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
  25 static struct kmem_cache *extent_state_cache;
  26 static struct kmem_cache *extent_buffer_cache;
  27 static struct bio_set *btrfs_bioset;
  28
  29 static inline bool extent_state_in_tree(const struct extent_state *state)
  30 {
  31         return !RB_EMPTY_NODE(&state->rb_node);
  32 }
  33
  34 #ifdef CONFIG_BTRFS_DEBUG
  35 static LIST_HEAD(buffers);
  36 static LIST_HEAD(states);
  37
  38 static DEFINE_SPINLOCK(leak_lock);
  39
  40 static inline
  41 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  42 {
  43         unsigned long flags;
  44
  45         spin_lock_irqsave(&leak_lock, flags);
  46         list_add(new, head);
  47         spin_unlock_irqrestore(&leak_lock, flags);
  48 }
  49
  50 static inline
  51 void btrfs_leak_debug_del(struct list_head *entry)
  52 {
  53         unsigned long flags;
  54
  55         spin_lock_irqsave(&leak_lock, flags);
  56         list_del(entry);
  57         spin_unlock_irqrestore(&leak_lock, flags);
  58 }
  59
  60 static inline
  61 void btrfs_leak_debug_check(void)
  62 {
  63         struct extent_state *state;
  64         struct extent_buffer *eb;
  65
  66         while (!list_empty(&states)) {
  67                 state = list_entry(states.next, struct extent_state, leak_list);
  68                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  69                        state->start, state->end, state->state,
  70                        extent_state_in_tree(state),
  71                        refcount_read(&state->refs));
  72                 list_del(&state->leak_list);
  73                 kmem_cache_free(extent_state_cache, state);
  74         }
  75
  76         while (!list_empty(&buffers)) {
  77                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  78                 pr_err("BTRFS: buffer leak start %llu len %lu refs %d\n",
  79                        eb->start, eb->len, atomic_read(&eb->refs));
  80                 list_del(&eb->leak_list);
  81                 kmem_cache_free(extent_buffer_cache, eb);
  82         }
  83 }
  84
  85 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  86         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  87 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  88                 struct extent_io_tree *tree, u64 start, u64 end)
  89 {
  90         if (tree->ops && tree->ops->check_extent_io_range)
  91                 tree->ops->check_extent_io_range(tree->private_data, caller,
  92                                                  start, end);
  93 }
  94 #else
  95 #define btrfs_leak_debug_add(new, head) do {} while (0)
  96 #define btrfs_leak_debug_del(entry)     do {} while (0)
  97 #define btrfs_leak_debug_check()        do {} while (0)
  98 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
  99 #endif
 100
 101 #define BUFFER_LRU_MAX 64
 102
 103 struct tree_entry {
 104         u64 start;
 105         u64 end;
 106         struct rb_node rb_node;
 107 };
 108
 109 struct extent_page_data {
 110         struct bio *bio;
 111         struct extent_io_tree *tree;
 112         get_extent_t *get_extent;
 113         unsigned long bio_flags;
 114
 115         /* tells writepage not to lock the state bits for this range
 116          * it still does the unlocking
 117          */
 118         unsigned int extent_locked:1;
 119
 120         /* tells the submit_bio code to use REQ_SYNC */
 121         unsigned int sync_io:1;
 122 };
 123
 124 static void add_extent_changeset(struct extent_state *state, unsigned bits,
 125                                  struct extent_changeset *changeset,
 126                                  int set)
 127 {
 128         int ret;
 129
 130         if (!changeset)
 131                 return;
 132         if (set && (state->state & bits) == bits)
 133                 return;
 134         if (!set && (state->state & bits) == 0)
 135                 return;
 136         changeset->bytes_changed += state->end - state->start + 1;
 137         ret = ulist_add(&changeset->range_changed, state->start, state->end,
 138                         GFP_ATOMIC);
 139         /* ENOMEM */
 140         BUG_ON(ret < 0);
 141 }
 142
 143 static noinline void flush_write_bio(void *data);
 144 static inline struct btrfs_fs_info *
 145 tree_fs_info(struct extent_io_tree *tree)
 146 {
 147         if (tree->ops)
 148                 return tree->ops->tree_fs_info(tree->private_data);
 149         return NULL;
 150 }
 151
 152 int __init extent_io_init(void)
 153 {
 154         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 155                         sizeof(struct extent_state), 0,
 156                         SLAB_MEM_SPREAD, NULL);
 157         if (!extent_state_cache)
 158                 return -ENOMEM;
 159
 160         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 161                         sizeof(struct extent_buffer), 0,
 162                         SLAB_MEM_SPREAD, NULL);
 163         if (!extent_buffer_cache)
 164                 goto free_state_cache;
 165
 166         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 167                                      offsetof(struct btrfs_io_bio, bio),
 168                                      BIOSET_NEED_BVECS);
 169         if (!btrfs_bioset)
 170                 goto free_buffer_cache;
 171
 172         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 173                 goto free_bioset;
 174
 175         return 0;
 176
 177 free_bioset:
 178         bioset_free(btrfs_bioset);
 179         btrfs_bioset = NULL;
 180
 181 free_buffer_cache:
 182         kmem_cache_destroy(extent_buffer_cache);
 183         extent_buffer_cache = NULL;
 184
 185 free_state_cache:
 186         kmem_cache_destroy(extent_state_cache);
 187         extent_state_cache = NULL;
 188         return -ENOMEM;
 189 }
 190
 191 void extent_io_exit(void)
 192 {
 193         btrfs_leak_debug_check();
 194
 195         /*
 196          * Make sure all delayed rcu free are flushed before we
 197          * destroy caches.
 198          */
 199         rcu_barrier();
 200         kmem_cache_destroy(extent_state_cache);
 201         kmem_cache_destroy(extent_buffer_cache);
 202         if (btrfs_bioset)
 203                 bioset_free(btrfs_bioset);
 204 }
 205
 206 void extent_io_tree_init(struct extent_io_tree *tree,
 207                          void *private_data)
 208 {
 209         tree->state = RB_ROOT;
 210         tree->ops = NULL;
 211         tree->dirty_bytes = 0;
 212         spin_lock_init(&tree->lock);
 213         tree->private_data = private_data;
 214 }
 215
 216 static struct extent_state *alloc_extent_state(gfp_t mask)
 217 {
 218         struct extent_state *state;
 219
 220         /*
 221          * The given mask might be not appropriate for the slab allocator,
 222          * drop the unsupported bits
 223          */
 224         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
 225         state = kmem_cache_alloc(extent_state_cache, mask);
 226         if (!state)
 227                 return state;
 228         state->state = 0;
 229         state->failrec = NULL;
 230         RB_CLEAR_NODE(&state->rb_node);
 231         btrfs_leak_debug_add(&state->leak_list, &states);
 232         refcount_set(&state->refs, 1);
 233         init_waitqueue_head(&state->wq);
 234         trace_alloc_extent_state(state, mask, _RET_IP_);
 235         return state;
 236 }
 237
 238 void free_extent_state(struct extent_state *state)
 239 {
 240         if (!state)
 241                 return;
 242         if (refcount_dec_and_test(&state->refs)) {
 243                 WARN_ON(extent_state_in_tree(state));
 244                 btrfs_leak_debug_del(&state->leak_list);
 245                 trace_free_extent_state(state, _RET_IP_);
 246                 kmem_cache_free(extent_state_cache, state);
 247         }
 248 }
 249
 250 static struct rb_node *tree_insert(struct rb_root *root,
 251                                    struct rb_node *search_start,
 252                                    u64 offset,
 253                                    struct rb_node *node,
 254                                    struct rb_node ***p_in,
 255                                    struct rb_node **parent_in)
 256 {
 257         struct rb_node **p;
 258         struct rb_node *parent = NULL;
 259         struct tree_entry *entry;
 260
 261         if (p_in && parent_in) {
 262                 p = *p_in;
 263                 parent = *parent_in;
 264                 goto do_insert;
 265         }
 266
 267         p = search_start ? &search_start : &root->rb_node;
 268         while (*p) {
 269                 parent = *p;
 270                 entry = rb_entry(parent, struct tree_entry, rb_node);
 271
 272                 if (offset < entry->start)
 273                         p = &(*p)->rb_left;
 274                 else if (offset > entry->end)
 275                         p = &(*p)->rb_right;
 276                 else
 277                         return parent;
 278         }
 279
 280 do_insert:
 281         rb_link_node(node, parent, p);
 282         rb_insert_color(node, root);
 283         return NULL;
 284 }
 285
 286 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 287                                       struct rb_node **prev_ret,
 288                                       struct rb_node **next_ret,
 289                                       struct rb_node ***p_ret,
 290                                       struct rb_node **parent_ret)
 291 {
 292         struct rb_root *root = &tree->state;
 293         struct rb_node **n = &root->rb_node;
 294         struct rb_node *prev = NULL;
 295         struct rb_node *orig_prev = NULL;
 296         struct tree_entry *entry;
 297         struct tree_entry *prev_entry = NULL;
 298
 299         while (*n) {
 300                 prev = *n;
 301                 entry = rb_entry(prev, struct tree_entry, rb_node);
 302                 prev_entry = entry;
 303
 304                 if (offset < entry->start)
 305                         n = &(*n)->rb_left;
 306                 else if (offset > entry->end)
 307                         n = &(*n)->rb_right;
 308                 else
 309                         return *n;
 310         }
 311
 312         if (p_ret)
 313                 *p_ret = n;
 314         if (parent_ret)
 315                 *parent_ret = prev;
 316
 317         if (prev_ret) {
 318                 orig_prev = prev;
 319                 while (prev && offset > prev_entry->end) {
 320                         prev = rb_next(prev);
 321                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 322                 }
 323                 *prev_ret = prev;
 324                 prev = orig_prev;
 325         }
 326
 327         if (next_ret) {
 328                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 329                 while (prev && offset < prev_entry->start) {
 330                         prev = rb_prev(prev);
 331                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 332                 }
 333                 *next_ret = prev;
 334         }
 335         return NULL;
 336 }
 337
 338 static inline struct rb_node *
 339 tree_search_for_insert(struct extent_io_tree *tree,
 340                        u64 offset,
 341                        struct rb_node ***p_ret,
 342                        struct rb_node **parent_ret)
 343 {
 344         struct rb_node *prev = NULL;
 345         struct rb_node *ret;
 346
 347         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 348         if (!ret)
 349                 return prev;
 350         return ret;
 351 }
 352
 353 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 354                                           u64 offset)
 355 {
 356         return tree_search_for_insert(tree, offset, NULL, NULL);
 357 }
 358
 359 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 360                      struct extent_state *other)
 361 {
 362         if (tree->ops && tree->ops->merge_extent_hook)
 363                 tree->ops->merge_extent_hook(tree->private_data, new, other);
 364 }
 365
 366 /*
 367  * utility function to look for merge candidates inside a given range.
 368  * Any extents with matching state are merged together into a single
 369  * extent in the tree.  Extents with EXTENT_IO in their state field
 370  * are not merged because the end_io handlers need to be able to do
 371  * operations on them without sleeping (or doing allocations/splits).
 372  *
 373  * This should be called with the tree lock held.
 374  */
 375 static void merge_state(struct extent_io_tree *tree,
 376                         struct extent_state *state)
 377 {
 378         struct extent_state *other;
 379         struct rb_node *other_node;
 380
 381         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 382                 return;
 383
 384         other_node = rb_prev(&state->rb_node);
 385         if (other_node) {
 386                 other = rb_entry(other_node, struct extent_state, rb_node);
 387                 if (other->end == state->start - 1 &&
 388                     other->state == state->state) {
 389                         merge_cb(tree, state, other);
 390                         state->start = other->start;
 391                         rb_erase(&other->rb_node, &tree->state);
 392                         RB_CLEAR_NODE(&other->rb_node);
 393                         free_extent_state(other);
 394                 }
 395         }
 396         other_node = rb_next(&state->rb_node);
 397         if (other_node) {
 398                 other = rb_entry(other_node, struct extent_state, rb_node);
 399                 if (other->start == state->end + 1 &&
 400                     other->state == state->state) {
 401                         merge_cb(tree, state, other);
 402                         state->end = other->end;
 403                         rb_erase(&other->rb_node, &tree->state);
 404                         RB_CLEAR_NODE(&other->rb_node);
 405                         free_extent_state(other);
 406                 }
 407         }
 408 }
 409
 410 static void set_state_cb(struct extent_io_tree *tree,
 411                          struct extent_state *state, unsigned *bits)
 412 {
 413         if (tree->ops && tree->ops->set_bit_hook)
 414                 tree->ops->set_bit_hook(tree->private_data, state, bits);
 415 }
 416
 417 static void clear_state_cb(struct extent_io_tree *tree,
 418                            struct extent_state *state, unsigned *bits)
 419 {
 420         if (tree->ops && tree->ops->clear_bit_hook)
 421                 tree->ops->clear_bit_hook(tree->private_data, state, bits);
 422 }
 423
 424 static void set_state_bits(struct extent_io_tree *tree,
 425                            struct extent_state *state, unsigned *bits,
 426                            struct extent_changeset *changeset);
 427
 428 /*
 429  * insert an extent_state struct into the tree.  'bits' are set on the
 430  * struct before it is inserted.
 431  *
 432  * This may return -EEXIST if the extent is already there, in which case the
 433  * state struct is freed.
 434  *
 435  * The tree lock is not taken internally.  This is a utility function and
 436  * probably isn't what you want to call (see set/clear_extent_bit).
 437  */
 438 static int insert_state(struct extent_io_tree *tree,
 439                         struct extent_state *state, u64 start, u64 end,
 440                         struct rb_node ***p,
 441                         struct rb_node **parent,
 442                         unsigned *bits, struct extent_changeset *changeset)
 443 {
 444         struct rb_node *node;
 445
 446         if (end < start)
 447                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 448                        end, start);
 449         state->start = start;
 450         state->end = end;
 451
 452         set_state_bits(tree, state, bits, changeset);
 453
 454         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 455         if (node) {
 456                 struct extent_state *found;
 457                 found = rb_entry(node, struct extent_state, rb_node);
 458                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
 459                        found->start, found->end, start, end);
 460                 return -EEXIST;
 461         }
 462         merge_state(tree, state);
 463         return 0;
 464 }
 465
 466 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 467                      u64 split)
 468 {
 469         if (tree->ops && tree->ops->split_extent_hook)
 470                 tree->ops->split_extent_hook(tree->private_data, orig, split);
 471 }
 472
 473 /*
 474  * split a given extent state struct in two, inserting the preallocated
 475  * struct 'prealloc' as the newly created second half.  'split' indicates an
 476  * offset inside 'orig' where it should be split.
 477  *
 478  * Before calling,
 479  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 480  * are two extent state structs in the tree:
 481  * prealloc: [orig->start, split - 1]
 482  * orig: [ split, orig->end ]
 483  *
 484  * The tree locks are not taken by this function. They need to be held
 485  * by the caller.
 486  */
 487 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 488                        struct extent_state *prealloc, u64 split)
 489 {
 490         struct rb_node *node;
 491
 492         split_cb(tree, orig, split);
 493
 494         prealloc->start = orig->start;
 495         prealloc->end = split - 1;
 496         prealloc->state = orig->state;
 497         orig->start = split;
 498
 499         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 500                            &prealloc->rb_node, NULL, NULL);
 501         if (node) {
 502                 free_extent_state(prealloc);
 503                 return -EEXIST;
 504         }
 505         return 0;
 506 }
 507
 508 static struct extent_state *next_state(struct extent_state *state)
 509 {
 510         struct rb_node *next = rb_next(&state->rb_node);
 511         if (next)
 512                 return rb_entry(next, struct extent_state, rb_node);
 513         else
 514                 return NULL;
 515 }
 516
 517 /*
 518  * utility function to clear some bits in an extent state struct.
 519  * it will optionally wake up any one waiting on this state (wake == 1).
 520  *
 521  * If no bits are set on the state struct after clearing things, the
 522  * struct is freed and removed from the tree
 523  */
 524 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 525                                             struct extent_state *state,
 526                                             unsigned *bits, int wake,
 527                                             struct extent_changeset *changeset)
 528 {
 529         struct extent_state *next;
 530         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 531
 532         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 533                 u64 range = state->end - state->start + 1;
 534                 WARN_ON(range > tree->dirty_bytes);
 535                 tree->dirty_bytes -= range;
 536         }
 537         clear_state_cb(tree, state, bits);
 538         add_extent_changeset(state, bits_to_clear, changeset, 0);
 539         state->state &= ~bits_to_clear;
 540         if (wake)
 541                 wake_up(&state->wq);
 542         if (state->state == 0) {
 543                 next = next_state(state);
 544                 if (extent_state_in_tree(state)) {
 545                         rb_erase(&state->rb_node, &tree->state);
 546                         RB_CLEAR_NODE(&state->rb_node);
 547                         free_extent_state(state);
 548                 } else {
 549                         WARN_ON(1);
 550                 }
 551         } else {
 552                 merge_state(tree, state);
 553                 next = next_state(state);
 554         }
 555         return next;
 556 }
 557
 558 static struct extent_state *
 559 alloc_extent_state_atomic(struct extent_state *prealloc)
 560 {
 561         if (!prealloc)
 562                 prealloc = alloc_extent_state(GFP_ATOMIC);
 563
 564         return prealloc;
 565 }
 566
 567 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 568 {
 569         btrfs_panic(tree_fs_info(tree), err,
 570                     "Locking error: Extent tree was modified by another thread while locked.");
 571 }
 572
 573 /*
 574  * clear some bits on a range in the tree.  This may require splitting
 575  * or inserting elements in the tree, so the gfp mask is used to
 576  * indicate which allocations or sleeping are allowed.
 577  *
 578  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 579  * the given range from the tree regardless of state (ie for truncate).
 580  *
 581  * the range [start, end] is inclusive.
 582  *
 583  * This takes the tree lock, and returns 0 on success and < 0 on error.
 584  */
 585 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 586                               unsigned bits, int wake, int delete,
 587                               struct extent_state **cached_state,
 588                               gfp_t mask, struct extent_changeset *changeset)
 589 {
 590         struct extent_state *state;
 591         struct extent_state *cached;
 592         struct extent_state *prealloc = NULL;
 593         struct rb_node *node;
 594         u64 last_end;
 595         int err;
 596         int clear = 0;
 597
 598         btrfs_debug_check_extent_io_range(tree, start, end);
 599
 600         if (bits & EXTENT_DELALLOC)
 601                 bits |= EXTENT_NORESERVE;
 602
 603         if (delete)
 604                 bits |= ~EXTENT_CTLBITS;
 605         bits |= EXTENT_FIRST_DELALLOC;
 606
 607         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 608                 clear = 1;
 609 again:
 610         if (!prealloc && gfpflags_allow_blocking(mask)) {
 611                 /*
 612                  * Don't care for allocation failure here because we might end
 613                  * up not needing the pre-allocated extent state at all, which
 614                  * is the case if we only have in the tree extent states that
 615                  * cover our input range and don't cover too any other range.
 616                  * If we end up needing a new extent state we allocate it later.
 617                  */
 618                 prealloc = alloc_extent_state(mask);
 619         }
 620
 621         spin_lock(&tree->lock);
 622         if (cached_state) {
 623                 cached = *cached_state;
 624
 625                 if (clear) {
 626                         *cached_state = NULL;
 627                         cached_state = NULL;
 628                 }
 629
 630                 if (cached && extent_state_in_tree(cached) &&
 631                     cached->start <= start && cached->end > start) {
 632                         if (clear)
 633                                 refcount_dec(&cached->refs);
 634                         state = cached;
 635                         goto hit_next;
 636                 }
 637                 if (clear)
 638                         free_extent_state(cached);
 639         }
 640         /*
 641          * this search will find the extents that end after
 642          * our range starts
 643          */
 644         node = tree_search(tree, start);
 645         if (!node)
 646                 goto out;
 647         state = rb_entry(node, struct extent_state, rb_node);
 648 hit_next:
 649         if (state->start > end)
 650                 goto out;
 651         WARN_ON(state->end < start);
 652         last_end = state->end;
 653
 654         /* the state doesn't have the wanted bits, go ahead */
 655         if (!(state->state & bits)) {
 656                 state = next_state(state);
 657                 goto next;
 658         }
 659
 660         /*
 661          *     | ---- desired range ---- |
 662          *  | state | or
 663          *  | ------------- state -------------- |
 664          *
 665          * We need to split the extent we found, and may flip
 666          * bits on second half.
 667          *
 668          * If the extent we found extends past our range, we
 669          * just split and search again.  It'll get split again
 670          * the next time though.
 671          *
 672          * If the extent we found is inside our range, we clear
 673          * the desired bit on it.
 674          */
 675
 676         if (state->start < start) {
 677                 prealloc = alloc_extent_state_atomic(prealloc);
 678                 BUG_ON(!prealloc);
 679                 err = split_state(tree, state, prealloc, start);
 680                 if (err)
 681                         extent_io_tree_panic(tree, err);
 682
 683                 prealloc = NULL;
 684                 if (err)
 685                         goto out;
 686                 if (state->end <= end) {
 687                         state = clear_state_bit(tree, state, &bits, wake,
 688                                                 changeset);
 689                         goto next;
 690                 }
 691                 goto search_again;
 692         }
 693         /*
 694          * | ---- desired range ---- |
 695          *                        | state |
 696          * We need to split the extent, and clear the bit
 697          * on the first half
 698          */
 699         if (state->start <= end && state->end > end) {
 700                 prealloc = alloc_extent_state_atomic(prealloc);
 701                 BUG_ON(!prealloc);
 702                 err = split_state(tree, state, prealloc, end + 1);
 703                 if (err)
 704                         extent_io_tree_panic(tree, err);
 705
 706                 if (wake)
 707                         wake_up(&state->wq);
 708
 709                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 710
 711                 prealloc = NULL;
 712                 goto out;
 713         }
 714
 715         state = clear_state_bit(tree, state, &bits, wake, changeset);
 716 next:
 717         if (last_end == (u64)-1)
 718                 goto out;
 719         start = last_end + 1;
 720         if (start <= end && state && !need_resched())
 721                 goto hit_next;
 722
 723 search_again:
 724         if (start > end)
 725                 goto out;
 726         spin_unlock(&tree->lock);
 727         if (gfpflags_allow_blocking(mask))
 728                 cond_resched();
 729         goto again;
 730
 731 out:
 732         spin_unlock(&tree->lock);
 733         if (prealloc)
 734                 free_extent_state(prealloc);
 735
 736         return 0;
 737
 738 }
 739
 740 static void wait_on_state(struct extent_io_tree *tree,
 741                           struct extent_state *state)
 742                 __releases(tree->lock)
 743                 __acquires(tree->lock)
 744 {
 745         DEFINE_WAIT(wait);
 746         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 747         spin_unlock(&tree->lock);
 748         schedule();
 749         spin_lock(&tree->lock);
 750         finish_wait(&state->wq, &wait);
 751 }
 752
 753 /*
 754  * waits for one or more bits to clear on a range in the state tree.
 755  * The range [start, end] is inclusive.
 756  * The tree lock is taken by this function
 757  */
 758 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 759                             unsigned long bits)
 760 {
 761         struct extent_state *state;
 762         struct rb_node *node;
 763
 764         btrfs_debug_check_extent_io_range(tree, start, end);
 765
 766         spin_lock(&tree->lock);
 767 again:
 768         while (1) {
 769                 /*
 770                  * this search will find all the extents that end after
 771                  * our range starts
 772                  */
 773                 node = tree_search(tree, start);
 774 process_node:
 775                 if (!node)
 776                         break;
 777
 778                 state = rb_entry(node, struct extent_state, rb_node);
 779
 780                 if (state->start > end)
 781                         goto out;
 782
 783                 if (state->state & bits) {
 784                         start = state->start;
 785                         refcount_inc(&state->refs);
 786                         wait_on_state(tree, state);
 787                         free_extent_state(state);
 788                         goto again;
 789                 }
 790                 start = state->end + 1;
 791
 792                 if (start > end)
 793                         break;
 794
 795                 if (!cond_resched_lock(&tree->lock)) {
 796                         node = rb_next(node);
 797                         goto process_node;
 798                 }
 799         }
 800 out:
 801         spin_unlock(&tree->lock);
 802 }
 803
 804 static void set_state_bits(struct extent_io_tree *tree,
 805                            struct extent_state *state,
 806                            unsigned *bits, struct extent_changeset *changeset)
 807 {
 808         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 809
 810         set_state_cb(tree, state, bits);
 811         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 812                 u64 range = state->end - state->start + 1;
 813                 tree->dirty_bytes += range;
 814         }
 815         add_extent_changeset(state, bits_to_set, changeset, 1);
 816         state->state |= bits_to_set;
 817 }
 818
 819 static void cache_state_if_flags(struct extent_state *state,
 820                                  struct extent_state **cached_ptr,
 821                                  unsigned flags)
 822 {
 823         if (cached_ptr && !(*cached_ptr)) {
 824                 if (!flags || (state->state & flags)) {
 825                         *cached_ptr = state;
 826                         refcount_inc(&state->refs);
 827                 }
 828         }
 829 }
 830
 831 static void cache_state(struct extent_state *state,
 832                         struct extent_state **cached_ptr)
 833 {
 834         return cache_state_if_flags(state, cached_ptr,
 835                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 836 }
 837
 838 /*
 839  * set some bits on a range in the tree.  This may require allocations or
 840  * sleeping, so the gfp mask is used to indicate what is allowed.
 841  *
 842  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 843  * part of the range already has the desired bits set.  The start of the
 844  * existing range is returned in failed_start in this case.
 845  *
 846  * [start, end] is inclusive This takes the tree lock.
 847  */
 848
 849 static int __must_check
 850 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 851                  unsigned bits, unsigned exclusive_bits,
 852                  u64 *failed_start, struct extent_state **cached_state,
 853                  gfp_t mask, struct extent_changeset *changeset)
 854 {
 855         struct extent_state *state;
 856         struct extent_state *prealloc = NULL;
 857         struct rb_node *node;
 858         struct rb_node **p;
 859         struct rb_node *parent;
 860         int err = 0;
 861         u64 last_start;
 862         u64 last_end;
 863
 864         btrfs_debug_check_extent_io_range(tree, start, end);
 865
 866         bits |= EXTENT_FIRST_DELALLOC;
 867 again:
 868         if (!prealloc && gfpflags_allow_blocking(mask)) {
 869                 /*
 870                  * Don't care for allocation failure here because we might end
 871                  * up not needing the pre-allocated extent state at all, which
 872                  * is the case if we only have in the tree extent states that
 873                  * cover our input range and don't cover too any other range.
 874                  * If we end up needing a new extent state we allocate it later.
 875                  */
 876                 prealloc = alloc_extent_state(mask);
 877         }
 878
 879         spin_lock(&tree->lock);
 880         if (cached_state && *cached_state) {
 881                 state = *cached_state;
 882                 if (state->start <= start && state->end > start &&
 883                     extent_state_in_tree(state)) {
 884                         node = &state->rb_node;
 885                         goto hit_next;
 886                 }
 887         }
 888         /*
 889          * this search will find all the extents that end after
 890          * our range starts.
 891          */
 892         node = tree_search_for_insert(tree, start, &p, &parent);
 893         if (!node) {
 894                 prealloc = alloc_extent_state_atomic(prealloc);
 895                 BUG_ON(!prealloc);
 896                 err = insert_state(tree, prealloc, start, end,
 897                                    &p, &parent, &bits, changeset);
 898                 if (err)
 899                         extent_io_tree_panic(tree, err);
 900
 901                 cache_state(prealloc, cached_state);
 902                 prealloc = NULL;
 903                 goto out;
 904         }
 905         state = rb_entry(node, struct extent_state, rb_node);
 906 hit_next:
 907         last_start = state->start;
 908         last_end = state->end;
 909
 910         /*
 911          * | ---- desired range ---- |
 912          * | state |
 913          *
 914          * Just lock what we found and keep going
 915          */
 916         if (state->start == start && state->end <= end) {
 917                 if (state->state & exclusive_bits) {
 918                         *failed_start = state->start;
 919                         err = -EEXIST;
 920                         goto out;
 921                 }
 922
 923                 set_state_bits(tree, state, &bits, changeset);
 924                 cache_state(state, cached_state);
 925                 merge_state(tree, state);
 926                 if (last_end == (u64)-1)
 927                         goto out;
 928                 start = last_end + 1;
 929                 state = next_state(state);
 930                 if (start < end && state && state->start == start &&
 931                     !need_resched())
 932                         goto hit_next;
 933                 goto search_again;
 934         }
 935
 936         /*
 937          *     | ---- desired range ---- |
 938          * | state |
 939          *   or
 940          * | ------------- state -------------- |
 941          *
 942          * We need to split the extent we found, and may flip bits on
 943          * second half.
 944          *
 945          * If the extent we found extends past our
 946          * range, we just split and search again.  It'll get split
 947          * again the next time though.
 948          *
 949          * If the extent we found is inside our range, we set the
 950          * desired bit on it.
 951          */
 952         if (state->start < start) {
 953                 if (state->state & exclusive_bits) {
 954                         *failed_start = start;
 955                         err = -EEXIST;
 956                         goto out;
 957                 }
 958
 959                 prealloc = alloc_extent_state_atomic(prealloc);
 960                 BUG_ON(!prealloc);
 961                 err = split_state(tree, state, prealloc, start);
 962                 if (err)
 963                         extent_io_tree_panic(tree, err);
 964
 965                 prealloc = NULL;
 966                 if (err)
 967                         goto out;
 968                 if (state->end <= end) {
 969                         set_state_bits(tree, state, &bits, changeset);
 970                         cache_state(state, cached_state);
 971                         merge_state(tree, state);
 972                         if (last_end == (u64)-1)
 973                                 goto out;
 974                         start = last_end + 1;
 975                         state = next_state(state);
 976                         if (start < end && state && state->start == start &&
 977                             !need_resched())
 978                                 goto hit_next;
 979                 }
 980                 goto search_again;
 981         }
 982         /*
 983          * | ---- desired range ---- |
 984          *     | state | or               | state |
 985          *
 986          * There's a hole, we need to insert something in it and
 987          * ignore the extent we found.
 988          */
 989         if (state->start > start) {
 990                 u64 this_end;
 991                 if (end < last_start)
 992                         this_end = end;
 993                 else
 994                         this_end = last_start - 1;
 995
 996                 prealloc = alloc_extent_state_atomic(prealloc);
 997                 BUG_ON(!prealloc);
 998
 999                 /*
1000                  * Avoid to free 'prealloc' if it can be merged with
1001                  * the later extent.
1002                  */
1003                 err = insert_state(tree, prealloc, start, this_end,
1004                                    NULL, NULL, &bits, changeset);
1005                 if (err)
1006                         extent_io_tree_panic(tree, err);
1007
1008                 cache_state(prealloc, cached_state);
1009                 prealloc = NULL;
1010                 start = this_end + 1;
1011                 goto search_again;
1012         }
1013         /*
1014          * | ---- desired range ---- |
1015          *                        | state |
1016          * We need to split the extent, and set the bit
1017          * on the first half
1018          */
1019         if (state->start <= end && state->end > end) {
1020                 if (state->state & exclusive_bits) {
1021                         *failed_start = start;
1022                         err = -EEXIST;
1023                         goto out;
1024                 }
1025
1026                 prealloc = alloc_extent_state_atomic(prealloc);
1027                 BUG_ON(!prealloc);
1028                 err = split_state(tree, state, prealloc, end + 1);
1029                 if (err)
1030                         extent_io_tree_panic(tree, err);
1031
1032                 set_state_bits(tree, prealloc, &bits, changeset);
1033                 cache_state(prealloc, cached_state);
1034                 merge_state(tree, prealloc);
1035                 prealloc = NULL;
1036                 goto out;
1037         }
1038
1039 search_again:
1040         if (start > end)
1041                 goto out;
1042         spin_unlock(&tree->lock);
1043         if (gfpflags_allow_blocking(mask))
1044                 cond_resched();
1045         goto again;
1046
1047 out:
1048         spin_unlock(&tree->lock);
1049         if (prealloc)
1050                 free_extent_state(prealloc);
1051
1052         return err;
1053
1054 }
1055
1056 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1057                    unsigned bits, u64 * failed_start,
1058                    struct extent_state **cached_state, gfp_t mask)
1059 {
1060         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1061                                 cached_state, mask, NULL);
1062 }
1063
1064
1065 /**
1066  * convert_extent_bit - convert all bits in a given range from one bit to
1067  *                      another
1068  * @tree:       the io tree to search
1069  * @start:      the start offset in bytes
1070  * @end:        the end offset in bytes (inclusive)
1071  * @bits:       the bits to set in this range
1072  * @clear_bits: the bits to clear in this range
1073  * @cached_state:       state that we're going to cache
1074  *
1075  * This will go through and set bits for the given range.  If any states exist
1076  * already in this range they are set with the given bit and cleared of the
1077  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1078  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1079  * boundary bits like LOCK.
1080  *
1081  * All allocations are done with GFP_NOFS.
1082  */
1083 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1084                        unsigned bits, unsigned clear_bits,
1085                        struct extent_state **cached_state)
1086 {
1087         struct extent_state *state;
1088         struct extent_state *prealloc = NULL;
1089         struct rb_node *node;
1090         struct rb_node **p;
1091         struct rb_node *parent;
1092         int err = 0;
1093         u64 last_start;
1094         u64 last_end;
1095         bool first_iteration = true;
1096
1097         btrfs_debug_check_extent_io_range(tree, start, end);
1098
1099 again:
1100         if (!prealloc) {
1101                 /*
1102                  * Best effort, don't worry if extent state allocation fails
1103                  * here for the first iteration. We might have a cached state
1104                  * that matches exactly the target range, in which case no
1105                  * extent state allocations are needed. We'll only know this
1106                  * after locking the tree.
1107                  */
1108                 prealloc = alloc_extent_state(GFP_NOFS);
1109                 if (!prealloc && !first_iteration)
1110                         return -ENOMEM;
1111         }
1112
1113         spin_lock(&tree->lock);
1114         if (cached_state && *cached_state) {
1115                 state = *cached_state;
1116                 if (state->start <= start && state->end > start &&
1117                     extent_state_in_tree(state)) {
1118                         node = &state->rb_node;
1119                         goto hit_next;
1120                 }
1121         }
1122
1123         /*
1124          * this search will find all the extents that end after
1125          * our range starts.
1126          */
1127         node = tree_search_for_insert(tree, start, &p, &parent);
1128         if (!node) {
1129                 prealloc = alloc_extent_state_atomic(prealloc);
1130                 if (!prealloc) {
1131                         err = -ENOMEM;
1132                         goto out;
1133                 }
1134                 err = insert_state(tree, prealloc, start, end,
1135                                    &p, &parent, &bits, NULL);
1136                 if (err)
1137                         extent_io_tree_panic(tree, err);
1138                 cache_state(prealloc, cached_state);
1139                 prealloc = NULL;
1140                 goto out;
1141         }
1142         state = rb_entry(node, struct extent_state, rb_node);
1143 hit_next:
1144         last_start = state->start;
1145         last_end = state->end;
1146
1147         /*
1148          * | ---- desired range ---- |
1149          * | state |
1150          *
1151          * Just lock what we found and keep going
1152          */
1153         if (state->start == start && state->end <= end) {
1154                 set_state_bits(tree, state, &bits, NULL);
1155                 cache_state(state, cached_state);
1156                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1157                 if (last_end == (u64)-1)
1158                         goto out;
1159                 start = last_end + 1;
1160                 if (start < end && state && state->start == start &&
1161                     !need_resched())
1162                         goto hit_next;
1163                 goto search_again;
1164         }
1165
1166         /*
1167          *     | ---- desired range ---- |
1168          * | state |
1169          *   or
1170          * | ------------- state -------------- |
1171          *
1172          * We need to split the extent we found, and may flip bits on
1173          * second half.
1174          *
1175          * If the extent we found extends past our
1176          * range, we just split and search again.  It'll get split
1177          * again the next time though.
1178          *
1179          * If the extent we found is inside our range, we set the
1180          * desired bit on it.
1181          */
1182         if (state->start < start) {
1183                 prealloc = alloc_extent_state_atomic(prealloc);
1184                 if (!prealloc) {
1185                         err = -ENOMEM;
1186                         goto out;
1187                 }
1188                 err = split_state(tree, state, prealloc, start);
1189                 if (err)
1190                         extent_io_tree_panic(tree, err);
1191                 prealloc = NULL;
1192                 if (err)
1193                         goto out;
1194                 if (state->end <= end) {
1195                         set_state_bits(tree, state, &bits, NULL);
1196                         cache_state(state, cached_state);
1197                         state = clear_state_bit(tree, state, &clear_bits, 0,
1198                                                 NULL);
1199                         if (last_end == (u64)-1)
1200                                 goto out;
1201                         start = last_end + 1;
1202                         if (start < end && state && state->start == start &&
1203                             !need_resched())
1204                                 goto hit_next;
1205                 }
1206                 goto search_again;
1207         }
1208         /*
1209          * | ---- desired range ---- |
1210          *     | state | or               | state |
1211          *
1212          * There's a hole, we need to insert something in it and
1213          * ignore the extent we found.
1214          */
1215         if (state->start > start) {
1216                 u64 this_end;
1217                 if (end < last_start)
1218                         this_end = end;
1219                 else
1220                         this_end = last_start - 1;
1221
1222                 prealloc = alloc_extent_state_atomic(prealloc);
1223                 if (!prealloc) {
1224                         err = -ENOMEM;
1225                         goto out;
1226                 }
1227
1228                 /*
1229                  * Avoid to free 'prealloc' if it can be merged with
1230                  * the later extent.
1231                  */
1232                 err = insert_state(tree, prealloc, start, this_end,
1233                                    NULL, NULL, &bits, NULL);
1234                 if (err)
1235                         extent_io_tree_panic(tree, err);
1236                 cache_state(prealloc, cached_state);
1237                 prealloc = NULL;
1238                 start = this_end + 1;
1239                 goto search_again;
1240         }
1241         /*
1242          * | ---- desired range ---- |
1243          *                        | state |
1244          * We need to split the extent, and set the bit
1245          * on the first half
1246          */
1247         if (state->start <= end && state->end > end) {
1248                 prealloc = alloc_extent_state_atomic(prealloc);
1249                 if (!prealloc) {
1250                         err = -ENOMEM;
1251                         goto out;
1252                 }
1253
1254                 err = split_state(tree, state, prealloc, end + 1);
1255                 if (err)
1256                         extent_io_tree_panic(tree, err);
1257
1258                 set_state_bits(tree, prealloc, &bits, NULL);
1259                 cache_state(prealloc, cached_state);
1260                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1261                 prealloc = NULL;
1262                 goto out;
1263         }
1264
1265 search_again:
1266         if (start > end)
1267                 goto out;
1268         spin_unlock(&tree->lock);
1269         cond_resched();
1270         first_iteration = false;
1271         goto again;
1272
1273 out:
1274         spin_unlock(&tree->lock);
1275         if (prealloc)
1276                 free_extent_state(prealloc);
1277
1278         return err;
1279 }
1280
1281 /* wrappers around set/clear extent bit */
1282 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1283                            unsigned bits, struct extent_changeset *changeset)
1284 {
1285         /*
1286          * We don't support EXTENT_LOCKED yet, as current changeset will
1287          * record any bits changed, so for EXTENT_LOCKED case, it will
1288          * either fail with -EEXIST or changeset will record the whole
1289          * range.
1290          */
1291         BUG_ON(bits & EXTENT_LOCKED);
1292
1293         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1294                                 changeset);
1295 }
1296
1297 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1298                      unsigned bits, int wake, int delete,
1299                      struct extent_state **cached, gfp_t mask)
1300 {
1301         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1302                                   cached, mask, NULL);
1303 }
1304
1305 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1306                 unsigned bits, struct extent_changeset *changeset)
1307 {
1308         /*
1309          * Don't support EXTENT_LOCKED case, same reason as
1310          * set_record_extent_bits().
1311          */
1312         BUG_ON(bits & EXTENT_LOCKED);
1313
1314         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1315                                   changeset);
1316 }
1317
1318 /*
1319  * either insert or lock state struct between start and end use mask to tell
1320  * us if waiting is desired.
1321  */
1322 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1323                      struct extent_state **cached_state)
1324 {
1325         int err;
1326         u64 failed_start;
1327
1328         while (1) {
1329                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1330                                        EXTENT_LOCKED, &failed_start,
1331                                        cached_state, GFP_NOFS, NULL);
1332                 if (err == -EEXIST) {
1333                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1334                         start = failed_start;
1335                 } else
1336                         break;
1337                 WARN_ON(start > end);
1338         }
1339         return err;
1340 }
1341
1342 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1343 {
1344         int err;
1345         u64 failed_start;
1346
1347         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1348                                &failed_start, NULL, GFP_NOFS, NULL);
1349         if (err == -EEXIST) {
1350                 if (failed_start > start)
1351                         clear_extent_bit(tree, start, failed_start - 1,
1352                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1353                 return 0;
1354         }
1355         return 1;
1356 }
1357
1358 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1359 {
1360         unsigned long index = start >> PAGE_SHIFT;
1361         unsigned long end_index = end >> PAGE_SHIFT;
1362         struct page *page;
1363
1364         while (index <= end_index) {
1365                 page = find_get_page(inode->i_mapping, index);
1366                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1367                 clear_page_dirty_for_io(page);
1368                 put_page(page);
1369                 index++;
1370         }
1371 }
1372
1373 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1374 {
1375         unsigned long index = start >> PAGE_SHIFT;
1376         unsigned long end_index = end >> PAGE_SHIFT;
1377         struct page *page;
1378
1379         while (index <= end_index) {
1380                 page = find_get_page(inode->i_mapping, index);
1381                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1382                 __set_page_dirty_nobuffers(page);
1383                 account_page_redirty(page);
1384                 put_page(page);
1385                 index++;
1386         }
1387 }
1388
1389 /*
1390  * helper function to set both pages and extents in the tree writeback
1391  */
1392 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1393 {
1394         tree->ops->set_range_writeback(tree->private_data, start, end);
1395 }
1396
1397 /* find the first state struct with 'bits' set after 'start', and
1398  * return it.  tree->lock must be held.  NULL will returned if
1399  * nothing was found after 'start'
1400  */
1401 static struct extent_state *
1402 find_first_extent_bit_state(struct extent_io_tree *tree,
1403                             u64 start, unsigned bits)
1404 {
1405         struct rb_node *node;
1406         struct extent_state *state;
1407
1408         /*
1409          * this search will find all the extents that end after
1410          * our range starts.
1411          */
1412         node = tree_search(tree, start);
1413         if (!node)
1414                 goto out;
1415
1416         while (1) {
1417                 state = rb_entry(node, struct extent_state, rb_node);
1418                 if (state->end >= start && (state->state & bits))
1419                         return state;
1420
1421                 node = rb_next(node);
1422                 if (!node)
1423                         break;
1424         }
1425 out:
1426         return NULL;
1427 }
1428
1429 /*
1430  * find the first offset in the io tree with 'bits' set. zero is
1431  * returned if we find something, and *start_ret and *end_ret are
1432  * set to reflect the state struct that was found.
1433  *
1434  * If nothing was found, 1 is returned. If found something, return 0.
1435  */
1436 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1437                           u64 *start_ret, u64 *end_ret, unsigned bits,
1438                           struct extent_state **cached_state)
1439 {
1440         struct extent_state *state;
1441         struct rb_node *n;
1442         int ret = 1;
1443
1444         spin_lock(&tree->lock);
1445         if (cached_state && *cached_state) {
1446                 state = *cached_state;
1447                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1448                         n = rb_next(&state->rb_node);
1449                         while (n) {
1450                                 state = rb_entry(n, struct extent_state,
1451                                                  rb_node);
1452                                 if (state->state & bits)
1453                                         goto got_it;
1454                                 n = rb_next(n);
1455                         }
1456                         free_extent_state(*cached_state);
1457                         *cached_state = NULL;
1458                         goto out;
1459                 }
1460                 free_extent_state(*cached_state);
1461                 *cached_state = NULL;
1462         }
1463
1464         state = find_first_extent_bit_state(tree, start, bits);
1465 got_it:
1466         if (state) {
1467                 cache_state_if_flags(state, cached_state, 0);
1468                 *start_ret = state->start;
1469                 *end_ret = state->end;
1470                 ret = 0;
1471         }
1472 out:
1473         spin_unlock(&tree->lock);
1474         return ret;
1475 }
1476
1477 /*
1478  * find a contiguous range of bytes in the file marked as delalloc, not
1479  * more than 'max_bytes'.  start and end are used to return the range,
1480  *
1481  * 1 is returned if we find something, 0 if nothing was in the tree
1482  */
1483 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1484                                         u64 *start, u64 *end, u64 max_bytes,
1485                                         struct extent_state **cached_state)
1486 {
1487         struct rb_node *node;
1488         struct extent_state *state;
1489         u64 cur_start = *start;
1490         u64 found = 0;
1491         u64 total_bytes = 0;
1492
1493         spin_lock(&tree->lock);
1494
1495         /*
1496          * this search will find all the extents that end after
1497          * our range starts.
1498          */
1499         node = tree_search(tree, cur_start);
1500         if (!node) {
1501                 if (!found)
1502                         *end = (u64)-1;
1503                 goto out;
1504         }
1505
1506         while (1) {
1507                 state = rb_entry(node, struct extent_state, rb_node);
1508                 if (found && (state->start != cur_start ||
1509                               (state->state & EXTENT_BOUNDARY))) {
1510                         goto out;
1511                 }
1512                 if (!(state->state & EXTENT_DELALLOC)) {
1513                         if (!found)
1514                                 *end = state->end;
1515                         goto out;
1516                 }
1517                 if (!found) {
1518                         *start = state->start;
1519                         *cached_state = state;
1520                         refcount_inc(&state->refs);
1521                 }
1522                 found++;
1523                 *end = state->end;
1524                 cur_start = state->end + 1;
1525                 node = rb_next(node);
1526                 total_bytes += state->end - state->start + 1;
1527                 if (total_bytes >= max_bytes)
1528                         break;
1529                 if (!node)
1530                         break;
1531         }
1532 out:
1533         spin_unlock(&tree->lock);
1534         return found;
1535 }
1536
1537 static int __process_pages_contig(struct address_space *mapping,
1538                                   struct page *locked_page,
1539                                   pgoff_t start_index, pgoff_t end_index,
1540                                   unsigned long page_ops, pgoff_t *index_ret);
1541
1542 static noinline void __unlock_for_delalloc(struct inode *inode,
1543                                            struct page *locked_page,
1544                                            u64 start, u64 end)
1545 {
1546         unsigned long index = start >> PAGE_SHIFT;
1547         unsigned long end_index = end >> PAGE_SHIFT;
1548
1549         ASSERT(locked_page);
1550         if (index == locked_page->index && end_index == index)
1551                 return;
1552
1553         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1554                                PAGE_UNLOCK, NULL);
1555 }
1556
1557 static noinline int lock_delalloc_pages(struct inode *inode,
1558                                         struct page *locked_page,
1559                                         u64 delalloc_start,
1560                                         u64 delalloc_end)
1561 {
1562         unsigned long index = delalloc_start >> PAGE_SHIFT;
1563         unsigned long index_ret = index;
1564         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1565         int ret;
1566
1567         ASSERT(locked_page);
1568         if (index == locked_page->index && index == end_index)
1569                 return 0;
1570
1571         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1572                                      end_index, PAGE_LOCK, &index_ret);
1573         if (ret == -EAGAIN)
1574                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1575                                       (u64)index_ret << PAGE_SHIFT);
1576         return ret;
1577 }
1578
1579 /*
1580  * find a contiguous range of bytes in the file marked as delalloc, not
1581  * more than 'max_bytes'.  start and end are used to return the range,
1582  *
1583  * 1 is returned if we find something, 0 if nothing was in the tree
1584  */
1585 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1586                                     struct extent_io_tree *tree,
1587                                     struct page *locked_page, u64 *start,
1588                                     u64 *end, u64 max_bytes)
1589 {
1590         u64 delalloc_start;
1591         u64 delalloc_end;
1592         u64 found;
1593         struct extent_state *cached_state = NULL;
1594         int ret;
1595         int loops = 0;
1596
1597 again:
1598         /* step one, find a bunch of delalloc bytes starting at start */
1599         delalloc_start = *start;
1600         delalloc_end = 0;
1601         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1602                                     max_bytes, &cached_state);
1603         if (!found || delalloc_end <= *start) {
1604                 *start = delalloc_start;
1605                 *end = delalloc_end;
1606                 free_extent_state(cached_state);
1607                 return 0;
1608         }
1609
1610         /*
1611          * start comes from the offset of locked_page.  We have to lock
1612          * pages in order, so we can't process delalloc bytes before
1613          * locked_page
1614          */
1615         if (delalloc_start < *start)
1616                 delalloc_start = *start;
1617
1618         /*
1619          * make sure to limit the number of pages we try to lock down
1620          */
1621         if (delalloc_end + 1 - delalloc_start > max_bytes)
1622                 delalloc_end = delalloc_start + max_bytes - 1;
1623
1624         /* step two, lock all the pages after the page that has start */
1625         ret = lock_delalloc_pages(inode, locked_page,
1626                                   delalloc_start, delalloc_end);
1627         if (ret == -EAGAIN) {
1628                 /* some of the pages are gone, lets avoid looping by
1629                  * shortening the size of the delalloc range we're searching
1630                  */
1631                 free_extent_state(cached_state);
1632                 cached_state = NULL;
1633                 if (!loops) {
1634                         max_bytes = PAGE_SIZE;
1635                         loops = 1;
1636                         goto again;
1637                 } else {
1638                         found = 0;
1639                         goto out_failed;
1640                 }
1641         }
1642         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1643
1644         /* step three, lock the state bits for the whole range */
1645         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1646
1647         /* then test to make sure it is all still delalloc */
1648         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1649                              EXTENT_DELALLOC, 1, cached_state);
1650         if (!ret) {
1651                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1652                                      &cached_state, GFP_NOFS);
1653                 __unlock_for_delalloc(inode, locked_page,
1654                               delalloc_start, delalloc_end);
1655                 cond_resched();
1656                 goto again;
1657         }
1658         free_extent_state(cached_state);
1659         *start = delalloc_start;
1660         *end = delalloc_end;
1661 out_failed:
1662         return found;
1663 }
1664
1665 static int __process_pages_contig(struct address_space *mapping,
1666                                   struct page *locked_page,
1667                                   pgoff_t start_index, pgoff_t end_index,
1668                                   unsigned long page_ops, pgoff_t *index_ret)
1669 {
1670         unsigned long nr_pages = end_index - start_index + 1;
1671         unsigned long pages_locked = 0;
1672         pgoff_t index = start_index;
1673         struct page *pages[16];
1674         unsigned ret;
1675         int err = 0;
1676         int i;
1677
1678         if (page_ops & PAGE_LOCK) {
1679                 ASSERT(page_ops == PAGE_LOCK);
1680                 ASSERT(index_ret && *index_ret == start_index);
1681         }
1682
1683         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1684                 mapping_set_error(mapping, -EIO);
1685
1686         while (nr_pages > 0) {
1687                 ret = find_get_pages_contig(mapping, index,
1688                                      min_t(unsigned long,
1689                                      nr_pages, ARRAY_SIZE(pages)), pages);
1690                 if (ret == 0) {
1691                         /*
1692                          * Only if we're going to lock these pages,
1693                          * can we find nothing at @index.
1694                          */
1695                         ASSERT(page_ops & PAGE_LOCK);
1696                         err = -EAGAIN;
1697                         goto out;
1698                 }
1699
1700                 for (i = 0; i < ret; i++) {
1701                         if (page_ops & PAGE_SET_PRIVATE2)
1702                                 SetPagePrivate2(pages[i]);
1703
1704                         if (pages[i] == locked_page) {
1705                                 put_page(pages[i]);
1706                                 pages_locked++;
1707                                 continue;
1708                         }
1709                         if (page_ops & PAGE_CLEAR_DIRTY)
1710                                 clear_page_dirty_for_io(pages[i]);
1711                         if (page_ops & PAGE_SET_WRITEBACK)
1712                                 set_page_writeback(pages[i]);
1713                         if (page_ops & PAGE_SET_ERROR)
1714                                 SetPageError(pages[i]);
1715                         if (page_ops & PAGE_END_WRITEBACK)
1716                                 end_page_writeback(pages[i]);
1717                         if (page_ops & PAGE_UNLOCK)
1718                                 unlock_page(pages[i]);
1719                         if (page_ops & PAGE_LOCK) {
1720                                 lock_page(pages[i]);
1721                                 if (!PageDirty(pages[i]) ||
1722                                     pages[i]->mapping != mapping) {
1723                                         unlock_page(pages[i]);
1724                                         for (; i < ret; i++)
1725                                                 put_page(pages[i]);
1726                                         err = -EAGAIN;
1727                                         goto out;
1728                                 }
1729                         }
1730                         put_page(pages[i]);
1731                         pages_locked++;
1732                 }
1733                 nr_pages -= ret;
1734                 index += ret;
1735                 cond_resched();
1736         }
1737 out:
1738         if (err && index_ret)
1739                 *index_ret = start_index + pages_locked - 1;
1740         return err;
1741 }
1742
1743 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1744                                  u64 delalloc_end, struct page *locked_page,
1745                                  unsigned clear_bits,
1746                                  unsigned long page_ops)
1747 {
1748         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1749                          NULL, GFP_NOFS);
1750
1751         __process_pages_contig(inode->i_mapping, locked_page,
1752                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1753                                page_ops, NULL);
1754 }
1755
1756 /*
1757  * count the number of bytes in the tree that have a given bit(s)
1758  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1759  * cached.  The total number found is returned.
1760  */
1761 u64 count_range_bits(struct extent_io_tree *tree,
1762                      u64 *start, u64 search_end, u64 max_bytes,
1763                      unsigned bits, int contig)
1764 {
1765         struct rb_node *node;
1766         struct extent_state *state;
1767         u64 cur_start = *start;
1768         u64 total_bytes = 0;
1769         u64 last = 0;
1770         int found = 0;
1771
1772         if (WARN_ON(search_end <= cur_start))
1773                 return 0;
1774
1775         spin_lock(&tree->lock);
1776         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1777                 total_bytes = tree->dirty_bytes;
1778                 goto out;
1779         }
1780         /*
1781          * this search will find all the extents that end after
1782          * our range starts.
1783          */
1784         node = tree_search(tree, cur_start);
1785         if (!node)
1786                 goto out;
1787
1788         while (1) {
1789                 state = rb_entry(node, struct extent_state, rb_node);
1790                 if (state->start > search_end)
1791                         break;
1792                 if (contig && found && state->start > last + 1)
1793                         break;
1794                 if (state->end >= cur_start && (state->state & bits) == bits) {
1795                         total_bytes += min(search_end, state->end) + 1 -
1796                                        max(cur_start, state->start);
1797                         if (total_bytes >= max_bytes)
1798                                 break;
1799                         if (!found) {
1800                                 *start = max(cur_start, state->start);
1801                                 found = 1;
1802                         }
1803                         last = state->end;
1804                 } else if (contig && found) {
1805                         break;
1806                 }
1807                 node = rb_next(node);
1808                 if (!node)
1809                         break;
1810         }
1811 out:
1812         spin_unlock(&tree->lock);
1813         return total_bytes;
1814 }
1815
1816 /*
1817  * set the private field for a given byte offset in the tree.  If there isn't
1818  * an extent_state there already, this does nothing.
1819  */
1820 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1821                 struct io_failure_record *failrec)
1822 {
1823         struct rb_node *node;
1824         struct extent_state *state;
1825         int ret = 0;
1826
1827         spin_lock(&tree->lock);
1828         /*
1829          * this search will find all the extents that end after
1830          * our range starts.
1831          */
1832         node = tree_search(tree, start);
1833         if (!node) {
1834                 ret = -ENOENT;
1835                 goto out;
1836         }
1837         state = rb_entry(node, struct extent_state, rb_node);
1838         if (state->start != start) {
1839                 ret = -ENOENT;
1840                 goto out;
1841         }
1842         state->failrec = failrec;
1843 out:
1844         spin_unlock(&tree->lock);
1845         return ret;
1846 }
1847
1848 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1849                 struct io_failure_record **failrec)
1850 {
1851         struct rb_node *node;
1852         struct extent_state *state;
1853         int ret = 0;
1854
1855         spin_lock(&tree->lock);
1856         /*
1857          * this search will find all the extents that end after
1858          * our range starts.
1859          */
1860         node = tree_search(tree, start);
1861         if (!node) {
1862                 ret = -ENOENT;
1863                 goto out;
1864         }
1865         state = rb_entry(node, struct extent_state, rb_node);
1866         if (state->start != start) {
1867                 ret = -ENOENT;
1868                 goto out;
1869         }
1870         *failrec = state->failrec;
1871 out:
1872         spin_unlock(&tree->lock);
1873         return ret;
1874 }
1875
1876 /*
1877  * searches a range in the state tree for a given mask.
1878  * If 'filled' == 1, this returns 1 only if every extent in the tree
1879  * has the bits set.  Otherwise, 1 is returned if any bit in the
1880  * range is found set.
1881  */
1882 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1883                    unsigned bits, int filled, struct extent_state *cached)
1884 {
1885         struct extent_state *state = NULL;
1886         struct rb_node *node;
1887         int bitset = 0;
1888
1889         spin_lock(&tree->lock);
1890         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1891             cached->end > start)
1892                 node = &cached->rb_node;
1893         else
1894                 node = tree_search(tree, start);
1895         while (node && start <= end) {
1896                 state = rb_entry(node, struct extent_state, rb_node);
1897
1898                 if (filled && state->start > start) {
1899                         bitset = 0;
1900                         break;
1901                 }
1902
1903                 if (state->start > end)
1904                         break;
1905
1906                 if (state->state & bits) {
1907                         bitset = 1;
1908                         if (!filled)
1909                                 break;
1910                 } else if (filled) {
1911                         bitset = 0;
1912                         break;
1913                 }
1914
1915                 if (state->end == (u64)-1)
1916                         break;
1917
1918                 start = state->end + 1;
1919                 if (start > end)
1920                         break;
1921                 node = rb_next(node);
1922                 if (!node) {
1923                         if (filled)
1924                                 bitset = 0;
1925                         break;
1926                 }
1927         }
1928         spin_unlock(&tree->lock);
1929         return bitset;
1930 }
1931
1932 /*
1933  * helper function to set a given page up to date if all the
1934  * extents in the tree for that page are up to date
1935  */
1936 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1937 {
1938         u64 start = page_offset(page);
1939         u64 end = start + PAGE_SIZE - 1;
1940         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1941                 SetPageUptodate(page);
1942 }
1943
1944 int free_io_failure(struct extent_io_tree *failure_tree,
1945                     struct extent_io_tree *io_tree,
1946                     struct io_failure_record *rec)
1947 {
1948         int ret;
1949         int err = 0;
1950
1951         set_state_failrec(failure_tree, rec->start, NULL);
1952         ret = clear_extent_bits(failure_tree, rec->start,
1953                                 rec->start + rec->len - 1,
1954                                 EXTENT_LOCKED | EXTENT_DIRTY);
1955         if (ret)
1956                 err = ret;
1957
1958         ret = clear_extent_bits(io_tree, rec->start,
1959                                 rec->start + rec->len - 1,
1960                                 EXTENT_DAMAGED);
1961         if (ret && !err)
1962                 err = ret;
1963
1964         kfree(rec);
1965         return err;
1966 }
1967
1968 /*
1969  * this bypasses the standard btrfs submit functions deliberately, as
1970  * the standard behavior is to write all copies in a raid setup. here we only
1971  * want to write the one bad copy. so we do the mapping for ourselves and issue
1972  * submit_bio directly.
1973  * to avoid any synchronization issues, wait for the data after writing, which
1974  * actually prevents the read that triggered the error from finishing.
1975  * currently, there can be no more than two copies of every data bit. thus,
1976  * exactly one rewrite is required.
1977  */
1978 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
1979                       u64 length, u64 logical, struct page *page,
1980                       unsigned int pg_offset, int mirror_num)
1981 {
1982         struct bio *bio;
1983         struct btrfs_device *dev;
1984         u64 map_length = 0;
1985         u64 sector;
1986         struct btrfs_bio *bbio = NULL;
1987         int ret;
1988
1989         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
1990         BUG_ON(!mirror_num);
1991
1992         bio = btrfs_io_bio_alloc(1);
1993         bio->bi_iter.bi_size = 0;
1994         map_length = length;
1995
1996         /*
1997          * Avoid races with device replace and make sure our bbio has devices
1998          * associated to its stripes that don't go away while we are doing the
1999          * read repair operation.
2000          */
2001         btrfs_bio_counter_inc_blocked(fs_info);
2002         if (btrfs_is_parity_mirror(fs_info, logical, length)) {
2003                 /*
2004                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2005                  * to update all raid stripes, but here we just want to correct
2006                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2007                  * stripe's dev and sector.
2008                  */
2009                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2010                                       &map_length, &bbio, 0);
2011                 if (ret) {
2012                         btrfs_bio_counter_dec(fs_info);
2013                         bio_put(bio);
2014                         return -EIO;
2015                 }
2016                 ASSERT(bbio->mirror_num == 1);
2017         } else {
2018                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2019                                       &map_length, &bbio, mirror_num);
2020                 if (ret) {
2021                         btrfs_bio_counter_dec(fs_info);
2022                         bio_put(bio);
2023                         return -EIO;
2024                 }
2025                 BUG_ON(mirror_num != bbio->mirror_num);
2026         }
2027
2028         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2029         bio->bi_iter.bi_sector = sector;
2030         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2031         btrfs_put_bbio(bbio);
2032         if (!dev || !dev->bdev || !dev->writeable) {
2033                 btrfs_bio_counter_dec(fs_info);
2034                 bio_put(bio);
2035                 return -EIO;
2036         }
2037         bio_set_dev(bio, dev->bdev);
2038         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2039         bio_add_page(bio, page, length, pg_offset);
2040
2041         if (btrfsic_submit_bio_wait(bio)) {
2042                 /* try to remap that extent elsewhere? */
2043                 btrfs_bio_counter_dec(fs_info);
2044                 bio_put(bio);
2045                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2046                 return -EIO;
2047         }
2048
2049         btrfs_info_rl_in_rcu(fs_info,
2050                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2051                                   ino, start,
2052                                   rcu_str_deref(dev->name), sector);
2053         btrfs_bio_counter_dec(fs_info);
2054         bio_put(bio);
2055         return 0;
2056 }
2057
2058 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2059                          struct extent_buffer *eb, int mirror_num)
2060 {
2061         u64 start = eb->start;
2062         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2063         int ret = 0;
2064
2065         if (sb_rdonly(fs_info->sb))
2066                 return -EROFS;
2067
2068         for (i = 0; i < num_pages; i++) {
2069                 struct page *p = eb->pages[i];
2070
2071                 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2072                                         start - page_offset(p), mirror_num);
2073                 if (ret)
2074                         break;
2075                 start += PAGE_SIZE;
2076         }
2077
2078         return ret;
2079 }
2080
2081 /*
2082  * each time an IO finishes, we do a fast check in the IO failure tree
2083  * to see if we need to process or clean up an io_failure_record
2084  */
2085 int clean_io_failure(struct btrfs_fs_info *fs_info,
2086                      struct extent_io_tree *failure_tree,
2087                      struct extent_io_tree *io_tree, u64 start,
2088                      struct page *page, u64 ino, unsigned int pg_offset)
2089 {
2090         u64 private;
2091         struct io_failure_record *failrec;
2092         struct extent_state *state;
2093         int num_copies;
2094         int ret;
2095
2096         private = 0;
2097         ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2098                                EXTENT_DIRTY, 0);
2099         if (!ret)
2100                 return 0;
2101
2102         ret = get_state_failrec(failure_tree, start, &failrec);
2103         if (ret)
2104                 return 0;
2105
2106         BUG_ON(!failrec->this_mirror);
2107
2108         if (failrec->in_validation) {
2109                 /* there was no real error, just free the record */
2110                 btrfs_debug(fs_info,
2111                         "clean_io_failure: freeing dummy error at %llu",
2112                         failrec->start);
2113                 goto out;
2114         }
2115         if (sb_rdonly(fs_info->sb))
2116                 goto out;
2117
2118         spin_lock(&io_tree->lock);
2119         state = find_first_extent_bit_state(io_tree,
2120                                             failrec->start,
2121                                             EXTENT_LOCKED);
2122         spin_unlock(&io_tree->lock);
2123
2124         if (state && state->start <= failrec->start &&
2125             state->end >= failrec->start + failrec->len - 1) {
2126                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2127                                               failrec->len);
2128                 if (num_copies > 1)  {
2129                         repair_io_failure(fs_info, ino, start, failrec->len,
2130                                           failrec->logical, page, pg_offset,
2131                                           failrec->failed_mirror);
2132                 }
2133         }
2134
2135 out:
2136         free_io_failure(failure_tree, io_tree, failrec);
2137
2138         return 0;
2139 }
2140
2141 /*
2142  * Can be called when
2143  * - hold extent lock
2144  * - under ordered extent
2145  * - the inode is freeing
2146  */
2147 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2148 {
2149         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2150         struct io_failure_record *failrec;
2151         struct extent_state *state, *next;
2152
2153         if (RB_EMPTY_ROOT(&failure_tree->state))
2154                 return;
2155
2156         spin_lock(&failure_tree->lock);
2157         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2158         while (state) {
2159                 if (state->start > end)
2160                         break;
2161
2162                 ASSERT(state->end <= end);
2163
2164                 next = next_state(state);
2165
2166                 failrec = state->failrec;
2167                 free_extent_state(state);
2168                 kfree(failrec);
2169
2170                 state = next;
2171         }
2172         spin_unlock(&failure_tree->lock);
2173 }
2174
2175 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2176                 struct io_failure_record **failrec_ret)
2177 {
2178         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2179         struct io_failure_record *failrec;
2180         struct extent_map *em;
2181         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2182         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2183         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2184         int ret;
2185         u64 logical;
2186
2187         ret = get_state_failrec(failure_tree, start, &failrec);
2188         if (ret) {
2189                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2190                 if (!failrec)
2191                         return -ENOMEM;
2192
2193                 failrec->start = start;
2194                 failrec->len = end - start + 1;
2195                 failrec->this_mirror = 0;
2196                 failrec->bio_flags = 0;
2197                 failrec->in_validation = 0;
2198
2199                 read_lock(&em_tree->lock);
2200                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2201                 if (!em) {
2202                         read_unlock(&em_tree->lock);
2203                         kfree(failrec);
2204                         return -EIO;
2205                 }
2206
2207                 if (em->start > start || em->start + em->len <= start) {
2208                         free_extent_map(em);
2209                         em = NULL;
2210                 }
2211                 read_unlock(&em_tree->lock);
2212                 if (!em) {
2213                         kfree(failrec);
2214                         return -EIO;
2215                 }
2216
2217                 logical = start - em->start;
2218                 logical = em->block_start + logical;
2219                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2220                         logical = em->block_start;
2221                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2222                         extent_set_compress_type(&failrec->bio_flags,
2223                                                  em->compress_type);
2224                 }
2225
2226                 btrfs_debug(fs_info,
2227                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2228                         logical, start, failrec->len);
2229
2230                 failrec->logical = logical;
2231                 free_extent_map(em);
2232
2233                 /* set the bits in the private failure tree */
2234                 ret = set_extent_bits(failure_tree, start, end,
2235                                         EXTENT_LOCKED | EXTENT_DIRTY);
2236                 if (ret >= 0)
2237                         ret = set_state_failrec(failure_tree, start, failrec);
2238                 /* set the bits in the inode's tree */
2239                 if (ret >= 0)
2240                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2241                 if (ret < 0) {
2242                         kfree(failrec);
2243                         return ret;
2244                 }
2245         } else {
2246                 btrfs_debug(fs_info,
2247                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2248                         failrec->logical, failrec->start, failrec->len,
2249                         failrec->in_validation);
2250                 /*
2251                  * when data can be on disk more than twice, add to failrec here
2252                  * (e.g. with a list for failed_mirror) to make
2253                  * clean_io_failure() clean all those errors at once.
2254                  */
2255         }
2256
2257         *failrec_ret = failrec;
2258
2259         return 0;
2260 }
2261
2262 bool btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2263                            struct io_failure_record *failrec, int failed_mirror)
2264 {
2265         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2266         int num_copies;
2267
2268         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2269         if (num_copies == 1) {
2270                 /*
2271                  * we only have a single copy of the data, so don't bother with
2272                  * all the retry and error correction code that follows. no
2273                  * matter what the error is, it is very likely to persist.
2274                  */
2275                 btrfs_debug(fs_info,
2276                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2277                         num_copies, failrec->this_mirror, failed_mirror);
2278                 return false;
2279         }
2280
2281         /*
2282          * there are two premises:
2283          *      a) deliver good data to the caller
2284          *      b) correct the bad sectors on disk
2285          */
2286         if (failed_bio->bi_vcnt > 1) {
2287                 /*
2288                  * to fulfill b), we need to know the exact failing sectors, as
2289                  * we don't want to rewrite any more than the failed ones. thus,
2290                  * we need separate read requests for the failed bio
2291                  *
2292                  * if the following BUG_ON triggers, our validation request got
2293                  * merged. we need separate requests for our algorithm to work.
2294                  */
2295                 BUG_ON(failrec->in_validation);
2296                 failrec->in_validation = 1;
2297                 failrec->this_mirror = failed_mirror;
2298         } else {
2299                 /*
2300                  * we're ready to fulfill a) and b) alongside. get a good copy
2301                  * of the failed sector and if we succeed, we have setup
2302                  * everything for repair_io_failure to do the rest for us.
2303                  */
2304                 if (failrec->in_validation) {
2305                         BUG_ON(failrec->this_mirror != failed_mirror);
2306                         failrec->in_validation = 0;
2307                         failrec->this_mirror = 0;
2308                 }
2309                 failrec->failed_mirror = failed_mirror;
2310                 failrec->this_mirror++;
2311                 if (failrec->this_mirror == failed_mirror)
2312                         failrec->this_mirror++;
2313         }
2314
2315         if (failrec->this_mirror > num_copies) {
2316                 btrfs_debug(fs_info,
2317                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2318                         num_copies, failrec->this_mirror, failed_mirror);
2319                 return false;
2320         }
2321
2322         return true;
2323 }
2324
2325
2326 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2327                                     struct io_failure_record *failrec,
2328                                     struct page *page, int pg_offset, int icsum,
2329                                     bio_end_io_t *endio_func, void *data)
2330 {
2331         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2332         struct bio *bio;
2333         struct btrfs_io_bio *btrfs_failed_bio;
2334         struct btrfs_io_bio *btrfs_bio;
2335
2336         bio = btrfs_io_bio_alloc(1);
2337         bio->bi_end_io = endio_func;
2338         bio->bi_iter.bi_sector = failrec->logical >> 9;
2339         bio_set_dev(bio, fs_info->fs_devices->latest_bdev);
2340         bio->bi_iter.bi_size = 0;
2341         bio->bi_private = data;
2342
2343         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2344         if (btrfs_failed_bio->csum) {
2345                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2346
2347                 btrfs_bio = btrfs_io_bio(bio);
2348                 btrfs_bio->csum = btrfs_bio->csum_inline;
2349                 icsum *= csum_size;
2350                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2351                        csum_size);
2352         }
2353
2354         bio_add_page(bio, page, failrec->len, pg_offset);
2355
2356         return bio;
2357 }
2358
2359 /*
2360  * this is a generic handler for readpage errors (default
2361  * readpage_io_failed_hook). if other copies exist, read those and write back
2362  * good data to the failed position. does not investigate in remapping the
2363  * failed extent elsewhere, hoping the device will be smart enough to do this as
2364  * needed
2365  */
2366
2367 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2368                               struct page *page, u64 start, u64 end,
2369                               int failed_mirror)
2370 {
2371         struct io_failure_record *failrec;
2372         struct inode *inode = page->mapping->host;
2373         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2374         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2375         struct bio *bio;
2376         int read_mode = 0;
2377         blk_status_t status;
2378         int ret;
2379
2380         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2381
2382         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2383         if (ret)
2384                 return ret;
2385
2386         if (!btrfs_check_repairable(inode, failed_bio, failrec,
2387                                     failed_mirror)) {
2388                 free_io_failure(failure_tree, tree, failrec);
2389                 return -EIO;
2390         }
2391
2392         if (failed_bio->bi_vcnt > 1)
2393                 read_mode |= REQ_FAILFAST_DEV;
2394
2395         phy_offset >>= inode->i_sb->s_blocksize_bits;
2396         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2397                                       start - page_offset(page),
2398                                       (int)phy_offset, failed_bio->bi_end_io,
2399                                       NULL);
2400         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2401
2402         btrfs_debug(btrfs_sb(inode->i_sb),
2403                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2404                 read_mode, failrec->this_mirror, failrec->in_validation);
2405
2406         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2407                                          failrec->bio_flags, 0);
2408         if (status) {
2409                 free_io_failure(failure_tree, tree, failrec);
2410                 bio_put(bio);
2411                 ret = blk_status_to_errno(status);
2412         }
2413
2414         return ret;
2415 }
2416
2417 /* lots and lots of room for performance fixes in the end_bio funcs */
2418
2419 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2420 {
2421         int uptodate = (err == 0);
2422         struct extent_io_tree *tree;
2423         int ret = 0;
2424
2425         tree = &BTRFS_I(page->mapping->host)->io_tree;
2426
2427         if (tree->ops && tree->ops->writepage_end_io_hook)
2428                 tree->ops->writepage_end_io_hook(page, start, end, NULL,
2429                                 uptodate);
2430
2431         if (!uptodate) {
2432                 ClearPageUptodate(page);
2433                 SetPageError(page);
2434                 ret = err < 0 ? err : -EIO;
2435                 mapping_set_error(page->mapping, ret);
2436         }
2437 }
2438
2439 /*
2440  * after a writepage IO is done, we need to:
2441  * clear the uptodate bits on error
2442  * clear the writeback bits in the extent tree for this IO
2443  * end_page_writeback if the page has no more pending IO
2444  *
2445  * Scheduling is not allowed, so the extent state tree is expected
2446  * to have one and only one object corresponding to this IO.
2447  */
2448 static void end_bio_extent_writepage(struct bio *bio)
2449 {
2450         int error = blk_status_to_errno(bio->bi_status);
2451         struct bio_vec *bvec;
2452         u64 start;
2453         u64 end;
2454         int i;
2455
2456         ASSERT(!bio_flagged(bio, BIO_CLONED));
2457         bio_for_each_segment_all(bvec, bio, i) {
2458                 struct page *page = bvec->bv_page;
2459                 struct inode *inode = page->mapping->host;
2460                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2461
2462                 /* We always issue full-page reads, but if some block
2463                  * in a page fails to read, blk_update_request() will
2464                  * advance bv_offset and adjust bv_len to compensate.
2465                  * Print a warning for nonzero offsets, and an error
2466                  * if they don't add up to a full page.  */
2467                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2468                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2469                                 btrfs_err(fs_info,
2470                                    "partial page write in btrfs with offset %u and length %u",
2471                                         bvec->bv_offset, bvec->bv_len);
2472                         else
2473                                 btrfs_info(fs_info,
2474                                    "incomplete page write in btrfs with offset %u and length %u",
2475                                         bvec->bv_offset, bvec->bv_len);
2476                 }
2477
2478                 start = page_offset(page);
2479                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2480
2481                 end_extent_writepage(page, error, start, end);
2482                 end_page_writeback(page);
2483         }
2484
2485         bio_put(bio);
2486 }
2487
2488 static void
2489 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2490                               int uptodate)
2491 {
2492         struct extent_state *cached = NULL;
2493         u64 end = start + len - 1;
2494
2495         if (uptodate && tree->track_uptodate)
2496                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2497         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2498 }
2499
2500 /*
2501  * after a readpage IO is done, we need to:
2502  * clear the uptodate bits on error
2503  * set the uptodate bits if things worked
2504  * set the page up to date if all extents in the tree are uptodate
2505  * clear the lock bit in the extent tree
2506  * unlock the page if there are no other extents locked for it
2507  *
2508  * Scheduling is not allowed, so the extent state tree is expected
2509  * to have one and only one object corresponding to this IO.
2510  */
2511 static void end_bio_extent_readpage(struct bio *bio)
2512 {
2513         struct bio_vec *bvec;
2514         int uptodate = !bio->bi_status;
2515         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2516         struct extent_io_tree *tree, *failure_tree;
2517         u64 offset = 0;
2518         u64 start;
2519         u64 end;
2520         u64 len;
2521         u64 extent_start = 0;
2522         u64 extent_len = 0;
2523         int mirror;
2524         int ret;
2525         int i;
2526
2527         ASSERT(!bio_flagged(bio, BIO_CLONED));
2528         bio_for_each_segment_all(bvec, bio, i) {
2529                 struct page *page = bvec->bv_page;
2530                 struct inode *inode = page->mapping->host;
2531                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2532
2533                 btrfs_debug(fs_info,
2534                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2535                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2536                         io_bio->mirror_num);
2537                 tree = &BTRFS_I(inode)->io_tree;
2538                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2539
2540                 /* We always issue full-page reads, but if some block
2541                  * in a page fails to read, blk_update_request() will
2542                  * advance bv_offset and adjust bv_len to compensate.
2543                  * Print a warning for nonzero offsets, and an error
2544                  * if they don't add up to a full page.  */
2545                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2546                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2547                                 btrfs_err(fs_info,
2548                                         "partial page read in btrfs with offset %u and length %u",
2549                                         bvec->bv_offset, bvec->bv_len);
2550                         else
2551                                 btrfs_info(fs_info,
2552                                         "incomplete page read in btrfs with offset %u and length %u",
2553                                         bvec->bv_offset, bvec->bv_len);
2554                 }
2555
2556                 start = page_offset(page);
2557                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2558                 len = bvec->bv_len;
2559
2560                 mirror = io_bio->mirror_num;
2561                 if (likely(uptodate && tree->ops)) {
2562                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2563                                                               page, start, end,
2564                                                               mirror);
2565                         if (ret)
2566                                 uptodate = 0;
2567                         else
2568                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2569                                                  failure_tree, tree, start,
2570                                                  page,
2571                                                  btrfs_ino(BTRFS_I(inode)), 0);
2572                 }
2573
2574                 if (likely(uptodate))
2575                         goto readpage_ok;
2576
2577                 if (tree->ops) {
2578                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2579                         if (ret == -EAGAIN) {
2580                                 /*
2581                                  * Data inode's readpage_io_failed_hook() always
2582                                  * returns -EAGAIN.
2583                                  *
2584                                  * The generic bio_readpage_error handles errors
2585                                  * the following way: If possible, new read
2586                                  * requests are created and submitted and will
2587                                  * end up in end_bio_extent_readpage as well (if
2588                                  * we're lucky, not in the !uptodate case). In
2589                                  * that case it returns 0 and we just go on with
2590                                  * the next page in our bio. If it can't handle
2591                                  * the error it will return -EIO and we remain
2592                                  * responsible for that page.
2593                                  */
2594                                 ret = bio_readpage_error(bio, offset, page,
2595                                                          start, end, mirror);
2596                                 if (ret == 0) {
2597                                         uptodate = !bio->bi_status;
2598                                         offset += len;
2599                                         continue;
2600                                 }
2601                         }
2602
2603                         /*
2604                          * metadata's readpage_io_failed_hook() always returns
2605                          * -EIO and fixes nothing.  -EIO is also returned if
2606                          * data inode error could not be fixed.
2607                          */
2608                         ASSERT(ret == -EIO);
2609                 }
2610 readpage_ok:
2611                 if (likely(uptodate)) {
2612                         loff_t i_size = i_size_read(inode);
2613                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2614                         unsigned off;
2615
2616                         /* Zero out the end if this page straddles i_size */
2617                         off = i_size & (PAGE_SIZE-1);
2618                         if (page->index == end_index && off)
2619                                 zero_user_segment(page, off, PAGE_SIZE);
2620                         SetPageUptodate(page);
2621                 } else {
2622                         ClearPageUptodate(page);
2623                         SetPageError(page);
2624                 }
2625                 unlock_page(page);
2626                 offset += len;
2627
2628                 if (unlikely(!uptodate)) {
2629                         if (extent_len) {
2630                                 endio_readpage_release_extent(tree,
2631                                                               extent_start,
2632                                                               extent_len, 1);
2633                                 extent_start = 0;
2634                                 extent_len = 0;
2635                         }
2636                         endio_readpage_release_extent(tree, start,
2637                                                       end - start + 1, 0);
2638                 } else if (!extent_len) {
2639                         extent_start = start;
2640                         extent_len = end + 1 - start;
2641                 } else if (extent_start + extent_len == start) {
2642                         extent_len += end + 1 - start;
2643                 } else {
2644                         endio_readpage_release_extent(tree, extent_start,
2645                                                       extent_len, uptodate);
2646                         extent_start = start;
2647                         extent_len = end + 1 - start;
2648                 }
2649         }
2650
2651         if (extent_len)
2652                 endio_readpage_release_extent(tree, extent_start, extent_len,
2653                                               uptodate);
2654         if (io_bio->end_io)
2655                 io_bio->end_io(io_bio, blk_status_to_errno(bio->bi_status));
2656         bio_put(bio);
2657 }
2658
2659 /*
2660  * Initialize the members up to but not including 'bio'. Use after allocating a
2661  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2662  * 'bio' because use of __GFP_ZERO is not supported.
2663  */
2664 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2665 {
2666         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2667 }
2668
2669 /*
2670  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2671  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2672  * for the appropriate container_of magic
2673  */
2674 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte)
2675 {
2676         struct bio *bio;
2677
2678         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, btrfs_bioset);
2679         bio_set_dev(bio, bdev);
2680         bio->bi_iter.bi_sector = first_byte >> 9;
2681         btrfs_io_bio_init(btrfs_io_bio(bio));
2682         return bio;
2683 }
2684
2685 struct bio *btrfs_bio_clone(struct bio *bio)
2686 {
2687         struct btrfs_io_bio *btrfs_bio;
2688         struct bio *new;
2689
2690         /* Bio allocation backed by a bioset does not fail */
2691         new = bio_clone_fast(bio, GFP_NOFS, btrfs_bioset);
2692         btrfs_bio = btrfs_io_bio(new);
2693         btrfs_io_bio_init(btrfs_bio);
2694         btrfs_bio->iter = bio->bi_iter;
2695         return new;
2696 }
2697
2698 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2699 {
2700         struct bio *bio;
2701
2702         /* Bio allocation backed by a bioset does not fail */
2703         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, btrfs_bioset);
2704         btrfs_io_bio_init(btrfs_io_bio(bio));
2705         return bio;
2706 }
2707
2708 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2709 {
2710         struct bio *bio;
2711         struct btrfs_io_bio *btrfs_bio;
2712
2713         /* this will never fail when it's backed by a bioset */
2714         bio = bio_clone_fast(orig, GFP_NOFS, btrfs_bioset);
2715         ASSERT(bio);
2716
2717         btrfs_bio = btrfs_io_bio(bio);
2718         btrfs_io_bio_init(btrfs_bio);
2719
2720         bio_trim(bio, offset >> 9, size >> 9);
2721         btrfs_bio->iter = bio->bi_iter;
2722         return bio;
2723 }
2724
2725 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2726                                        unsigned long bio_flags)
2727 {
2728         blk_status_t ret = 0;
2729         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2730         struct page *page = bvec->bv_page;
2731         struct extent_io_tree *tree = bio->bi_private;
2732         u64 start;
2733
2734         start = page_offset(page) + bvec->bv_offset;
2735
2736         bio->bi_private = NULL;
2737         bio_get(bio);
2738
2739         if (tree->ops)
2740                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
2741                                            mirror_num, bio_flags, start);
2742         else
2743                 btrfsic_submit_bio(bio);
2744
2745         bio_put(bio);
2746         return blk_status_to_errno(ret);
2747 }
2748
2749 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2750                      unsigned long offset, size_t size, struct bio *bio,
2751                      unsigned long bio_flags)
2752 {
2753         int ret = 0;
2754         if (tree->ops)
2755                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2756                                                 bio_flags);
2757         return ret;
2758
2759 }
2760
2761 /*
2762  * @opf:        bio REQ_OP_* and REQ_* flags as one value
2763  */
2764 static int submit_extent_page(unsigned int opf, struct extent_io_tree *tree,
2765                               struct writeback_control *wbc,
2766                               struct page *page, sector_t sector,
2767                               size_t size, unsigned long offset,
2768                               struct block_device *bdev,
2769                               struct bio **bio_ret,
2770                               bio_end_io_t end_io_func,
2771                               int mirror_num,
2772                               unsigned long prev_bio_flags,
2773                               unsigned long bio_flags,
2774                               bool force_bio_submit)
2775 {
2776         int ret = 0;
2777         struct bio *bio;
2778         int contig = 0;
2779         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2780         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2781
2782         if (bio_ret && *bio_ret) {
2783                 bio = *bio_ret;
2784                 if (old_compressed)
2785                         contig = bio->bi_iter.bi_sector == sector;
2786                 else
2787                         contig = bio_end_sector(bio) == sector;
2788
2789                 if (prev_bio_flags != bio_flags || !contig ||
2790                     force_bio_submit ||
2791                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2792                     bio_add_page(bio, page, page_size, offset) < page_size) {
2793                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2794                         if (ret < 0) {
2795                                 *bio_ret = NULL;
2796                                 return ret;
2797                         }
2798                         bio = NULL;
2799                 } else {
2800                         if (wbc)
2801                                 wbc_account_io(wbc, page, page_size);
2802                         return 0;
2803                 }
2804         }
2805
2806         bio = btrfs_bio_alloc(bdev, (u64)sector << 9);
2807         bio_add_page(bio, page, page_size, offset);
2808         bio->bi_end_io = end_io_func;
2809         bio->bi_private = tree;
2810         bio->bi_write_hint = page->mapping->host->i_write_hint;
2811         bio->bi_opf = opf;
2812         if (wbc) {
2813                 wbc_init_bio(wbc, bio);
2814                 wbc_account_io(wbc, page, page_size);
2815         }
2816
2817         if (bio_ret)
2818                 *bio_ret = bio;
2819         else
2820                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2821
2822         return ret;
2823 }
2824
2825 static void attach_extent_buffer_page(struct extent_buffer *eb,
2826                                       struct page *page)
2827 {
2828         if (!PagePrivate(page)) {
2829                 SetPagePrivate(page);
2830                 get_page(page);
2831                 set_page_private(page, (unsigned long)eb);
2832         } else {
2833                 WARN_ON(page->private != (unsigned long)eb);
2834         }
2835 }
2836
2837 void set_page_extent_mapped(struct page *page)
2838 {
2839         if (!PagePrivate(page)) {
2840                 SetPagePrivate(page);
2841                 get_page(page);
2842                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2843         }
2844 }
2845
2846 static struct extent_map *
2847 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2848                  u64 start, u64 len, get_extent_t *get_extent,
2849                  struct extent_map **em_cached)
2850 {
2851         struct extent_map *em;
2852
2853         if (em_cached && *em_cached) {
2854                 em = *em_cached;
2855                 if (extent_map_in_tree(em) && start >= em->start &&
2856                     start < extent_map_end(em)) {
2857                         refcount_inc(&em->refs);
2858                         return em;
2859                 }
2860
2861                 free_extent_map(em);
2862                 *em_cached = NULL;
2863         }
2864
2865         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2866         if (em_cached && !IS_ERR_OR_NULL(em)) {
2867                 BUG_ON(*em_cached);
2868                 refcount_inc(&em->refs);
2869                 *em_cached = em;
2870         }
2871         return em;
2872 }
2873 /*
2874  * basic readpage implementation.  Locked extent state structs are inserted
2875  * into the tree that are removed when the IO is done (by the end_io
2876  * handlers)
2877  * XXX JDM: This needs looking at to ensure proper page locking
2878  * return 0 on success, otherwise return error
2879  */
2880 static int __do_readpage(struct extent_io_tree *tree,
2881                          struct page *page,
2882                          get_extent_t *get_extent,
2883                          struct extent_map **em_cached,
2884                          struct bio **bio, int mirror_num,
2885                          unsigned long *bio_flags, unsigned int read_flags,
2886                          u64 *prev_em_start)
2887 {
2888         struct inode *inode = page->mapping->host;
2889         u64 start = page_offset(page);
2890         u64 page_end = start + PAGE_SIZE - 1;
2891         u64 end;
2892         u64 cur = start;
2893         u64 extent_offset;
2894         u64 last_byte = i_size_read(inode);
2895         u64 block_start;
2896         u64 cur_end;
2897         sector_t sector;
2898         struct extent_map *em;
2899         struct block_device *bdev;
2900         int ret = 0;
2901         int nr = 0;
2902         size_t pg_offset = 0;
2903         size_t iosize;
2904         size_t disk_io_size;
2905         size_t blocksize = inode->i_sb->s_blocksize;
2906         unsigned long this_bio_flag = 0;
2907
2908         set_page_extent_mapped(page);
2909
2910         end = page_end;
2911         if (!PageUptodate(page)) {
2912                 if (cleancache_get_page(page) == 0) {
2913                         BUG_ON(blocksize != PAGE_SIZE);
2914                         unlock_extent(tree, start, end);
2915                         goto out;
2916                 }
2917         }
2918
2919         if (page->index == last_byte >> PAGE_SHIFT) {
2920                 char *userpage;
2921                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2922
2923                 if (zero_offset) {
2924                         iosize = PAGE_SIZE - zero_offset;
2925                         userpage = kmap_atomic(page);
2926                         memset(userpage + zero_offset, 0, iosize);
2927                         flush_dcache_page(page);
2928                         kunmap_atomic(userpage);
2929                 }
2930         }
2931         while (cur <= end) {
2932                 bool force_bio_submit = false;
2933
2934                 if (cur >= last_byte) {
2935                         char *userpage;
2936                         struct extent_state *cached = NULL;
2937
2938                         iosize = PAGE_SIZE - pg_offset;
2939                         userpage = kmap_atomic(page);
2940                         memset(userpage + pg_offset, 0, iosize);
2941                         flush_dcache_page(page);
2942                         kunmap_atomic(userpage);
2943                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2944                                             &cached, GFP_NOFS);
2945                         unlock_extent_cached(tree, cur,
2946                                              cur + iosize - 1,
2947                                              &cached, GFP_NOFS);
2948                         break;
2949                 }
2950                 em = __get_extent_map(inode, page, pg_offset, cur,
2951                                       end - cur + 1, get_extent, em_cached);
2952                 if (IS_ERR_OR_NULL(em)) {
2953                         SetPageError(page);
2954                         unlock_extent(tree, cur, end);
2955                         break;
2956                 }
2957                 extent_offset = cur - em->start;
2958                 BUG_ON(extent_map_end(em) <= cur);
2959                 BUG_ON(end < cur);
2960
2961                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2962                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2963                         extent_set_compress_type(&this_bio_flag,
2964                                                  em->compress_type);
2965                 }
2966
2967                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2968                 cur_end = min(extent_map_end(em) - 1, end);
2969                 iosize = ALIGN(iosize, blocksize);
2970                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2971                         disk_io_size = em->block_len;
2972                         sector = em->block_start >> 9;
2973                 } else {
2974                         sector = (em->block_start + extent_offset) >> 9;
2975                         disk_io_size = iosize;
2976                 }
2977                 bdev = em->bdev;
2978                 block_start = em->block_start;
2979                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2980                         block_start = EXTENT_MAP_HOLE;
2981
2982                 /*
2983                  * If we have a file range that points to a compressed extent
2984                  * and it's followed by a consecutive file range that points to
2985                  * to the same compressed extent (possibly with a different
2986                  * offset and/or length, so it either points to the whole extent
2987                  * or only part of it), we must make sure we do not submit a
2988                  * single bio to populate the pages for the 2 ranges because
2989                  * this makes the compressed extent read zero out the pages
2990                  * belonging to the 2nd range. Imagine the following scenario:
2991                  *
2992                  *  File layout
2993                  *  [0 - 8K]                     [8K - 24K]
2994                  *    |                               |
2995                  *    |                               |
2996                  * points to extent X,         points to extent X,
2997                  * offset 4K, length of 8K     offset 0, length 16K
2998                  *
2999                  * [extent X, compressed length = 4K uncompressed length = 16K]
3000                  *
3001                  * If the bio to read the compressed extent covers both ranges,
3002                  * it will decompress extent X into the pages belonging to the
3003                  * first range and then it will stop, zeroing out the remaining
3004                  * pages that belong to the other range that points to extent X.
3005                  * So here we make sure we submit 2 bios, one for the first
3006                  * range and another one for the third range. Both will target
3007                  * the same physical extent from disk, but we can't currently
3008                  * make the compressed bio endio callback populate the pages
3009                  * for both ranges because each compressed bio is tightly
3010                  * coupled with a single extent map, and each range can have
3011                  * an extent map with a different offset value relative to the
3012                  * uncompressed data of our extent and different lengths. This
3013                  * is a corner case so we prioritize correctness over
3014                  * non-optimal behavior (submitting 2 bios for the same extent).
3015                  */
3016                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3017                     prev_em_start && *prev_em_start != (u64)-1 &&
3018                     *prev_em_start != em->start)
3019                         force_bio_submit = true;
3020
3021                 if (prev_em_start)
3022                         *prev_em_start = em->start;
3023
3024                 free_extent_map(em);
3025                 em = NULL;
3026
3027                 /* we've found a hole, just zero and go on */
3028                 if (block_start == EXTENT_MAP_HOLE) {
3029                         char *userpage;
3030                         struct extent_state *cached = NULL;
3031
3032                         userpage = kmap_atomic(page);
3033                         memset(userpage + pg_offset, 0, iosize);
3034                         flush_dcache_page(page);
3035                         kunmap_atomic(userpage);
3036
3037                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3038                                             &cached, GFP_NOFS);
3039                         unlock_extent_cached(tree, cur,
3040                                              cur + iosize - 1,
3041                                              &cached, GFP_NOFS);
3042                         cur = cur + iosize;
3043                         pg_offset += iosize;
3044                         continue;
3045                 }
3046                 /* the get_extent function already copied into the page */
3047                 if (test_range_bit(tree, cur, cur_end,
3048                                    EXTENT_UPTODATE, 1, NULL)) {
3049                         check_page_uptodate(tree, page);
3050                         unlock_extent(tree, cur, cur + iosize - 1);
3051                         cur = cur + iosize;
3052                         pg_offset += iosize;
3053                         continue;
3054                 }
3055                 /* we have an inline extent but it didn't get marked up
3056                  * to date.  Error out
3057                  */
3058                 if (block_start == EXTENT_MAP_INLINE) {
3059                         SetPageError(page);
3060                         unlock_extent(tree, cur, cur + iosize - 1);
3061                         cur = cur + iosize;
3062                         pg_offset += iosize;
3063                         continue;
3064                 }
3065
3066                 ret = submit_extent_page(REQ_OP_READ | read_flags, tree, NULL,
3067                                          page, sector, disk_io_size, pg_offset,
3068                                          bdev, bio,
3069                                          end_bio_extent_readpage, mirror_num,
3070                                          *bio_flags,
3071                                          this_bio_flag,
3072                                          force_bio_submit);
3073                 if (!ret) {
3074                         nr++;
3075                         *bio_flags = this_bio_flag;
3076                 } else {
3077                         SetPageError(page);
3078                         unlock_extent(tree, cur, cur + iosize - 1);
3079                         goto out;
3080                 }
3081                 cur = cur + iosize;
3082                 pg_offset += iosize;
3083         }
3084 out:
3085         if (!nr) {
3086                 if (!PageError(page))
3087                         SetPageUptodate(page);
3088                 unlock_page(page);
3089         }
3090         return ret;
3091 }
3092
3093 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3094                                              struct page *pages[], int nr_pages,
3095                                              u64 start, u64 end,
3096                                              get_extent_t *get_extent,
3097                                              struct extent_map **em_cached,
3098                                              struct bio **bio, int mirror_num,
3099                                              unsigned long *bio_flags,
3100                                              u64 *prev_em_start)
3101 {
3102         struct inode *inode;
3103         struct btrfs_ordered_extent *ordered;
3104         int index;
3105
3106         inode = pages[0]->mapping->host;
3107         while (1) {
3108                 lock_extent(tree, start, end);
3109                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3110                                                      end - start + 1);
3111                 if (!ordered)
3112                         break;
3113                 unlock_extent(tree, start, end);
3114                 btrfs_start_ordered_extent(inode, ordered, 1);
3115                 btrfs_put_ordered_extent(ordered);
3116         }
3117
3118         for (index = 0; index < nr_pages; index++) {
3119                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3120                               mirror_num, bio_flags, 0, prev_em_start);
3121                 put_page(pages[index]);
3122         }
3123 }
3124
3125 static void __extent_readpages(struct extent_io_tree *tree,
3126                                struct page *pages[],
3127                                int nr_pages, get_extent_t *get_extent,
3128                                struct extent_map **em_cached,
3129                                struct bio **bio, int mirror_num,
3130                                unsigned long *bio_flags,
3131                                u64 *prev_em_start)
3132 {
3133         u64 start = 0;
3134         u64 end = 0;
3135         u64 page_start;
3136         int index;
3137         int first_index = 0;
3138
3139         for (index = 0; index < nr_pages; index++) {
3140                 page_start = page_offset(pages[index]);
3141                 if (!end) {
3142                         start = page_start;
3143                         end = start + PAGE_SIZE - 1;
3144                         first_index = index;
3145                 } else if (end + 1 == page_start) {
3146                         end += PAGE_SIZE;
3147                 } else {
3148                         __do_contiguous_readpages(tree, &pages[first_index],
3149                                                   index - first_index, start,
3150                                                   end, get_extent, em_cached,
3151                                                   bio, mirror_num, bio_flags,
3152                                                   prev_em_start);
3153                         start = page_start;
3154                         end = start + PAGE_SIZE - 1;
3155                         first_index = index;
3156                 }
3157         }
3158
3159         if (end)
3160                 __do_contiguous_readpages(tree, &pages[first_index],
3161                                           index - first_index, start,
3162                                           end, get_extent, em_cached, bio,
3163                                           mirror_num, bio_flags,
3164                                           prev_em_start);
3165 }
3166
3167 static int __extent_read_full_page(struct extent_io_tree *tree,
3168                                    struct page *page,
3169                                    get_extent_t *get_extent,
3170                                    struct bio **bio, int mirror_num,
3171                                    unsigned long *bio_flags,
3172                                    unsigned int read_flags)
3173 {
3174         struct inode *inode = page->mapping->host;
3175         struct btrfs_ordered_extent *ordered;
3176         u64 start = page_offset(page);
3177         u64 end = start + PAGE_SIZE - 1;
3178         int ret;
3179
3180         while (1) {
3181                 lock_extent(tree, start, end);
3182                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3183                                                 PAGE_SIZE);
3184                 if (!ordered)
3185                         break;
3186                 unlock_extent(tree, start, end);
3187                 btrfs_start_ordered_extent(inode, ordered, 1);
3188                 btrfs_put_ordered_extent(ordered);
3189         }
3190
3191         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3192                             bio_flags, read_flags, NULL);
3193         return ret;
3194 }
3195
3196 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3197                             get_extent_t *get_extent, int mirror_num)
3198 {
3199         struct bio *bio = NULL;
3200         unsigned long bio_flags = 0;
3201         int ret;
3202
3203         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3204                                       &bio_flags, 0);
3205         if (bio)
3206                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3207         return ret;
3208 }
3209
3210 static void update_nr_written(struct writeback_control *wbc,
3211                               unsigned long nr_written)
3212 {
3213         wbc->nr_to_write -= nr_written;
3214 }
3215
3216 /*
3217  * helper for __extent_writepage, doing all of the delayed allocation setup.
3218  *
3219  * This returns 1 if our fill_delalloc function did all the work required
3220  * to write the page (copy into inline extent).  In this case the IO has
3221  * been started and the page is already unlocked.
3222  *
3223  * This returns 0 if all went well (page still locked)
3224  * This returns < 0 if there were errors (page still locked)
3225  */
3226 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3227                               struct page *page, struct writeback_control *wbc,
3228                               struct extent_page_data *epd,
3229                               u64 delalloc_start,
3230                               unsigned long *nr_written)
3231 {
3232         struct extent_io_tree *tree = epd->tree;
3233         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3234         u64 nr_delalloc;
3235         u64 delalloc_to_write = 0;
3236         u64 delalloc_end = 0;
3237         int ret;
3238         int page_started = 0;
3239
3240         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3241                 return 0;
3242
3243         while (delalloc_end < page_end) {
3244                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3245                                                page,
3246                                                &delalloc_start,
3247                                                &delalloc_end,
3248                                                BTRFS_MAX_EXTENT_SIZE);
3249                 if (nr_delalloc == 0) {
3250                         delalloc_start = delalloc_end + 1;
3251                         continue;
3252                 }
3253                 ret = tree->ops->fill_delalloc(inode, page,
3254                                                delalloc_start,
3255                                                delalloc_end,
3256                                                &page_started,
3257                                                nr_written);
3258                 /* File system has been set read-only */
3259                 if (ret) {
3260                         SetPageError(page);
3261                         /* fill_delalloc should be return < 0 for error
3262                          * but just in case, we use > 0 here meaning the
3263                          * IO is started, so we don't want to return > 0
3264                          * unless things are going well.
3265                          */
3266                         ret = ret < 0 ? ret : -EIO;
3267                         goto done;
3268                 }
3269                 /*
3270                  * delalloc_end is already one less than the total length, so
3271                  * we don't subtract one from PAGE_SIZE
3272                  */
3273                 delalloc_to_write += (delalloc_end - delalloc_start +
3274                                       PAGE_SIZE) >> PAGE_SHIFT;
3275                 delalloc_start = delalloc_end + 1;
3276         }
3277         if (wbc->nr_to_write < delalloc_to_write) {
3278                 int thresh = 8192;
3279
3280                 if (delalloc_to_write < thresh * 2)
3281                         thresh = delalloc_to_write;
3282                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3283                                          thresh);
3284         }
3285
3286         /* did the fill delalloc function already unlock and start
3287          * the IO?
3288          */
3289         if (page_started) {
3290                 /*
3291                  * we've unlocked the page, so we can't update
3292                  * the mapping's writeback index, just update
3293                  * nr_to_write.
3294                  */
3295                 wbc->nr_to_write -= *nr_written;
3296                 return 1;
3297         }
3298
3299         ret = 0;
3300
3301 done:
3302         return ret;
3303 }
3304
3305 /*
3306  * helper for __extent_writepage.  This calls the writepage start hooks,
3307  * and does the loop to map the page into extents and bios.
3308  *
3309  * We return 1 if the IO is started and the page is unlocked,
3310  * 0 if all went well (page still locked)
3311  * < 0 if there were errors (page still locked)
3312  */
3313 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3314                                  struct page *page,
3315                                  struct writeback_control *wbc,
3316                                  struct extent_page_data *epd,
3317                                  loff_t i_size,
3318                                  unsigned long nr_written,
3319                                  unsigned int write_flags, int *nr_ret)
3320 {
3321         struct extent_io_tree *tree = epd->tree;
3322         u64 start = page_offset(page);
3323         u64 page_end = start + PAGE_SIZE - 1;
3324         u64 end;
3325         u64 cur = start;
3326         u64 extent_offset;
3327         u64 block_start;
3328         u64 iosize;
3329         sector_t sector;
3330         struct extent_map *em;
3331         struct block_device *bdev;
3332         size_t pg_offset = 0;
3333         size_t blocksize;
3334         int ret = 0;
3335         int nr = 0;
3336         bool compressed;
3337
3338         if (tree->ops && tree->ops->writepage_start_hook) {
3339                 ret = tree->ops->writepage_start_hook(page, start,
3340                                                       page_end);
3341                 if (ret) {
3342                         /* Fixup worker will requeue */
3343                         if (ret == -EBUSY)
3344                                 wbc->pages_skipped++;
3345                         else
3346                                 redirty_page_for_writepage(wbc, page);
3347
3348                         update_nr_written(wbc, nr_written);
3349                         unlock_page(page);
3350                         return 1;
3351                 }
3352         }
3353
3354         /*
3355          * we don't want to touch the inode after unlocking the page,
3356          * so we update the mapping writeback index now
3357          */
3358         update_nr_written(wbc, nr_written + 1);
3359
3360         end = page_end;
3361         if (i_size <= start) {
3362                 if (tree->ops && tree->ops->writepage_end_io_hook)
3363                         tree->ops->writepage_end_io_hook(page, start,
3364                                                          page_end, NULL, 1);
3365                 goto done;
3366         }
3367
3368         blocksize = inode->i_sb->s_blocksize;
3369
3370         while (cur <= end) {
3371                 u64 em_end;
3372
3373                 if (cur >= i_size) {
3374                         if (tree->ops && tree->ops->writepage_end_io_hook)
3375                                 tree->ops->writepage_end_io_hook(page, cur,
3376                                                          page_end, NULL, 1);
3377                         break;
3378                 }
3379                 em = epd->get_extent(BTRFS_I(inode), page, pg_offset, cur,
3380                                      end - cur + 1, 1);
3381                 if (IS_ERR_OR_NULL(em)) {
3382                         SetPageError(page);
3383                         ret = PTR_ERR_OR_ZERO(em);
3384                         break;
3385                 }
3386
3387                 extent_offset = cur - em->start;
3388                 em_end = extent_map_end(em);
3389                 BUG_ON(em_end <= cur);
3390                 BUG_ON(end < cur);
3391                 iosize = min(em_end - cur, end - cur + 1);
3392                 iosize = ALIGN(iosize, blocksize);
3393                 sector = (em->block_start + extent_offset) >> 9;
3394                 bdev = em->bdev;
3395                 block_start = em->block_start;
3396                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3397                 free_extent_map(em);
3398                 em = NULL;
3399
3400                 /*
3401                  * compressed and inline extents are written through other
3402                  * paths in the FS
3403                  */
3404                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3405                     block_start == EXTENT_MAP_INLINE) {
3406                         /*
3407                          * end_io notification does not happen here for
3408                          * compressed extents
3409                          */
3410                         if (!compressed && tree->ops &&
3411                             tree->ops->writepage_end_io_hook)
3412                                 tree->ops->writepage_end_io_hook(page, cur,
3413                                                          cur + iosize - 1,
3414                                                          NULL, 1);
3415                         else if (compressed) {
3416                                 /* we don't want to end_page_writeback on
3417                                  * a compressed extent.  this happens
3418                                  * elsewhere
3419                                  */
3420                                 nr++;
3421                         }
3422
3423                         cur += iosize;
3424                         pg_offset += iosize;
3425                         continue;
3426                 }
3427
3428                 set_range_writeback(tree, cur, cur + iosize - 1);
3429                 if (!PageWriteback(page)) {
3430                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3431                                    "page %lu not writeback, cur %llu end %llu",
3432                                page->index, cur, end);
3433                 }
3434
3435                 ret = submit_extent_page(REQ_OP_WRITE | write_flags, tree, wbc,
3436                                          page, sector, iosize, pg_offset,
3437                                          bdev, &epd->bio,
3438                                          end_bio_extent_writepage,
3439                                          0, 0, 0, false);
3440                 if (ret) {
3441                         SetPageError(page);
3442                         if (PageWriteback(page))
3443                                 end_page_writeback(page);
3444                 }
3445
3446                 cur = cur + iosize;
3447                 pg_offset += iosize;
3448                 nr++;
3449         }
3450 done:
3451         *nr_ret = nr;
3452         return ret;
3453 }
3454
3455 /*
3456  * the writepage semantics are similar to regular writepage.  extent
3457  * records are inserted to lock ranges in the tree, and as dirty areas
3458  * are found, they are marked writeback.  Then the lock bits are removed
3459  * and the end_io handler clears the writeback ranges
3460  */
3461 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3462                               void *data)
3463 {
3464         struct inode *inode = page->mapping->host;
3465         struct extent_page_data *epd = data;
3466         u64 start = page_offset(page);
3467         u64 page_end = start + PAGE_SIZE - 1;
3468         int ret;
3469         int nr = 0;
3470         size_t pg_offset = 0;
3471         loff_t i_size = i_size_read(inode);
3472         unsigned long end_index = i_size >> PAGE_SHIFT;
3473         unsigned int write_flags = 0;
3474         unsigned long nr_written = 0;
3475
3476         write_flags = wbc_to_write_flags(wbc);
3477
3478         trace___extent_writepage(page, inode, wbc);
3479
3480         WARN_ON(!PageLocked(page));
3481
3482         ClearPageError(page);
3483
3484         pg_offset = i_size & (PAGE_SIZE - 1);
3485         if (page->index > end_index ||
3486            (page->index == end_index && !pg_offset)) {
3487                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3488                 unlock_page(page);
3489                 return 0;
3490         }
3491
3492         if (page->index == end_index) {
3493                 char *userpage;
3494
3495                 userpage = kmap_atomic(page);
3496                 memset(userpage + pg_offset, 0,
3497                        PAGE_SIZE - pg_offset);
3498                 kunmap_atomic(userpage);
3499                 flush_dcache_page(page);
3500         }
3501
3502         pg_offset = 0;
3503
3504         set_page_extent_mapped(page);
3505
3506         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3507         if (ret == 1)
3508                 goto done_unlocked;
3509         if (ret)
3510                 goto done;
3511
3512         ret = __extent_writepage_io(inode, page, wbc, epd,
3513                                     i_size, nr_written, write_flags, &nr);
3514         if (ret == 1)
3515                 goto done_unlocked;
3516
3517 done:
3518         if (nr == 0) {
3519                 /* make sure the mapping tag for page dirty gets cleared */
3520                 set_page_writeback(page);
3521                 end_page_writeback(page);
3522         }
3523         if (PageError(page)) {
3524                 ret = ret < 0 ? ret : -EIO;
3525                 end_extent_writepage(page, ret, start, page_end);
3526         }
3527         unlock_page(page);
3528         return ret;
3529
3530 done_unlocked:
3531         return 0;
3532 }
3533
3534 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3535 {
3536         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3537                        TASK_UNINTERRUPTIBLE);
3538 }
3539
3540 static noinline_for_stack int
3541 lock_extent_buffer_for_io(struct extent_buffer *eb,
3542                           struct btrfs_fs_info *fs_info,
3543                           struct extent_page_data *epd)
3544 {
3545         unsigned long i, num_pages;
3546         int flush = 0;
3547         int ret = 0;
3548
3549         if (!btrfs_try_tree_write_lock(eb)) {
3550                 flush = 1;
3551                 flush_write_bio(epd);
3552                 btrfs_tree_lock(eb);
3553         }
3554
3555         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3556                 btrfs_tree_unlock(eb);
3557                 if (!epd->sync_io)
3558                         return 0;
3559                 if (!flush) {
3560                         flush_write_bio(epd);
3561                         flush = 1;
3562                 }
3563                 while (1) {
3564                         wait_on_extent_buffer_writeback(eb);
3565                         btrfs_tree_lock(eb);
3566                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3567                                 break;
3568                         btrfs_tree_unlock(eb);
3569                 }
3570         }
3571
3572         /*
3573          * We need to do this to prevent races in people who check if the eb is
3574          * under IO since we can end up having no IO bits set for a short period
3575          * of time.
3576          */
3577         spin_lock(&eb->refs_lock);
3578         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3579                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3580                 spin_unlock(&eb->refs_lock);
3581                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3582                 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
3583                                          -eb->len,
3584                                          fs_info->dirty_metadata_batch);
3585                 ret = 1;
3586         } else {
3587                 spin_unlock(&eb->refs_lock);
3588         }
3589
3590         btrfs_tree_unlock(eb);
3591
3592         if (!ret)
3593                 return ret;
3594
3595         num_pages = num_extent_pages(eb->start, eb->len);
3596         for (i = 0; i < num_pages; i++) {
3597                 struct page *p = eb->pages[i];
3598
3599                 if (!trylock_page(p)) {
3600                         if (!flush) {
3601                                 flush_write_bio(epd);
3602                                 flush = 1;
3603                         }
3604                         lock_page(p);
3605                 }
3606         }
3607
3608         return ret;
3609 }
3610
3611 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3612 {
3613         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3614         smp_mb__after_atomic();
3615         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3616 }
3617
3618 static void set_btree_ioerr(struct page *page)
3619 {
3620         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3621
3622         SetPageError(page);
3623         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3624                 return;
3625
3626         /*
3627          * If writeback for a btree extent that doesn't belong to a log tree
3628          * failed, increment the counter transaction->eb_write_errors.
3629          * We do this because while the transaction is running and before it's
3630          * committing (when we call filemap_fdata[write|wait]_range against
3631          * the btree inode), we might have
3632          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3633          * returns an error or an error happens during writeback, when we're
3634          * committing the transaction we wouldn't know about it, since the pages
3635          * can be no longer dirty nor marked anymore for writeback (if a
3636          * subsequent modification to the extent buffer didn't happen before the
3637          * transaction commit), which makes filemap_fdata[write|wait]_range not
3638          * able to find the pages tagged with SetPageError at transaction
3639          * commit time. So if this happens we must abort the transaction,
3640          * otherwise we commit a super block with btree roots that point to
3641          * btree nodes/leafs whose content on disk is invalid - either garbage
3642          * or the content of some node/leaf from a past generation that got
3643          * cowed or deleted and is no longer valid.
3644          *
3645          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3646          * not be enough - we need to distinguish between log tree extents vs
3647          * non-log tree extents, and the next filemap_fdatawait_range() call
3648          * will catch and clear such errors in the mapping - and that call might
3649          * be from a log sync and not from a transaction commit. Also, checking
3650          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3651          * not done and would not be reliable - the eb might have been released
3652          * from memory and reading it back again means that flag would not be
3653          * set (since it's a runtime flag, not persisted on disk).
3654          *
3655          * Using the flags below in the btree inode also makes us achieve the
3656          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3657          * writeback for all dirty pages and before filemap_fdatawait_range()
3658          * is called, the writeback for all dirty pages had already finished
3659          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3660          * filemap_fdatawait_range() would return success, as it could not know
3661          * that writeback errors happened (the pages were no longer tagged for
3662          * writeback).
3663          */
3664         switch (eb->log_index) {
3665         case -1:
3666                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3667                 break;
3668         case 0:
3669                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3670                 break;
3671         case 1:
3672                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3673                 break;
3674         default:
3675                 BUG(); /* unexpected, logic error */
3676         }
3677 }
3678
3679 static void end_bio_extent_buffer_writepage(struct bio *bio)
3680 {
3681         struct bio_vec *bvec;
3682         struct extent_buffer *eb;
3683         int i, done;
3684
3685         ASSERT(!bio_flagged(bio, BIO_CLONED));
3686         bio_for_each_segment_all(bvec, bio, i) {
3687                 struct page *page = bvec->bv_page;
3688
3689                 eb = (struct extent_buffer *)page->private;
3690                 BUG_ON(!eb);
3691                 done = atomic_dec_and_test(&eb->io_pages);
3692
3693                 if (bio->bi_status ||
3694                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3695                         ClearPageUptodate(page);
3696                         set_btree_ioerr(page);
3697                 }
3698
3699                 end_page_writeback(page);
3700
3701                 if (!done)
3702                         continue;
3703
3704                 end_extent_buffer_writeback(eb);
3705         }
3706
3707         bio_put(bio);
3708 }
3709
3710 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3711                         struct btrfs_fs_info *fs_info,
3712                         struct writeback_control *wbc,
3713                         struct extent_page_data *epd)
3714 {
3715         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3716         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3717         u64 offset = eb->start;
3718         u32 nritems;
3719         unsigned long i, num_pages;
3720         unsigned long bio_flags = 0;
3721         unsigned long start, end;
3722         unsigned int write_flags = wbc_to_write_flags(wbc) | REQ_META;
3723         int ret = 0;
3724
3725         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3726         num_pages = num_extent_pages(eb->start, eb->len);
3727         atomic_set(&eb->io_pages, num_pages);
3728         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3729                 bio_flags = EXTENT_BIO_TREE_LOG;
3730
3731         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3732         nritems = btrfs_header_nritems(eb);
3733         if (btrfs_header_level(eb) > 0) {
3734                 end = btrfs_node_key_ptr_offset(nritems);
3735
3736                 memzero_extent_buffer(eb, end, eb->len - end);
3737         } else {
3738                 /*
3739                  * leaf:
3740                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3741                  */
3742                 start = btrfs_item_nr_offset(nritems);
3743                 end = BTRFS_LEAF_DATA_OFFSET + leaf_data_end(fs_info, eb);
3744                 memzero_extent_buffer(eb, start, end - start);
3745         }
3746
3747         for (i = 0; i < num_pages; i++) {
3748                 struct page *p = eb->pages[i];
3749
3750                 clear_page_dirty_for_io(p);
3751                 set_page_writeback(p);
3752                 ret = submit_extent_page(REQ_OP_WRITE | write_flags, tree, wbc,
3753                                          p, offset >> 9, PAGE_SIZE, 0, bdev,
3754                                          &epd->bio,
3755                                          end_bio_extent_buffer_writepage,
3756                                          0, epd->bio_flags, bio_flags, false);
3757                 epd->bio_flags = bio_flags;
3758                 if (ret) {
3759                         set_btree_ioerr(p);
3760                         if (PageWriteback(p))
3761                                 end_page_writeback(p);
3762                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3763                                 end_extent_buffer_writeback(eb);
3764                         ret = -EIO;
3765                         break;
3766                 }
3767                 offset += PAGE_SIZE;
3768                 update_nr_written(wbc, 1);
3769                 unlock_page(p);
3770         }
3771
3772         if (unlikely(ret)) {
3773                 for (; i < num_pages; i++) {
3774                         struct page *p = eb->pages[i];
3775                         clear_page_dirty_for_io(p);
3776                         unlock_page(p);
3777                 }
3778         }
3779
3780         return ret;
3781 }
3782
3783 int btree_write_cache_pages(struct address_space *mapping,
3784                                    struct writeback_control *wbc)
3785 {
3786         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3787         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3788         struct extent_buffer *eb, *prev_eb = NULL;
3789         struct extent_page_data epd = {
3790                 .bio = NULL,
3791                 .tree = tree,
3792                 .extent_locked = 0,
3793                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3794                 .bio_flags = 0,
3795         };
3796         int ret = 0;
3797         int done = 0;
3798         int nr_to_write_done = 0;
3799         struct pagevec pvec;
3800         int nr_pages;
3801         pgoff_t index;
3802         pgoff_t end;            /* Inclusive */
3803         int scanned = 0;
3804         int tag;
3805
3806         pagevec_init(&pvec, 0);
3807         if (wbc->range_cyclic) {
3808                 index = mapping->writeback_index; /* Start from prev offset */
3809                 end = -1;
3810         } else {
3811                 index = wbc->range_start >> PAGE_SHIFT;
3812                 end = wbc->range_end >> PAGE_SHIFT;
3813                 scanned = 1;
3814         }
3815         if (wbc->sync_mode == WB_SYNC_ALL)
3816                 tag = PAGECACHE_TAG_TOWRITE;
3817         else
3818                 tag = PAGECACHE_TAG_DIRTY;
3819 retry:
3820         if (wbc->sync_mode == WB_SYNC_ALL)
3821                 tag_pages_for_writeback(mapping, index, end);
3822         while (!done && !nr_to_write_done && (index <= end) &&
3823                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3824                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3825                 unsigned i;
3826
3827                 scanned = 1;
3828                 for (i = 0; i < nr_pages; i++) {
3829                         struct page *page = pvec.pages[i];
3830
3831                         if (!PagePrivate(page))
3832                                 continue;
3833
3834                         if (!wbc->range_cyclic && page->index > end) {
3835                                 done = 1;
3836                                 break;
3837                         }
3838
3839                         spin_lock(&mapping->private_lock);
3840                         if (!PagePrivate(page)) {
3841                                 spin_unlock(&mapping->private_lock);
3842                                 continue;
3843                         }
3844
3845                         eb = (struct extent_buffer *)page->private;
3846
3847                         /*
3848                          * Shouldn't happen and normally this would be a BUG_ON
3849                          * but no sense in crashing the users box for something
3850                          * we can survive anyway.
3851                          */
3852                         if (WARN_ON(!eb)) {
3853                                 spin_unlock(&mapping->private_lock);
3854                                 continue;
3855                         }
3856
3857                         if (eb == prev_eb) {
3858                                 spin_unlock(&mapping->private_lock);
3859                                 continue;
3860                         }
3861
3862                         ret = atomic_inc_not_zero(&eb->refs);
3863                         spin_unlock(&mapping->private_lock);
3864                         if (!ret)
3865                                 continue;
3866
3867                         prev_eb = eb;
3868                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3869                         if (!ret) {
3870                                 free_extent_buffer(eb);
3871                                 continue;
3872                         } else if (ret < 0) {
3873                                 done = 1;
3874                                 free_extent_buffer(eb);
3875                                 break;
3876                         }
3877
3878                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3879                         if (ret) {
3880                                 done = 1;
3881                                 free_extent_buffer(eb);
3882                                 break;
3883                         }
3884                         free_extent_buffer(eb);
3885
3886                         /*
3887                          * the filesystem may choose to bump up nr_to_write.
3888                          * We have to make sure to honor the new nr_to_write
3889                          * at any time
3890                          */
3891                         nr_to_write_done = wbc->nr_to_write <= 0;
3892                 }
3893                 pagevec_release(&pvec);
3894                 cond_resched();
3895         }
3896         if (!scanned && !done) {
3897                 /*
3898                  * We hit the last page and there is more work to be done: wrap
3899                  * back to the start of the file
3900                  */
3901                 scanned = 1;
3902                 index = 0;
3903                 goto retry;
3904         }
3905         flush_write_bio(&epd);
3906         return ret;
3907 }
3908
3909 /**
3910  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3911  * @mapping: address space structure to write
3912  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3913  * @writepage: function called for each page
3914  * @data: data passed to writepage function
3915  *
3916  * If a page is already under I/O, write_cache_pages() skips it, even
3917  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3918  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3919  * and msync() need to guarantee that all the data which was dirty at the time
3920  * the call was made get new I/O started against them.  If wbc->sync_mode is
3921  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3922  * existing IO to complete.
3923  */
3924 static int extent_write_cache_pages(struct address_space *mapping,
3925                              struct writeback_control *wbc,
3926                              writepage_t writepage, void *data,
3927                              void (*flush_fn)(void *))
3928 {
3929         struct inode *inode = mapping->host;
3930         int ret = 0;
3931         int done = 0;
3932         int nr_to_write_done = 0;
3933         struct pagevec pvec;
3934         int nr_pages;
3935         pgoff_t index;
3936         pgoff_t end;            /* Inclusive */
3937         pgoff_t done_index;
3938         int range_whole = 0;
3939         int scanned = 0;
3940         int tag;
3941
3942         /*
3943          * We have to hold onto the inode so that ordered extents can do their
3944          * work when the IO finishes.  The alternative to this is failing to add
3945          * an ordered extent if the igrab() fails there and that is a huge pain
3946          * to deal with, so instead just hold onto the inode throughout the
3947          * writepages operation.  If it fails here we are freeing up the inode
3948          * anyway and we'd rather not waste our time writing out stuff that is
3949          * going to be truncated anyway.
3950          */
3951         if (!igrab(inode))
3952                 return 0;
3953
3954         pagevec_init(&pvec, 0);
3955         if (wbc->range_cyclic) {
3956                 index = mapping->writeback_index; /* Start from prev offset */
3957                 end = -1;
3958         } else {
3959                 index = wbc->range_start >> PAGE_SHIFT;
3960                 end = wbc->range_end >> PAGE_SHIFT;
3961                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3962                         range_whole = 1;
3963                 scanned = 1;
3964         }
3965         if (wbc->sync_mode == WB_SYNC_ALL)
3966                 tag = PAGECACHE_TAG_TOWRITE;
3967         else
3968                 tag = PAGECACHE_TAG_DIRTY;
3969 retry:
3970         if (wbc->sync_mode == WB_SYNC_ALL)
3971                 tag_pages_for_writeback(mapping, index, end);
3972         done_index = index;
3973         while (!done && !nr_to_write_done && (index <= end) &&
3974                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3975                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3976                 unsigned i;
3977
3978                 scanned = 1;
3979                 for (i = 0; i < nr_pages; i++) {
3980                         struct page *page = pvec.pages[i];
3981
3982                         done_index = page->index;
3983                         /*
3984                          * At this point we hold neither mapping->tree_lock nor
3985                          * lock on the page itself: the page may be truncated or
3986                          * invalidated (changing page->mapping to NULL), or even
3987                          * swizzled back from swapper_space to tmpfs file
3988                          * mapping
3989                          */
3990                         if (!trylock_page(page)) {
3991                                 flush_fn(data);
3992                                 lock_page(page);
3993                         }
3994
3995                         if (unlikely(page->mapping != mapping)) {
3996                                 unlock_page(page);
3997                                 continue;
3998                         }
3999
4000                         if (!wbc->range_cyclic && page->index > end) {
4001                                 done = 1;
4002                                 unlock_page(page);
4003                                 continue;
4004                         }
4005
4006                         if (wbc->sync_mode != WB_SYNC_NONE) {
4007                                 if (PageWriteback(page))
4008                                         flush_fn(data);
4009                                 wait_on_page_writeback(page);
4010                         }
4011
4012                         if (PageWriteback(page) ||
4013                             !clear_page_dirty_for_io(page)) {
4014                                 unlock_page(page);
4015                                 continue;
4016                         }
4017
4018                         ret = (*writepage)(page, wbc, data);
4019
4020                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4021                                 unlock_page(page);
4022                                 ret = 0;
4023                         }
4024                         if (ret < 0) {
4025                                 /*
4026                                  * done_index is set past this page,
4027                                  * so media errors will not choke
4028                                  * background writeout for the entire
4029                                  * file. This has consequences for
4030                                  * range_cyclic semantics (ie. it may
4031                                  * not be suitable for data integrity
4032                                  * writeout).
4033                                  */
4034                                 done_index = page->index + 1;
4035                                 done = 1;
4036                                 break;
4037                         }
4038
4039                         /*
4040                          * the filesystem may choose to bump up nr_to_write.
4041                          * We have to make sure to honor the new nr_to_write
4042                          * at any time
4043                          */
4044                         nr_to_write_done = wbc->nr_to_write <= 0;
4045                 }
4046                 pagevec_release(&pvec);
4047                 cond_resched();
4048         }
4049         if (!scanned && !done) {
4050                 /*
4051                  * We hit the last page and there is more work to be done: wrap
4052                  * back to the start of the file
4053                  */
4054                 scanned = 1;
4055                 index = 0;
4056
4057                 /*
4058                  * If we're looping we could run into a page that is locked by a
4059                  * writer and that writer could be waiting on writeback for a
4060                  * page in our current bio, and thus deadlock, so flush the
4061                  * write bio here.
4062                  */
4063                 flush_write_bio(data);
4064                 goto retry;
4065         }
4066
4067         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4068                 mapping->writeback_index = done_index;
4069
4070         btrfs_add_delayed_iput(inode);
4071         return ret;
4072 }
4073
4074 static void flush_epd_write_bio(struct extent_page_data *epd)
4075 {
4076         if (epd->bio) {
4077                 int ret;
4078
4079                 ret = submit_one_bio(epd->bio, 0, epd->bio_flags);
4080                 BUG_ON(ret < 0); /* -ENOMEM */
4081                 epd->bio = NULL;
4082         }
4083 }
4084
4085 static noinline void flush_write_bio(void *data)
4086 {
4087         struct extent_page_data *epd = data;
4088         flush_epd_write_bio(epd);
4089 }
4090
4091 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4092                           get_extent_t *get_extent,
4093                           struct writeback_control *wbc)
4094 {
4095         int ret;
4096         struct extent_page_data epd = {
4097                 .bio = NULL,
4098                 .tree = tree,
4099                 .get_extent = get_extent,
4100                 .extent_locked = 0,
4101                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4102                 .bio_flags = 0,
4103         };
4104
4105         ret = __extent_writepage(page, wbc, &epd);
4106
4107         flush_epd_write_bio(&epd);
4108         return ret;
4109 }
4110
4111 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4112                               u64 start, u64 end, get_extent_t *get_extent,
4113                               int mode)
4114 {
4115         int ret = 0;
4116         struct address_space *mapping = inode->i_mapping;
4117         struct page *page;
4118         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4119                 PAGE_SHIFT;
4120
4121         struct extent_page_data epd = {
4122                 .bio = NULL,
4123                 .tree = tree,
4124                 .get_extent = get_extent,
4125                 .extent_locked = 1,
4126                 .sync_io = mode == WB_SYNC_ALL,
4127                 .bio_flags = 0,
4128         };
4129         struct writeback_control wbc_writepages = {
4130                 .sync_mode      = mode,
4131                 .nr_to_write    = nr_pages * 2,
4132                 .range_start    = start,
4133                 .range_end      = end + 1,
4134         };
4135
4136         while (start <= end) {
4137                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4138                 if (clear_page_dirty_for_io(page))
4139                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4140                 else {
4141                         if (tree->ops && tree->ops->writepage_end_io_hook)
4142                                 tree->ops->writepage_end_io_hook(page, start,
4143                                                  start + PAGE_SIZE - 1,
4144                                                  NULL, 1);
4145                         unlock_page(page);
4146                 }
4147                 put_page(page);
4148                 start += PAGE_SIZE;
4149         }
4150
4151         flush_epd_write_bio(&epd);
4152         return ret;
4153 }
4154
4155 int extent_writepages(struct extent_io_tree *tree,
4156                       struct address_space *mapping,
4157                       get_extent_t *get_extent,
4158                       struct writeback_control *wbc)
4159 {
4160         int ret = 0;
4161         struct extent_page_data epd = {
4162                 .bio = NULL,
4163                 .tree = tree,
4164                 .get_extent = get_extent,
4165                 .extent_locked = 0,
4166                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4167                 .bio_flags = 0,
4168         };
4169
4170         ret = extent_write_cache_pages(mapping, wbc, __extent_writepage, &epd,
4171                                        flush_write_bio);
4172         flush_epd_write_bio(&epd);
4173         return ret;
4174 }
4175
4176 int extent_readpages(struct extent_io_tree *tree,
4177                      struct address_space *mapping,
4178                      struct list_head *pages, unsigned nr_pages,
4179                      get_extent_t get_extent)
4180 {
4181         struct bio *bio = NULL;
4182         unsigned page_idx;
4183         unsigned long bio_flags = 0;
4184         struct page *pagepool[16];
4185         struct page *page;
4186         struct extent_map *em_cached = NULL;
4187         int nr = 0;
4188         u64 prev_em_start = (u64)-1;
4189
4190         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4191                 page = list_entry(pages->prev, struct page, lru);
4192
4193                 prefetchw(&page->flags);
4194                 list_del(&page->lru);
4195                 if (add_to_page_cache_lru(page, mapping,
4196                                         page->index,
4197                                         readahead_gfp_mask(mapping))) {
4198                         put_page(page);
4199                         continue;
4200                 }
4201
4202                 pagepool[nr++] = page;
4203                 if (nr < ARRAY_SIZE(pagepool))
4204                         continue;
4205                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4206                                    &bio, 0, &bio_flags, &prev_em_start);
4207                 nr = 0;
4208         }
4209         if (nr)
4210                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4211                                    &bio, 0, &bio_flags, &prev_em_start);
4212
4213         if (em_cached)
4214                 free_extent_map(em_cached);
4215
4216         BUG_ON(!list_empty(pages));
4217         if (bio)
4218                 return submit_one_bio(bio, 0, bio_flags);
4219         return 0;
4220 }
4221
4222 /*
4223  * basic invalidatepage code, this waits on any locked or writeback
4224  * ranges corresponding to the page, and then deletes any extent state
4225  * records from the tree
4226  */
4227 int extent_invalidatepage(struct extent_io_tree *tree,
4228                           struct page *page, unsigned long offset)
4229 {
4230         struct extent_state *cached_state = NULL;
4231         u64 start = page_offset(page);
4232         u64 end = start + PAGE_SIZE - 1;
4233         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4234
4235         start += ALIGN(offset, blocksize);
4236         if (start > end)
4237                 return 0;
4238
4239         lock_extent_bits(tree, start, end, &cached_state);
4240         wait_on_page_writeback(page);
4241         clear_extent_bit(tree, start, end,
4242                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4243                          EXTENT_DO_ACCOUNTING,
4244                          1, 1, &cached_state, GFP_NOFS);
4245         return 0;
4246 }
4247
4248 /*
4249  * a helper for releasepage, this tests for areas of the page that
4250  * are locked or under IO and drops the related state bits if it is safe
4251  * to drop the page.
4252  */
4253 static int try_release_extent_state(struct extent_map_tree *map,
4254                                     struct extent_io_tree *tree,
4255                                     struct page *page, gfp_t mask)
4256 {
4257         u64 start = page_offset(page);
4258         u64 end = start + PAGE_SIZE - 1;
4259         int ret = 1;
4260
4261         if (test_range_bit(tree, start, end,
4262                            EXTENT_IOBITS, 0, NULL))
4263                 ret = 0;
4264         else {
4265                 /*
4266                  * at this point we can safely clear everything except the
4267                  * locked bit and the nodatasum bit
4268                  */
4269                 ret = clear_extent_bit(tree, start, end,
4270                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4271                                  0, 0, NULL, mask);
4272
4273                 /* if clear_extent_bit failed for enomem reasons,
4274                  * we can't allow the release to continue.
4275                  */
4276                 if (ret < 0)
4277                         ret = 0;
4278                 else
4279                         ret = 1;
4280         }
4281         return ret;
4282 }
4283
4284 /*
4285  * a helper for releasepage.  As long as there are no locked extents
4286  * in the range corresponding to the page, both state records and extent
4287  * map records are removed
4288  */
4289 int try_release_extent_mapping(struct extent_map_tree *map,
4290                                struct extent_io_tree *tree, struct page *page,
4291                                gfp_t mask)
4292 {
4293         struct extent_map *em;
4294         u64 start = page_offset(page);
4295         u64 end = start + PAGE_SIZE - 1;
4296         struct btrfs_inode *btrfs_inode = BTRFS_I(page->mapping->host);
4297
4298         if (gfpflags_allow_blocking(mask) &&
4299             page->mapping->host->i_size > SZ_16M) {
4300                 u64 len;
4301                 while (start <= end) {
4302                         len = end - start + 1;
4303                         write_lock(&map->lock);
4304                         em = lookup_extent_mapping(map, start, len);
4305                         if (!em) {
4306                                 write_unlock(&map->lock);
4307                                 break;
4308                         }
4309                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4310                             em->start != start) {
4311                                 write_unlock(&map->lock);
4312                                 free_extent_map(em);
4313                                 break;
4314                         }
4315                         if (!test_range_bit(tree, em->start,
4316                                             extent_map_end(em) - 1,
4317                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4318                                             0, NULL)) {
4319                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4320                                         &btrfs_inode->runtime_flags);
4321                                 remove_extent_mapping(map, em);
4322                                 /* once for the rb tree */
4323                                 free_extent_map(em);
4324                         }
4325                         start = extent_map_end(em);
4326                         write_unlock(&map->lock);
4327
4328                         /* once for us */
4329                         free_extent_map(em);
4330
4331                         cond_resched(); /* Allow large-extent preemption. */
4332                 }
4333         }
4334         return try_release_extent_state(map, tree, page, mask);
4335 }
4336
4337 /*
4338  * helper function for fiemap, which doesn't want to see any holes.
4339  * This maps until we find something past 'last'
4340  */
4341 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4342                                                 u64 offset,
4343                                                 u64 last,
4344                                                 get_extent_t *get_extent)
4345 {
4346         u64 sectorsize = btrfs_inode_sectorsize(inode);
4347         struct extent_map *em;
4348         u64 len;
4349
4350         if (offset >= last)
4351                 return NULL;
4352
4353         while (1) {
4354                 len = last - offset;
4355                 if (len == 0)
4356                         break;
4357                 len = ALIGN(len, sectorsize);
4358                 em = get_extent(BTRFS_I(inode), NULL, 0, offset, len, 0);
4359                 if (IS_ERR_OR_NULL(em))
4360                         return em;
4361
4362                 /* if this isn't a hole return it */
4363                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4364                     em->block_start != EXTENT_MAP_HOLE) {
4365                         return em;
4366                 }
4367
4368                 /* this is a hole, advance to the next extent */
4369                 offset = extent_map_end(em);
4370                 free_extent_map(em);
4371                 if (offset >= last)
4372                         break;
4373         }
4374         return NULL;
4375 }
4376
4377 /*
4378  * To cache previous fiemap extent
4379  *
4380  * Will be used for merging fiemap extent
4381  */
4382 struct fiemap_cache {
4383         u64 offset;
4384         u64 phys;
4385         u64 len;
4386         u32 flags;
4387         bool cached;
4388 };
4389
4390 /*
4391  * Helper to submit fiemap extent.
4392  *
4393  * Will try to merge current fiemap extent specified by @offset, @phys,
4394  * @len and @flags with cached one.
4395  * And only when we fails to merge, cached one will be submitted as
4396  * fiemap extent.
4397  *
4398  * Return value is the same as fiemap_fill_next_extent().
4399  */
4400 static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
4401                                 struct fiemap_cache *cache,
4402                                 u64 offset, u64 phys, u64 len, u32 flags)
4403 {
4404         int ret = 0;
4405
4406         if (!cache->cached)
4407                 goto assign;
4408
4409         /*
4410          * Sanity check, extent_fiemap() should have ensured that new
4411          * fiemap extent won't overlap with cahced one.
4412          * Not recoverable.
4413          *
4414          * NOTE: Physical address can overlap, due to compression
4415          */
4416         if (cache->offset + cache->len > offset) {
4417                 WARN_ON(1);
4418                 return -EINVAL;
4419         }
4420
4421         /*
4422          * Only merges fiemap extents if
4423          * 1) Their logical addresses are continuous
4424          *
4425          * 2) Their physical addresses are continuous
4426          *    So truly compressed (physical size smaller than logical size)
4427          *    extents won't get merged with each other
4428          *
4429          * 3) Share same flags except FIEMAP_EXTENT_LAST
4430          *    So regular extent won't get merged with prealloc extent
4431          */
4432         if (cache->offset + cache->len  == offset &&
4433             cache->phys + cache->len == phys  &&
4434             (cache->flags & ~FIEMAP_EXTENT_LAST) ==
4435                         (flags & ~FIEMAP_EXTENT_LAST)) {
4436                 cache->len += len;
4437                 cache->flags |= flags;
4438                 goto try_submit_last;
4439         }
4440
4441         /* Not mergeable, need to submit cached one */
4442         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4443                                       cache->len, cache->flags);
4444         cache->cached = false;
4445         if (ret)
4446                 return ret;
4447 assign:
4448         cache->cached = true;
4449         cache->offset = offset;
4450         cache->phys = phys;
4451         cache->len = len;
4452         cache->flags = flags;
4453 try_submit_last:
4454         if (cache->flags & FIEMAP_EXTENT_LAST) {
4455                 ret = fiemap_fill_next_extent(fieinfo, cache->offset,
4456                                 cache->phys, cache->len, cache->flags);
4457                 cache->cached = false;
4458         }
4459         return ret;
4460 }
4461
4462 /*
4463  * Emit last fiemap cache
4464  *
4465  * The last fiemap cache may still be cached in the following case:
4466  * 0                  4k                    8k
4467  * |<- Fiemap range ->|
4468  * |<------------  First extent ----------->|
4469  *
4470  * In this case, the first extent range will be cached but not emitted.
4471  * So we must emit it before ending extent_fiemap().
4472  */
4473 static int emit_last_fiemap_cache(struct btrfs_fs_info *fs_info,
4474                                   struct fiemap_extent_info *fieinfo,
4475                                   struct fiemap_cache *cache)
4476 {
4477         int ret;
4478
4479         if (!cache->cached)
4480                 return 0;
4481
4482         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4483                                       cache->len, cache->flags);
4484         cache->cached = false;
4485         if (ret > 0)
4486                 ret = 0;
4487         return ret;
4488 }
4489
4490 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4491                 __u64 start, __u64 len, get_extent_t *get_extent)
4492 {
4493         int ret = 0;
4494         u64 off = start;
4495         u64 max = start + len;
4496         u32 flags = 0;
4497         u32 found_type;
4498         u64 last;
4499         u64 last_for_get_extent = 0;
4500         u64 disko = 0;
4501         u64 isize = i_size_read(inode);
4502         struct btrfs_key found_key;
4503         struct extent_map *em = NULL;
4504         struct extent_state *cached_state = NULL;
4505         struct btrfs_path *path;
4506         struct btrfs_root *root = BTRFS_I(inode)->root;
4507         struct fiemap_cache cache = { 0 };
4508         int end = 0;
4509         u64 em_start = 0;
4510         u64 em_len = 0;
4511         u64 em_end = 0;
4512
4513         if (len == 0)
4514                 return -EINVAL;
4515
4516         path = btrfs_alloc_path();
4517         if (!path)
4518                 return -ENOMEM;
4519         path->leave_spinning = 1;
4520
4521         start = round_down(start, btrfs_inode_sectorsize(inode));
4522         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4523
4524         /*
4525          * lookup the last file extent.  We're not using i_size here
4526          * because there might be preallocation past i_size
4527          */
4528         ret = btrfs_lookup_file_extent(NULL, root, path,
4529                         btrfs_ino(BTRFS_I(inode)), -1, 0);
4530         if (ret < 0) {
4531                 btrfs_free_path(path);
4532                 return ret;
4533         } else {
4534                 WARN_ON(!ret);
4535                 if (ret == 1)
4536                         ret = 0;
4537         }
4538
4539         path->slots[0]--;
4540         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4541         found_type = found_key.type;
4542
4543         /* No extents, but there might be delalloc bits */
4544         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4545             found_type != BTRFS_EXTENT_DATA_KEY) {
4546                 /* have to trust i_size as the end */
4547                 last = (u64)-1;
4548                 last_for_get_extent = isize;
4549         } else {
4550                 /*
4551                  * remember the start of the last extent.  There are a
4552                  * bunch of different factors that go into the length of the
4553                  * extent, so its much less complex to remember where it started
4554                  */
4555                 last = found_key.offset;
4556                 last_for_get_extent = last + 1;
4557         }
4558         btrfs_release_path(path);
4559
4560         /*
4561          * we might have some extents allocated but more delalloc past those
4562          * extents.  so, we trust isize unless the start of the last extent is
4563          * beyond isize
4564          */
4565         if (last < isize) {
4566                 last = (u64)-1;
4567                 last_for_get_extent = isize;
4568         }
4569
4570         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4571                          &cached_state);
4572
4573         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4574                                    get_extent);
4575         if (!em)
4576                 goto out;
4577         if (IS_ERR(em)) {
4578                 ret = PTR_ERR(em);
4579                 goto out;
4580         }
4581
4582         while (!end) {
4583                 u64 offset_in_extent = 0;
4584
4585                 /* break if the extent we found is outside the range */
4586                 if (em->start >= max || extent_map_end(em) < off)
4587                         break;
4588
4589                 /*
4590                  * get_extent may return an extent that starts before our
4591                  * requested range.  We have to make sure the ranges
4592                  * we return to fiemap always move forward and don't
4593                  * overlap, so adjust the offsets here
4594                  */
4595                 em_start = max(em->start, off);
4596
4597                 /*
4598                  * record the offset from the start of the extent
4599                  * for adjusting the disk offset below.  Only do this if the
4600                  * extent isn't compressed since our in ram offset may be past
4601                  * what we have actually allocated on disk.
4602                  */
4603                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4604                         offset_in_extent = em_start - em->start;
4605                 em_end = extent_map_end(em);
4606                 em_len = em_end - em_start;
4607                 disko = 0;
4608                 flags = 0;
4609
4610                 /*
4611                  * bump off for our next call to get_extent
4612                  */
4613                 off = extent_map_end(em);
4614                 if (off >= max)
4615                         end = 1;
4616
4617                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4618                         end = 1;
4619                         flags |= FIEMAP_EXTENT_LAST;
4620                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4621                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4622                                   FIEMAP_EXTENT_NOT_ALIGNED);
4623                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4624                         flags |= (FIEMAP_EXTENT_DELALLOC |
4625                                   FIEMAP_EXTENT_UNKNOWN);
4626                 } else if (fieinfo->fi_extents_max) {
4627                         u64 bytenr = em->block_start -
4628                                 (em->start - em->orig_start);
4629
4630                         disko = em->block_start + offset_in_extent;
4631
4632                         /*
4633                          * As btrfs supports shared space, this information
4634                          * can be exported to userspace tools via
4635                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4636                          * then we're just getting a count and we can skip the
4637                          * lookup stuff.
4638                          */
4639                         ret = btrfs_check_shared(root,
4640                                                  btrfs_ino(BTRFS_I(inode)),
4641                                                  bytenr);
4642                         if (ret < 0)
4643                                 goto out_free;
4644                         if (ret)
4645                                 flags |= FIEMAP_EXTENT_SHARED;
4646                         ret = 0;
4647                 }
4648                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4649                         flags |= FIEMAP_EXTENT_ENCODED;
4650                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4651                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4652
4653                 free_extent_map(em);
4654                 em = NULL;
4655                 if ((em_start >= last) || em_len == (u64)-1 ||
4656                    (last == (u64)-1 && isize <= em_end)) {
4657                         flags |= FIEMAP_EXTENT_LAST;
4658                         end = 1;
4659                 }
4660
4661                 /* now scan forward to see if this is really the last extent. */
4662                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4663                                            get_extent);
4664                 if (IS_ERR(em)) {
4665                         ret = PTR_ERR(em);
4666                         goto out;
4667                 }
4668                 if (!em) {
4669                         flags |= FIEMAP_EXTENT_LAST;
4670                         end = 1;
4671                 }
4672                 ret = emit_fiemap_extent(fieinfo, &cache, em_start, disko,
4673                                            em_len, flags);
4674                 if (ret) {
4675                         if (ret == 1)
4676                                 ret = 0;
4677                         goto out_free;
4678                 }
4679         }
4680 out_free:
4681         if (!ret)
4682                 ret = emit_last_fiemap_cache(root->fs_info, fieinfo, &cache);
4683         free_extent_map(em);
4684 out:
4685         btrfs_free_path(path);
4686         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4687                              &cached_state, GFP_NOFS);
4688         return ret;
4689 }
4690
4691 static void __free_extent_buffer(struct extent_buffer *eb)
4692 {
4693         btrfs_leak_debug_del(&eb->leak_list);
4694         kmem_cache_free(extent_buffer_cache, eb);
4695 }
4696
4697 int extent_buffer_under_io(struct extent_buffer *eb)
4698 {
4699         return (atomic_read(&eb->io_pages) ||
4700                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4701                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4702 }
4703
4704 /*
4705  * Helper for releasing extent buffer page.
4706  */
4707 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4708 {
4709         unsigned long index;
4710         struct page *page;
4711         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4712
4713         BUG_ON(extent_buffer_under_io(eb));
4714
4715         index = num_extent_pages(eb->start, eb->len);
4716         if (index == 0)
4717                 return;
4718
4719         do {
4720                 index--;
4721                 page = eb->pages[index];
4722                 if (!page)
4723                         continue;
4724                 if (mapped)
4725                         spin_lock(&page->mapping->private_lock);
4726                 /*
4727                  * We do this since we'll remove the pages after we've
4728                  * removed the eb from the radix tree, so we could race
4729                  * and have this page now attached to the new eb.  So
4730                  * only clear page_private if it's still connected to
4731                  * this eb.
4732                  */
4733                 if (PagePrivate(page) &&
4734                     page->private == (unsigned long)eb) {
4735                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4736                         BUG_ON(PageDirty(page));
4737                         BUG_ON(PageWriteback(page));
4738                         /*
4739                          * We need to make sure we haven't be attached
4740                          * to a new eb.
4741                          */
4742                         ClearPagePrivate(page);
4743                         set_page_private(page, 0);
4744                         /* One for the page private */
4745                         put_page(page);
4746                 }
4747
4748                 if (mapped)
4749                         spin_unlock(&page->mapping->private_lock);
4750
4751                 /* One for when we allocated the page */
4752                 put_page(page);
4753         } while (index != 0);
4754 }
4755
4756 /*
4757  * Helper for releasing the extent buffer.
4758  */
4759 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4760 {
4761         btrfs_release_extent_buffer_page(eb);
4762         __free_extent_buffer(eb);
4763 }
4764
4765 static struct extent_buffer *
4766 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4767                       unsigned long len)
4768 {
4769         struct extent_buffer *eb = NULL;
4770
4771         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4772         eb->start = start;
4773         eb->len = len;
4774         eb->fs_info = fs_info;
4775         eb->bflags = 0;
4776         rwlock_init(&eb->lock);
4777         atomic_set(&eb->write_locks, 0);
4778         atomic_set(&eb->read_locks, 0);
4779         atomic_set(&eb->blocking_readers, 0);
4780         atomic_set(&eb->blocking_writers, 0);
4781         atomic_set(&eb->spinning_readers, 0);
4782         atomic_set(&eb->spinning_writers, 0);
4783         eb->lock_nested = 0;
4784         init_waitqueue_head(&eb->write_lock_wq);
4785         init_waitqueue_head(&eb->read_lock_wq);
4786
4787         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4788
4789         spin_lock_init(&eb->refs_lock);
4790         atomic_set(&eb->refs, 1);
4791         atomic_set(&eb->io_pages, 0);
4792
4793         /*
4794          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4795          */
4796         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4797                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4798         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4799
4800         return eb;
4801 }
4802
4803 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4804 {
4805         unsigned long i;
4806         struct page *p;
4807         struct extent_buffer *new;
4808         unsigned long num_pages = num_extent_pages(src->start, src->len);
4809
4810         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4811         if (new == NULL)
4812                 return NULL;
4813
4814         for (i = 0; i < num_pages; i++) {
4815                 p = alloc_page(GFP_NOFS);
4816                 if (!p) {
4817                         btrfs_release_extent_buffer(new);
4818                         return NULL;
4819                 }
4820                 attach_extent_buffer_page(new, p);
4821                 WARN_ON(PageDirty(p));
4822                 SetPageUptodate(p);
4823                 new->pages[i] = p;
4824                 copy_page(page_address(p), page_address(src->pages[i]));
4825         }
4826
4827         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4828         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4829
4830         return new;
4831 }
4832
4833 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4834                                                   u64 start, unsigned long len)
4835 {
4836         struct extent_buffer *eb;
4837         unsigned long num_pages;
4838         unsigned long i;
4839
4840         num_pages = num_extent_pages(start, len);
4841
4842         eb = __alloc_extent_buffer(fs_info, start, len);
4843         if (!eb)
4844                 return NULL;
4845
4846         for (i = 0; i < num_pages; i++) {
4847                 eb->pages[i] = alloc_page(GFP_NOFS);
4848                 if (!eb->pages[i])
4849                         goto err;
4850         }
4851         set_extent_buffer_uptodate(eb);
4852         btrfs_set_header_nritems(eb, 0);
4853         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4854
4855         return eb;
4856 err:
4857         for (; i > 0; i--)
4858                 __free_page(eb->pages[i - 1]);
4859         __free_extent_buffer(eb);
4860         return NULL;
4861 }
4862
4863 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4864                                                 u64 start)
4865 {
4866         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4867 }
4868
4869 static void check_buffer_tree_ref(struct extent_buffer *eb)
4870 {
4871         int refs;
4872         /*
4873          * The TREE_REF bit is first set when the extent_buffer is added
4874          * to the radix tree. It is also reset, if unset, when a new reference
4875          * is created by find_extent_buffer.
4876          *
4877          * It is only cleared in two cases: freeing the last non-tree
4878          * reference to the extent_buffer when its STALE bit is set or
4879          * calling releasepage when the tree reference is the only reference.
4880          *
4881          * In both cases, care is taken to ensure that the extent_buffer's
4882          * pages are not under io. However, releasepage can be concurrently
4883          * called with creating new references, which is prone to race
4884          * conditions between the calls to check_buffer_tree_ref in those
4885          * codepaths and clearing TREE_REF in try_release_extent_buffer.
4886          *
4887          * The actual lifetime of the extent_buffer in the radix tree is
4888          * adequately protected by the refcount, but the TREE_REF bit and
4889          * its corresponding reference are not. To protect against this
4890          * class of races, we call check_buffer_tree_ref from the codepaths
4891          * which trigger io after they set eb->io_pages. Note that once io is
4892          * initiated, TREE_REF can no longer be cleared, so that is the
4893          * moment at which any such race is best fixed.
4894          */
4895         refs = atomic_read(&eb->refs);
4896         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4897                 return;
4898
4899         spin_lock(&eb->refs_lock);
4900         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4901                 atomic_inc(&eb->refs);
4902         spin_unlock(&eb->refs_lock);
4903 }
4904
4905 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4906                 struct page *accessed)
4907 {
4908         unsigned long num_pages, i;
4909
4910         check_buffer_tree_ref(eb);
4911
4912         num_pages = num_extent_pages(eb->start, eb->len);
4913         for (i = 0; i < num_pages; i++) {
4914                 struct page *p = eb->pages[i];
4915
4916                 if (p != accessed)
4917                         mark_page_accessed(p);
4918         }
4919 }
4920
4921 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4922                                          u64 start)
4923 {
4924         struct extent_buffer *eb;
4925
4926         rcu_read_lock();
4927         eb = radix_tree_lookup(&fs_info->buffer_radix,
4928                                start >> PAGE_SHIFT);
4929         if (eb && atomic_inc_not_zero(&eb->refs)) {
4930                 rcu_read_unlock();
4931                 /*
4932                  * Lock our eb's refs_lock to avoid races with
4933                  * free_extent_buffer. When we get our eb it might be flagged
4934                  * with EXTENT_BUFFER_STALE and another task running
4935                  * free_extent_buffer might have seen that flag set,
4936                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4937                  * writeback flags not set) and it's still in the tree (flag
4938                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4939                  * of decrementing the extent buffer's reference count twice.
4940                  * So here we could race and increment the eb's reference count,
4941                  * clear its stale flag, mark it as dirty and drop our reference
4942                  * before the other task finishes executing free_extent_buffer,
4943                  * which would later result in an attempt to free an extent
4944                  * buffer that is dirty.
4945                  */
4946                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4947                         spin_lock(&eb->refs_lock);
4948                         spin_unlock(&eb->refs_lock);
4949                 }
4950                 mark_extent_buffer_accessed(eb, NULL);
4951                 return eb;
4952         }
4953         rcu_read_unlock();
4954
4955         return NULL;
4956 }
4957
4958 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4959 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4960                                         u64 start)
4961 {
4962         struct extent_buffer *eb, *exists = NULL;
4963         int ret;
4964
4965         eb = find_extent_buffer(fs_info, start);
4966         if (eb)
4967                 return eb;
4968         eb = alloc_dummy_extent_buffer(fs_info, start);
4969         if (!eb)
4970                 return ERR_PTR(-ENOMEM);
4971         eb->fs_info = fs_info;
4972 again:
4973         ret = radix_tree_preload(GFP_NOFS);
4974         if (ret) {
4975                 exists = ERR_PTR(ret);
4976                 goto free_eb;
4977         }
4978         spin_lock(&fs_info->buffer_lock);
4979         ret = radix_tree_insert(&fs_info->buffer_radix,
4980                                 start >> PAGE_SHIFT, eb);
4981         spin_unlock(&fs_info->buffer_lock);
4982         radix_tree_preload_end();
4983         if (ret == -EEXIST) {
4984                 exists = find_extent_buffer(fs_info, start);
4985                 if (exists)
4986                         goto free_eb;
4987                 else
4988                         goto again;
4989         }
4990         check_buffer_tree_ref(eb);
4991         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4992
4993         /*
4994          * We will free dummy extent buffer's if they come into
4995          * free_extent_buffer with a ref count of 2, but if we are using this we
4996          * want the buffers to stay in memory until we're done with them, so
4997          * bump the ref count again.
4998          */
4999         atomic_inc(&eb->refs);
5000         return eb;
5001 free_eb:
5002         btrfs_release_extent_buffer(eb);
5003         return exists;
5004 }
5005 #endif
5006
5007 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
5008                                           u64 start)
5009 {
5010         unsigned long len = fs_info->nodesize;
5011         unsigned long num_pages = num_extent_pages(start, len);
5012         unsigned long i;
5013         unsigned long index = start >> PAGE_SHIFT;
5014         struct extent_buffer *eb;
5015         struct extent_buffer *exists = NULL;
5016         struct page *p;
5017         struct address_space *mapping = fs_info->btree_inode->i_mapping;
5018         int uptodate = 1;
5019         int ret;
5020
5021         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
5022                 btrfs_err(fs_info, "bad tree block start %llu", start);
5023                 return ERR_PTR(-EINVAL);
5024         }
5025
5026         eb = find_extent_buffer(fs_info, start);
5027         if (eb)
5028                 return eb;
5029
5030         eb = __alloc_extent_buffer(fs_info, start, len);
5031         if (!eb)
5032                 return ERR_PTR(-ENOMEM);
5033
5034         for (i = 0; i < num_pages; i++, index++) {
5035                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
5036                 if (!p) {
5037                         exists = ERR_PTR(-ENOMEM);
5038                         goto free_eb;
5039                 }
5040
5041                 spin_lock(&mapping->private_lock);
5042                 if (PagePrivate(p)) {
5043                         /*
5044                          * We could have already allocated an eb for this page
5045                          * and attached one so lets see if we can get a ref on
5046                          * the existing eb, and if we can we know it's good and
5047                          * we can just return that one, else we know we can just
5048                          * overwrite page->private.
5049                          */
5050                         exists = (struct extent_buffer *)p->private;
5051                         if (atomic_inc_not_zero(&exists->refs)) {
5052                                 spin_unlock(&mapping->private_lock);
5053                                 unlock_page(p);
5054                                 put_page(p);
5055                                 mark_extent_buffer_accessed(exists, p);
5056                                 goto free_eb;
5057                         }
5058                         exists = NULL;
5059
5060                         /*
5061                          * Do this so attach doesn't complain and we need to
5062                          * drop the ref the old guy had.
5063                          */
5064                         ClearPagePrivate(p);
5065                         WARN_ON(PageDirty(p));
5066                         put_page(p);
5067                 }
5068                 attach_extent_buffer_page(eb, p);
5069                 spin_unlock(&mapping->private_lock);
5070                 WARN_ON(PageDirty(p));
5071                 eb->pages[i] = p;
5072                 if (!PageUptodate(p))
5073                         uptodate = 0;
5074
5075                 /*
5076                  * see below about how we avoid a nasty race with release page
5077                  * and why we unlock later
5078                  */
5079         }
5080         if (uptodate)
5081                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5082 again:
5083         ret = radix_tree_preload(GFP_NOFS);
5084         if (ret) {
5085                 exists = ERR_PTR(ret);
5086                 goto free_eb;
5087         }
5088
5089         spin_lock(&fs_info->buffer_lock);
5090         ret = radix_tree_insert(&fs_info->buffer_radix,
5091                                 start >> PAGE_SHIFT, eb);
5092         spin_unlock(&fs_info->buffer_lock);
5093         radix_tree_preload_end();
5094         if (ret == -EEXIST) {
5095                 exists = find_extent_buffer(fs_info, start);
5096                 if (exists)
5097                         goto free_eb;
5098                 else
5099                         goto again;
5100         }
5101         /* add one reference for the tree */
5102         check_buffer_tree_ref(eb);
5103         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5104
5105         /*
5106          * there is a race where release page may have
5107          * tried to find this extent buffer in the radix
5108          * but failed.  It will tell the VM it is safe to
5109          * reclaim the, and it will clear the page private bit.
5110          * We must make sure to set the page private bit properly
5111          * after the extent buffer is in the radix tree so
5112          * it doesn't get lost
5113          */
5114         SetPageChecked(eb->pages[0]);
5115         for (i = 1; i < num_pages; i++) {
5116                 p = eb->pages[i];
5117                 ClearPageChecked(p);
5118                 unlock_page(p);
5119         }
5120         unlock_page(eb->pages[0]);
5121         return eb;
5122
5123 free_eb:
5124         WARN_ON(!atomic_dec_and_test(&eb->refs));
5125         for (i = 0; i < num_pages; i++) {
5126                 if (eb->pages[i])
5127                         unlock_page(eb->pages[i]);
5128         }
5129
5130         btrfs_release_extent_buffer(eb);
5131         return exists;
5132 }
5133
5134 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5135 {
5136         struct extent_buffer *eb =
5137                         container_of(head, struct extent_buffer, rcu_head);
5138
5139         __free_extent_buffer(eb);
5140 }
5141
5142 /* Expects to have eb->eb_lock already held */
5143 static int release_extent_buffer(struct extent_buffer *eb)
5144 {
5145         WARN_ON(atomic_read(&eb->refs) == 0);
5146         if (atomic_dec_and_test(&eb->refs)) {
5147                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5148                         struct btrfs_fs_info *fs_info = eb->fs_info;
5149
5150                         spin_unlock(&eb->refs_lock);
5151
5152                         spin_lock(&fs_info->buffer_lock);
5153                         radix_tree_delete(&fs_info->buffer_radix,
5154                                           eb->start >> PAGE_SHIFT);
5155                         spin_unlock(&fs_info->buffer_lock);
5156                 } else {
5157                         spin_unlock(&eb->refs_lock);
5158                 }
5159
5160                 /* Should be safe to release our pages at this point */
5161                 btrfs_release_extent_buffer_page(eb);
5162 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5163                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5164                         __free_extent_buffer(eb);
5165                         return 1;
5166                 }
5167 #endif
5168                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5169                 return 1;
5170         }
5171         spin_unlock(&eb->refs_lock);
5172
5173         return 0;
5174 }
5175
5176 void free_extent_buffer(struct extent_buffer *eb)
5177 {
5178         int refs;
5179         int old;
5180         if (!eb)
5181                 return;
5182
5183         while (1) {
5184                 refs = atomic_read(&eb->refs);
5185                 if (refs <= 3)
5186                         break;
5187                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5188                 if (old == refs)
5189                         return;
5190         }
5191
5192         spin_lock(&eb->refs_lock);
5193         if (atomic_read(&eb->refs) == 2 &&
5194             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5195                 atomic_dec(&eb->refs);
5196
5197         if (atomic_read(&eb->refs) == 2 &&
5198             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5199             !extent_buffer_under_io(eb) &&
5200             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5201                 atomic_dec(&eb->refs);
5202
5203         /*
5204          * I know this is terrible, but it's temporary until we stop tracking
5205          * the uptodate bits and such for the extent buffers.
5206          */
5207         release_extent_buffer(eb);
5208 }
5209
5210 void free_extent_buffer_stale(struct extent_buffer *eb)
5211 {
5212         if (!eb)
5213                 return;
5214
5215         spin_lock(&eb->refs_lock);
5216         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5217
5218         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5219             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5220                 atomic_dec(&eb->refs);
5221         release_extent_buffer(eb);
5222 }
5223
5224 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5225 {
5226         unsigned long i;
5227         unsigned long num_pages;
5228         struct page *page;
5229
5230         num_pages = num_extent_pages(eb->start, eb->len);
5231
5232         for (i = 0; i < num_pages; i++) {
5233                 page = eb->pages[i];
5234                 if (!PageDirty(page))
5235                         continue;
5236
5237                 lock_page(page);
5238                 WARN_ON(!PagePrivate(page));
5239
5240                 clear_page_dirty_for_io(page);
5241                 spin_lock_irq(&page->mapping->tree_lock);
5242                 if (!PageDirty(page)) {
5243                         radix_tree_tag_clear(&page->mapping->page_tree,
5244                                                 page_index(page),
5245                                                 PAGECACHE_TAG_DIRTY);
5246                 }
5247                 spin_unlock_irq(&page->mapping->tree_lock);
5248                 ClearPageError(page);
5249                 unlock_page(page);
5250         }
5251         WARN_ON(atomic_read(&eb->refs) == 0);
5252 }
5253
5254 int set_extent_buffer_dirty(struct extent_buffer *eb)
5255 {
5256         unsigned long i;
5257         unsigned long num_pages;
5258         int was_dirty = 0;
5259
5260         check_buffer_tree_ref(eb);
5261
5262         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5263
5264         num_pages = num_extent_pages(eb->start, eb->len);
5265         WARN_ON(atomic_read(&eb->refs) == 0);
5266         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5267
5268         for (i = 0; i < num_pages; i++)
5269                 set_page_dirty(eb->pages[i]);
5270         return was_dirty;
5271 }
5272
5273 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5274 {
5275         unsigned long i;
5276         struct page *page;
5277         unsigned long num_pages;
5278
5279         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5280         num_pages = num_extent_pages(eb->start, eb->len);
5281         for (i = 0; i < num_pages; i++) {
5282                 page = eb->pages[i];
5283                 if (page)
5284                         ClearPageUptodate(page);
5285         }
5286 }
5287
5288 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5289 {
5290         unsigned long i;
5291         struct page *page;
5292         unsigned long num_pages;
5293
5294         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5295         num_pages = num_extent_pages(eb->start, eb->len);
5296         for (i = 0; i < num_pages; i++) {
5297                 page = eb->pages[i];
5298                 SetPageUptodate(page);
5299         }
5300 }
5301
5302 int extent_buffer_uptodate(struct extent_buffer *eb)
5303 {
5304         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5305 }
5306
5307 int read_extent_buffer_pages(struct extent_io_tree *tree,
5308                              struct extent_buffer *eb, int wait,
5309                              get_extent_t *get_extent, int mirror_num)
5310 {
5311         unsigned long i;
5312         struct page *page;
5313         int err;
5314         int ret = 0;
5315         int locked_pages = 0;
5316         int all_uptodate = 1;
5317         unsigned long num_pages;
5318         unsigned long num_reads = 0;
5319         struct bio *bio = NULL;
5320         unsigned long bio_flags = 0;
5321
5322         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5323                 return 0;
5324
5325         num_pages = num_extent_pages(eb->start, eb->len);
5326         for (i = 0; i < num_pages; i++) {
5327                 page = eb->pages[i];
5328                 if (wait == WAIT_NONE) {
5329                         if (!trylock_page(page))
5330                                 goto unlock_exit;
5331                 } else {
5332                         lock_page(page);
5333                 }
5334                 locked_pages++;
5335         }
5336         /*
5337          * We need to firstly lock all pages to make sure that
5338          * the uptodate bit of our pages won't be affected by
5339          * clear_extent_buffer_uptodate().
5340          */
5341         for (i = 0; i < num_pages; i++) {
5342                 page = eb->pages[i];
5343                 if (!PageUptodate(page)) {
5344                         num_reads++;
5345                         all_uptodate = 0;
5346                 }
5347         }
5348
5349         if (all_uptodate) {
5350                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5351                 goto unlock_exit;
5352         }
5353
5354         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5355         eb->read_mirror = 0;
5356         atomic_set(&eb->io_pages, num_reads);
5357         /*
5358          * It is possible for releasepage to clear the TREE_REF bit before we
5359          * set io_pages. See check_buffer_tree_ref for a more detailed comment.
5360          */
5361         check_buffer_tree_ref(eb);
5362         for (i = 0; i < num_pages; i++) {
5363                 page = eb->pages[i];
5364
5365                 if (!PageUptodate(page)) {
5366                         if (ret) {
5367                                 atomic_dec(&eb->io_pages);
5368                                 unlock_page(page);
5369                                 continue;
5370                         }
5371
5372                         ClearPageError(page);
5373                         err = __extent_read_full_page(tree, page,
5374                                                       get_extent, &bio,
5375                                                       mirror_num, &bio_flags,
5376                                                       REQ_META);
5377                         if (err) {
5378                                 ret = err;
5379                                 /*
5380                                  * We use &bio in above __extent_read_full_page,
5381                                  * so we ensure that if it returns error, the
5382                                  * current page fails to add itself to bio and
5383                                  * it's been unlocked.
5384                                  *
5385                                  * We must dec io_pages by ourselves.
5386                                  */
5387                                 atomic_dec(&eb->io_pages);
5388                         }
5389                 } else {
5390                         unlock_page(page);
5391                 }
5392         }
5393
5394         if (bio) {
5395                 err = submit_one_bio(bio, mirror_num, bio_flags);
5396                 if (err)
5397                         return err;
5398         }
5399
5400         if (ret || wait != WAIT_COMPLETE)
5401                 return ret;
5402
5403         for (i = 0; i < num_pages; i++) {
5404                 page = eb->pages[i];
5405                 wait_on_page_locked(page);
5406                 if (!PageUptodate(page))
5407                         ret = -EIO;
5408         }
5409
5410         return ret;
5411
5412 unlock_exit:
5413         while (locked_pages > 0) {
5414                 locked_pages--;
5415                 page = eb->pages[locked_pages];
5416                 unlock_page(page);
5417         }
5418         return ret;
5419 }
5420
5421 void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
5422                         unsigned long start, unsigned long len)
5423 {
5424         size_t cur;
5425         size_t offset;
5426         struct page *page;
5427         char *kaddr;
5428         char *dst = (char *)dstv;
5429         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5430         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5431
5432         if (start + len > eb->len) {
5433                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5434                      eb->start, eb->len, start, len);
5435                 memset(dst, 0, len);
5436                 return;
5437         }
5438
5439         offset = (start_offset + start) & (PAGE_SIZE - 1);
5440
5441         while (len > 0) {
5442                 page = eb->pages[i];
5443
5444                 cur = min(len, (PAGE_SIZE - offset));
5445                 kaddr = page_address(page);
5446                 memcpy(dst, kaddr + offset, cur);
5447
5448                 dst += cur;
5449                 len -= cur;
5450                 offset = 0;
5451                 i++;
5452         }
5453 }
5454
5455 int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
5456                                        void __user *dstv,
5457                                        unsigned long start, unsigned long len)
5458 {
5459         size_t cur;
5460         size_t offset;
5461         struct page *page;
5462         char *kaddr;
5463         char __user *dst = (char __user *)dstv;
5464         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5465         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5466         int ret = 0;
5467
5468         WARN_ON(start > eb->len);
5469         WARN_ON(start + len > eb->start + eb->len);
5470
5471         offset = (start_offset + start) & (PAGE_SIZE - 1);
5472
5473         while (len > 0) {
5474                 page = eb->pages[i];
5475
5476                 cur = min(len, (PAGE_SIZE - offset));
5477                 kaddr = page_address(page);
5478                 if (probe_user_write(dst, kaddr + offset, cur)) {
5479                         ret = -EFAULT;
5480                         break;
5481                 }
5482
5483                 dst += cur;
5484                 len -= cur;
5485                 offset = 0;
5486                 i++;
5487         }
5488
5489         return ret;
5490 }
5491
5492 /*
5493  * return 0 if the item is found within a page.
5494  * return 1 if the item spans two pages.
5495  * return -EINVAL otherwise.
5496  */
5497 int map_private_extent_buffer(const struct extent_buffer *eb,
5498                               unsigned long start, unsigned long min_len,
5499                               char **map, unsigned long *map_start,
5500                               unsigned long *map_len)
5501 {
5502         size_t offset = start & (PAGE_SIZE - 1);
5503         char *kaddr;
5504         struct page *p;
5505         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5506         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5507         unsigned long end_i = (start_offset + start + min_len - 1) >>
5508                 PAGE_SHIFT;
5509
5510         if (start + min_len > eb->len) {
5511                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5512                        eb->start, eb->len, start, min_len);
5513                 return -EINVAL;
5514         }
5515
5516         if (i != end_i)
5517                 return 1;
5518
5519         if (i == 0) {
5520                 offset = start_offset;
5521                 *map_start = 0;
5522         } else {
5523                 offset = 0;
5524                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5525         }
5526
5527         p = eb->pages[i];
5528         kaddr = page_address(p);
5529         *map = kaddr + offset;
5530         *map_len = PAGE_SIZE - offset;
5531         return 0;
5532 }
5533
5534 int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
5535                          unsigned long start, unsigned long len)
5536 {
5537         size_t cur;
5538         size_t offset;
5539         struct page *page;
5540         char *kaddr;
5541         char *ptr = (char *)ptrv;
5542         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5543         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5544         int ret = 0;
5545
5546         WARN_ON(start > eb->len);
5547         WARN_ON(start + len > eb->start + eb->len);
5548
5549         offset = (start_offset + start) & (PAGE_SIZE - 1);
5550
5551         while (len > 0) {
5552                 page = eb->pages[i];
5553
5554                 cur = min(len, (PAGE_SIZE - offset));
5555
5556                 kaddr = page_address(page);
5557                 ret = memcmp(ptr, kaddr + offset, cur);
5558                 if (ret)
5559                         break;
5560
5561                 ptr += cur;
5562                 len -= cur;
5563                 offset = 0;
5564                 i++;
5565         }
5566         return ret;
5567 }
5568
5569 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5570                 const void *srcv)
5571 {
5572         char *kaddr;
5573
5574         WARN_ON(!PageUptodate(eb->pages[0]));
5575         kaddr = page_address(eb->pages[0]);
5576         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5577                         BTRFS_FSID_SIZE);
5578 }
5579
5580 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv)
5581 {
5582         char *kaddr;
5583
5584         WARN_ON(!PageUptodate(eb->pages[0]));
5585         kaddr = page_address(eb->pages[0]);
5586         memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5587                         BTRFS_FSID_SIZE);
5588 }
5589
5590 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5591                          unsigned long start, unsigned long len)
5592 {
5593         size_t cur;
5594         size_t offset;
5595         struct page *page;
5596         char *kaddr;
5597         char *src = (char *)srcv;
5598         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5599         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5600
5601         WARN_ON(start > eb->len);
5602         WARN_ON(start + len > eb->start + eb->len);
5603
5604         offset = (start_offset + start) & (PAGE_SIZE - 1);
5605
5606         while (len > 0) {
5607                 page = eb->pages[i];
5608                 WARN_ON(!PageUptodate(page));
5609
5610                 cur = min(len, PAGE_SIZE - offset);
5611                 kaddr = page_address(page);
5612                 memcpy(kaddr + offset, src, cur);
5613
5614                 src += cur;
5615                 len -= cur;
5616                 offset = 0;
5617                 i++;
5618         }
5619 }
5620
5621 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5622                 unsigned long len)
5623 {
5624         size_t cur;
5625         size_t offset;
5626         struct page *page;
5627         char *kaddr;
5628         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5629         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5630
5631         WARN_ON(start > eb->len);
5632         WARN_ON(start + len > eb->start + eb->len);
5633
5634         offset = (start_offset + start) & (PAGE_SIZE - 1);
5635
5636         while (len > 0) {
5637                 page = eb->pages[i];
5638                 WARN_ON(!PageUptodate(page));
5639
5640                 cur = min(len, PAGE_SIZE - offset);
5641                 kaddr = page_address(page);
5642                 memset(kaddr + offset, 0, cur);
5643
5644                 len -= cur;
5645                 offset = 0;
5646                 i++;
5647         }
5648 }
5649
5650 void copy_extent_buffer_full(struct extent_buffer *dst,
5651                              struct extent_buffer *src)
5652 {
5653         int i;
5654         unsigned num_pages;
5655
5656         ASSERT(dst->len == src->len);
5657
5658         num_pages = num_extent_pages(dst->start, dst->len);
5659         for (i = 0; i < num_pages; i++)
5660                 copy_page(page_address(dst->pages[i]),
5661                                 page_address(src->pages[i]));
5662 }
5663
5664 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5665                         unsigned long dst_offset, unsigned long src_offset,
5666                         unsigned long len)
5667 {
5668         u64 dst_len = dst->len;
5669         size_t cur;
5670         size_t offset;
5671         struct page *page;
5672         char *kaddr;
5673         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5674         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5675
5676         WARN_ON(src->len != dst_len);
5677
5678         offset = (start_offset + dst_offset) &
5679                 (PAGE_SIZE - 1);
5680
5681         while (len > 0) {
5682                 page = dst->pages[i];
5683                 WARN_ON(!PageUptodate(page));
5684
5685                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5686
5687                 kaddr = page_address(page);
5688                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5689
5690                 src_offset += cur;
5691                 len -= cur;
5692                 offset = 0;
5693                 i++;
5694         }
5695 }
5696
5697 void le_bitmap_set(u8 *map, unsigned int start, int len)
5698 {
5699         u8 *p = map + BIT_BYTE(start);
5700         const unsigned int size = start + len;
5701         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5702         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5703
5704         while (len - bits_to_set >= 0) {
5705                 *p |= mask_to_set;
5706                 len -= bits_to_set;
5707                 bits_to_set = BITS_PER_BYTE;
5708                 mask_to_set = ~0;
5709                 p++;
5710         }
5711         if (len) {
5712                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5713                 *p |= mask_to_set;
5714         }
5715 }
5716
5717 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5718 {
5719         u8 *p = map + BIT_BYTE(start);
5720         const unsigned int size = start + len;
5721         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5722         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5723
5724         while (len - bits_to_clear >= 0) {
5725                 *p &= ~mask_to_clear;
5726                 len -= bits_to_clear;
5727                 bits_to_clear = BITS_PER_BYTE;
5728                 mask_to_clear = ~0;
5729                 p++;
5730         }
5731         if (len) {
5732                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5733                 *p &= ~mask_to_clear;
5734         }
5735 }
5736
5737 /*
5738  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5739  * given bit number
5740  * @eb: the extent buffer
5741  * @start: offset of the bitmap item in the extent buffer
5742  * @nr: bit number
5743  * @page_index: return index of the page in the extent buffer that contains the
5744  * given bit number
5745  * @page_offset: return offset into the page given by page_index
5746  *
5747  * This helper hides the ugliness of finding the byte in an extent buffer which
5748  * contains a given bit.
5749  */
5750 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5751                                     unsigned long start, unsigned long nr,
5752                                     unsigned long *page_index,
5753                                     size_t *page_offset)
5754 {
5755         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5756         size_t byte_offset = BIT_BYTE(nr);
5757         size_t offset;
5758
5759         /*
5760          * The byte we want is the offset of the extent buffer + the offset of
5761          * the bitmap item in the extent buffer + the offset of the byte in the
5762          * bitmap item.
5763          */
5764         offset = start_offset + start + byte_offset;
5765
5766         *page_index = offset >> PAGE_SHIFT;
5767         *page_offset = offset & (PAGE_SIZE - 1);
5768 }
5769
5770 /**
5771  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5772  * @eb: the extent buffer
5773  * @start: offset of the bitmap item in the extent buffer
5774  * @nr: bit number to test
5775  */
5776 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5777                            unsigned long nr)
5778 {
5779         u8 *kaddr;
5780         struct page *page;
5781         unsigned long i;
5782         size_t offset;
5783
5784         eb_bitmap_offset(eb, start, nr, &i, &offset);
5785         page = eb->pages[i];
5786         WARN_ON(!PageUptodate(page));
5787         kaddr = page_address(page);
5788         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5789 }
5790
5791 /**
5792  * extent_buffer_bitmap_set - set an area of a bitmap
5793  * @eb: the extent buffer
5794  * @start: offset of the bitmap item in the extent buffer
5795  * @pos: bit number of the first bit
5796  * @len: number of bits to set
5797  */
5798 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5799                               unsigned long pos, unsigned long len)
5800 {
5801         u8 *kaddr;
5802         struct page *page;
5803         unsigned long i;
5804         size_t offset;
5805         const unsigned int size = pos + len;
5806         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5807         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5808
5809         eb_bitmap_offset(eb, start, pos, &i, &offset);
5810         page = eb->pages[i];
5811         WARN_ON(!PageUptodate(page));
5812         kaddr = page_address(page);
5813
5814         while (len >= bits_to_set) {
5815                 kaddr[offset] |= mask_to_set;
5816                 len -= bits_to_set;
5817                 bits_to_set = BITS_PER_BYTE;
5818                 mask_to_set = ~0;
5819                 if (++offset >= PAGE_SIZE && len > 0) {
5820                         offset = 0;
5821                         page = eb->pages[++i];
5822                         WARN_ON(!PageUptodate(page));
5823                         kaddr = page_address(page);
5824                 }
5825         }
5826         if (len) {
5827                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5828                 kaddr[offset] |= mask_to_set;
5829         }
5830 }
5831
5832
5833 /**
5834  * extent_buffer_bitmap_clear - clear an area of a bitmap
5835  * @eb: the extent buffer
5836  * @start: offset of the bitmap item in the extent buffer
5837  * @pos: bit number of the first bit
5838  * @len: number of bits to clear
5839  */
5840 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5841                                 unsigned long pos, unsigned long len)
5842 {
5843         u8 *kaddr;
5844         struct page *page;
5845         unsigned long i;
5846         size_t offset;
5847         const unsigned int size = pos + len;
5848         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5849         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5850
5851         eb_bitmap_offset(eb, start, pos, &i, &offset);
5852         page = eb->pages[i];
5853         WARN_ON(!PageUptodate(page));
5854         kaddr = page_address(page);
5855
5856         while (len >= bits_to_clear) {
5857                 kaddr[offset] &= ~mask_to_clear;
5858                 len -= bits_to_clear;
5859                 bits_to_clear = BITS_PER_BYTE;
5860                 mask_to_clear = ~0;
5861                 if (++offset >= PAGE_SIZE && len > 0) {
5862                         offset = 0;
5863                         page = eb->pages[++i];
5864                         WARN_ON(!PageUptodate(page));
5865                         kaddr = page_address(page);
5866                 }
5867         }
5868         if (len) {
5869                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5870                 kaddr[offset] &= ~mask_to_clear;
5871         }
5872 }
5873
5874 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5875 {
5876         unsigned long distance = (src > dst) ? src - dst : dst - src;
5877         return distance < len;
5878 }
5879
5880 static void copy_pages(struct page *dst_page, struct page *src_page,
5881                        unsigned long dst_off, unsigned long src_off,
5882                        unsigned long len)
5883 {
5884         char *dst_kaddr = page_address(dst_page);
5885         char *src_kaddr;
5886         int must_memmove = 0;
5887
5888         if (dst_page != src_page) {
5889                 src_kaddr = page_address(src_page);
5890         } else {
5891                 src_kaddr = dst_kaddr;
5892                 if (areas_overlap(src_off, dst_off, len))
5893                         must_memmove = 1;
5894         }
5895
5896         if (must_memmove)
5897                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5898         else
5899                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5900 }
5901
5902 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5903                            unsigned long src_offset, unsigned long len)
5904 {
5905         struct btrfs_fs_info *fs_info = dst->fs_info;
5906         size_t cur;
5907         size_t dst_off_in_page;
5908         size_t src_off_in_page;
5909         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5910         unsigned long dst_i;
5911         unsigned long src_i;
5912
5913         if (src_offset + len > dst->len) {
5914                 btrfs_err(fs_info,
5915                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5916                          src_offset, len, dst->len);
5917                 BUG_ON(1);
5918         }
5919         if (dst_offset + len > dst->len) {
5920                 btrfs_err(fs_info,
5921                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5922                          dst_offset, len, dst->len);
5923                 BUG_ON(1);
5924         }
5925
5926         while (len > 0) {
5927                 dst_off_in_page = (start_offset + dst_offset) &
5928                         (PAGE_SIZE - 1);
5929                 src_off_in_page = (start_offset + src_offset) &
5930                         (PAGE_SIZE - 1);
5931
5932                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5933                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5934
5935                 cur = min(len, (unsigned long)(PAGE_SIZE -
5936                                                src_off_in_page));
5937                 cur = min_t(unsigned long, cur,
5938                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5939
5940                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5941                            dst_off_in_page, src_off_in_page, cur);
5942
5943                 src_offset += cur;
5944                 dst_offset += cur;
5945                 len -= cur;
5946         }
5947 }
5948
5949 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5950                            unsigned long src_offset, unsigned long len)
5951 {
5952         struct btrfs_fs_info *fs_info = dst->fs_info;
5953         size_t cur;
5954         size_t dst_off_in_page;
5955         size_t src_off_in_page;
5956         unsigned long dst_end = dst_offset + len - 1;
5957         unsigned long src_end = src_offset + len - 1;
5958         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5959         unsigned long dst_i;
5960         unsigned long src_i;
5961
5962         if (src_offset + len > dst->len) {
5963                 btrfs_err(fs_info,
5964                           "memmove bogus src_offset %lu move len %lu len %lu",
5965                           src_offset, len, dst->len);
5966                 BUG_ON(1);
5967         }
5968         if (dst_offset + len > dst->len) {
5969                 btrfs_err(fs_info,
5970                           "memmove bogus dst_offset %lu move len %lu len %lu",
5971                           dst_offset, len, dst->len);
5972                 BUG_ON(1);
5973         }
5974         if (dst_offset < src_offset) {
5975                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5976                 return;
5977         }
5978         while (len > 0) {
5979                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5980                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5981
5982                 dst_off_in_page = (start_offset + dst_end) &
5983                         (PAGE_SIZE - 1);
5984                 src_off_in_page = (start_offset + src_end) &
5985                         (PAGE_SIZE - 1);
5986
5987                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5988                 cur = min(cur, dst_off_in_page + 1);
5989                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5990                            dst_off_in_page - cur + 1,
5991                            src_off_in_page - cur + 1, cur);
5992
5993                 dst_end -= cur;
5994                 src_end -= cur;
5995                 len -= cur;
5996         }
5997 }
5998
5999 int try_release_extent_buffer(struct page *page)
6000 {
6001         struct extent_buffer *eb;
6002
6003         /*
6004          * We need to make sure nobody is attaching this page to an eb right
6005          * now.
6006          */
6007         spin_lock(&page->mapping->private_lock);
6008         if (!PagePrivate(page)) {
6009                 spin_unlock(&page->mapping->private_lock);
6010                 return 1;
6011         }
6012
6013         eb = (struct extent_buffer *)page->private;
6014         BUG_ON(!eb);
6015
6016         /*
6017          * This is a little awful but should be ok, we need to make sure that
6018          * the eb doesn't disappear out from under us while we're looking at
6019          * this page.
6020          */
6021         spin_lock(&eb->refs_lock);
6022         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
6023                 spin_unlock(&eb->refs_lock);
6024                 spin_unlock(&page->mapping->private_lock);
6025                 return 0;
6026         }
6027         spin_unlock(&page->mapping->private_lock);
6028
6029         /*
6030          * If tree ref isn't set then we know the ref on this eb is a real ref,
6031          * so just return, this page will likely be freed soon anyway.
6032          */
6033         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
6034                 spin_unlock(&eb->refs_lock);
6035                 return 0;
6036         }
6037
6038         return release_extent_buffer(eb);
6039 }