
5832         size_t cur;
5833         size_t dst_off_in_page;
5834         size_t src_off_in_page;
5835         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5836         unsigned long dst_i;
5837         unsigned long src_i;
5838
5839         if (src_offset + len > dst->len) {
5840                 btrfs_err(fs_info,
5841                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5842                          src_offset, len, dst->len);
5843                 BUG_ON(1);
5844         }
5845         if (dst_offset + len > dst->len) {
5846                 btrfs_err(fs_info,
5847                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5848                          dst_offset, len, dst->len);
5849                 BUG_ON(1);
5850         }
5851
5852         while (len > 0) {
5853                 dst_off_in_page = (start_offset + dst_offset) &
5854                         (PAGE_SIZE - 1);
5855                 src_off_in_page = (start_offset + src_offset) &
5856                         (PAGE_SIZE - 1);
5857
5858                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5859                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5860
5861                 cur = min(len, (unsigned long)(PAGE_SIZE -
5862                                                src_off_in_page));
5863                 cur = min_t(unsigned long, cur,
5864                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5865
5866                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5867                            dst_off_in_page, src_off_in_page, cur);
5868
5869                 src_offset += cur;
5870                 dst_offset += cur;
5871                 len -= cur;
5872         }
5873 }
5874
5875 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5876                            unsigned long src_offset, unsigned long len)
5877 {
5878         struct btrfs_fs_info *fs_info = dst->fs_info;
5879         size_t cur;
5880         size_t dst_off_in_page;
5881         size_t src_off_in_page;
5882         unsigned long dst_end = dst_offset + len - 1;
5883         unsigned long src_end = src_offset + len - 1;
5884         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5885         unsigned long dst_i;
5886         unsigned long src_i;
5887
5888         if (src_offset + len > dst->len) {
5889                 btrfs_err(fs_info,
5890                           "memmove bogus src_offset %lu move len %lu len %lu",
5891                           src_offset, len, dst->len);
5892                 BUG_ON(1);
5893         }
5894         if (dst_offset + len > dst->len) {
5895                 btrfs_err(fs_info,
5896                           "memmove bogus dst_offset %lu move len %lu len %lu",
5897                           dst_offset, len, dst->len);
5898                 BUG_ON(1);
5899         }
5900         if (dst_offset < src_offset) {
5901                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5902                 return;
5903         }
5904         while (len > 0) {
5905                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5906                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5907
5908                 dst_off_in_page = (start_offset + dst_end) &
5909                         (PAGE_SIZE - 1);
5910                 src_off_in_page = (start_offset + src_end) &
5911                         (PAGE_SIZE - 1);
5912
5913                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5914                 cur = min(cur, dst_off_in_page + 1);
5915                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5916                            dst_off_in_page - cur + 1,
5917                            src_off_in_page - cur + 1, cur);
5918
5919                 dst_end -= cur;
5920                 src_end -= cur;
5921                 len -= cur;
5922         }
5923 }
5924
5925 int try_release_extent_buffer(struct page *page)
5926 {
5927         struct extent_buffer *eb;
5928
5929         /*
5930          * We need to make sure nobody is attaching this page to an eb right
5931          * now.
5932          */
5933         spin_lock(&page->mapping->private_lock);
5934         if (!PagePrivate(page)) {
5935                 spin_unlock(&page->mapping->private_lock);
5936                 return 1;
5937         }
5938
5939         eb = (struct extent_buffer *)page->private;
5940         BUG_ON(!eb);
5941
5942         /*
5943          * This is a little awful but should be ok, we need to make sure that
5944          * the eb doesn't disappear out from under us while we're looking at
5945          * this page.
5946          */
5947         spin_lock(&eb->refs_lock);
5948         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5949                 spin_unlock(&eb->refs_lock);
5950                 spin_unlock(&page->mapping->private_lock);
5951                 return 0;
5952         }
5953         spin_unlock(&page->mapping->private_lock);
5954
5955         /*
5956          * If tree ref isn't set then we know the ref on this eb is a real ref,
5957          * so just return, this page will likely be freed soon anyway.
5958          */
5959         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5960                 spin_unlock(&eb->refs_lock);
5961                 return 0;
5962         }
5963
5964         return release_extent_buffer(eb);
5965 }