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
 114         /* tells writepage not to lock the state bits for this range
 115          * it still does the unlocking
 116          */
 117         unsigned int extent_locked:1;
 118
 119         /* tells the submit_bio code to use REQ_SYNC */
 120         unsigned int sync_io:1;
 121 };
 122
 123 static void add_extent_changeset(struct extent_state *state, unsigned bits,
 124                                  struct extent_changeset *changeset,
 125                                  int set)
 126 {
 127         int ret;
 128
 129         if (!changeset)
 130                 return;
 131         if (set && (state->state & bits) == bits)
 132                 return;
 133         if (!set && (state->state & bits) == 0)
 134                 return;
 135         changeset->bytes_changed += state->end - state->start + 1;
 136         ret = ulist_add(&changeset->range_changed, state->start, state->end,
 137                         GFP_ATOMIC);
 138         /* ENOMEM */
 139         BUG_ON(ret < 0);
 140 }
 141
 142 static noinline void flush_write_bio(void *data);
 143 static inline struct btrfs_fs_info *
 144 tree_fs_info(struct extent_io_tree *tree)
 145 {
 146         if (tree->ops)
 147                 return tree->ops->tree_fs_info(tree->private_data);
 148         return NULL;
 149 }
 150
 151 int __init extent_io_init(void)
 152 {
 153         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 154                         sizeof(struct extent_state), 0,
 155                         SLAB_MEM_SPREAD, NULL);
 156         if (!extent_state_cache)
 157                 return -ENOMEM;
 158
 159         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 160                         sizeof(struct extent_buffer), 0,
 161                         SLAB_MEM_SPREAD, NULL);
 162         if (!extent_buffer_cache)
 163                 goto free_state_cache;
 164
 165         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 166                                      offsetof(struct btrfs_io_bio, bio),
 167                                      BIOSET_NEED_BVECS);
 168         if (!btrfs_bioset)
 169                 goto free_buffer_cache;
 170
 171         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 172                 goto free_bioset;
 173
 174         return 0;
 175
 176 free_bioset:
 177         bioset_free(btrfs_bioset);
 178         btrfs_bioset = NULL;
 179
 180 free_buffer_cache:
 181         kmem_cache_destroy(extent_buffer_cache);
 182         extent_buffer_cache = NULL;
 183
 184 free_state_cache:
 185         kmem_cache_destroy(extent_state_cache);
 186         extent_state_cache = NULL;
 187         return -ENOMEM;
 188 }
 189
 190 void extent_io_exit(void)
 191 {
 192         btrfs_leak_debug_check();
 193
 194         /*
 195          * Make sure all delayed rcu free are flushed before we
 196          * destroy caches.
 197          */
 198         rcu_barrier();
 199         kmem_cache_destroy(extent_state_cache);
 200         kmem_cache_destroy(extent_buffer_cache);
 201         if (btrfs_bioset)
 202                 bioset_free(btrfs_bioset);
 203 }
 204
 205 void extent_io_tree_init(struct extent_io_tree *tree,
 206                          void *private_data)
 207 {
 208         tree->state = RB_ROOT;
 209         tree->ops = NULL;
 210         tree->dirty_bytes = 0;
 211         spin_lock_init(&tree->lock);
 212         tree->private_data = private_data;
 213 }
 214
 215 static struct extent_state *alloc_extent_state(gfp_t mask)
 216 {
 217         struct extent_state *state;
 218
 219         /*
 220          * The given mask might be not appropriate for the slab allocator,
 221          * drop the unsupported bits
 222          */
 223         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
 224         state = kmem_cache_alloc(extent_state_cache, mask);
 225         if (!state)
 226                 return state;
 227         state->state = 0;
 228         state->failrec = NULL;
 229         RB_CLEAR_NODE(&state->rb_node);
 230         btrfs_leak_debug_add(&state->leak_list, &states);
 231         refcount_set(&state->refs, 1);
 232         init_waitqueue_head(&state->wq);
 233         trace_alloc_extent_state(state, mask, _RET_IP_);
 234         return state;
 235 }
 236
 237 void free_extent_state(struct extent_state *state)
 238 {
 239         if (!state)
 240                 return;
 241         if (refcount_dec_and_test(&state->refs)) {
 242                 WARN_ON(extent_state_in_tree(state));
 243                 btrfs_leak_debug_del(&state->leak_list);
 244                 trace_free_extent_state(state, _RET_IP_);
 245                 kmem_cache_free(extent_state_cache, state);
 246         }
 247 }
 248
 249 static struct rb_node *tree_insert(struct rb_root *root,
 250                                    struct rb_node *search_start,
 251                                    u64 offset,
 252                                    struct rb_node *node,
 253                                    struct rb_node ***p_in,
 254                                    struct rb_node **parent_in)
 255 {
 256         struct rb_node **p;
 257         struct rb_node *parent = NULL;
 258         struct tree_entry *entry;
 259
 260         if (p_in && parent_in) {
 261                 p = *p_in;
 262                 parent = *parent_in;
 263                 goto do_insert;
 264         }
 265
 266         p = search_start ? &search_start : &root->rb_node;
 267         while (*p) {
 268                 parent = *p;
 269                 entry = rb_entry(parent, struct tree_entry, rb_node);
 270
 271                 if (offset < entry->start)
 272                         p = &(*p)->rb_left;
 273                 else if (offset > entry->end)
 274                         p = &(*p)->rb_right;
 275                 else
 276                         return parent;
 277         }
 278
 279 do_insert:
 280         rb_link_node(node, parent, p);
 281         rb_insert_color(node, root);
 282         return NULL;
 283 }
 284
 285 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 286                                       struct rb_node **prev_ret,
 287                                       struct rb_node **next_ret,
 288                                       struct rb_node ***p_ret,
 289                                       struct rb_node **parent_ret)
 290 {
 291         struct rb_root *root = &tree->state;
 292         struct rb_node **n = &root->rb_node;
 293         struct rb_node *prev = NULL;
 294         struct rb_node *orig_prev = NULL;
 295         struct tree_entry *entry;
 296         struct tree_entry *prev_entry = NULL;
 297
 298         while (*n) {
 299                 prev = *n;
 300                 entry = rb_entry(prev, struct tree_entry, rb_node);
 301                 prev_entry = entry;
 302
 303                 if (offset < entry->start)
 304                         n = &(*n)->rb_left;
 305                 else if (offset > entry->end)
 306                         n = &(*n)->rb_right;
 307                 else
 308                         return *n;
 309         }
 310
 311         if (p_ret)
 312                 *p_ret = n;
 313         if (parent_ret)
 314                 *parent_ret = prev;
 315
 316         if (prev_ret) {
 317                 orig_prev = prev;
 318                 while (prev && offset > prev_entry->end) {
 319                         prev = rb_next(prev);
 320                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 321                 }
 322                 *prev_ret = prev;
 323                 prev = orig_prev;
 324         }
 325
 326         if (next_ret) {
 327                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 328                 while (prev && offset < prev_entry->start) {
 329                         prev = rb_prev(prev);
 330                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 331                 }
 332                 *next_ret = prev;
 333         }
 334         return NULL;
 335 }
 336
 337 static inline struct rb_node *
 338 tree_search_for_insert(struct extent_io_tree *tree,
 339                        u64 offset,
 340                        struct rb_node ***p_ret,
 341                        struct rb_node **parent_ret)
 342 {
 343         struct rb_node *prev = NULL;
 344         struct rb_node *ret;
 345
 346         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 347         if (!ret)
 348                 return prev;
 349         return ret;
 350 }
 351
 352 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 353                                           u64 offset)
 354 {
 355         return tree_search_for_insert(tree, offset, NULL, NULL);
 356 }
 357
 358 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 359                      struct extent_state *other)
 360 {
 361         if (tree->ops && tree->ops->merge_extent_hook)
 362                 tree->ops->merge_extent_hook(tree->private_data, new, other);
 363 }
 364
 365 /*
 366  * utility function to look for merge candidates inside a given range.
 367  * Any extents with matching state are merged together into a single
 368  * extent in the tree.  Extents with EXTENT_IO in their state field
 369  * are not merged because the end_io handlers need to be able to do
 370  * operations on them without sleeping (or doing allocations/splits).
 371  *
 372  * This should be called with the tree lock held.
 373  */
 374 static void merge_state(struct extent_io_tree *tree,
 375                         struct extent_state *state)
 376 {
 377         struct extent_state *other;
 378         struct rb_node *other_node;
 379
 380         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 381                 return;
 382
 383         other_node = rb_prev(&state->rb_node);
 384         if (other_node) {
 385                 other = rb_entry(other_node, struct extent_state, rb_node);
 386                 if (other->end == state->start - 1 &&
 387                     other->state == state->state) {
 388                         merge_cb(tree, state, other);
 389                         state->start = other->start;
 390                         rb_erase(&other->rb_node, &tree->state);
 391                         RB_CLEAR_NODE(&other->rb_node);
 392                         free_extent_state(other);
 393                 }
 394         }
 395         other_node = rb_next(&state->rb_node);
 396         if (other_node) {
 397                 other = rb_entry(other_node, struct extent_state, rb_node);
 398                 if (other->start == state->end + 1 &&
 399                     other->state == state->state) {
 400                         merge_cb(tree, state, other);
 401                         state->end = other->end;
 402                         rb_erase(&other->rb_node, &tree->state);
 403                         RB_CLEAR_NODE(&other->rb_node);
 404                         free_extent_state(other);
 405                 }
 406         }
 407 }
 408
 409 static void set_state_cb(struct extent_io_tree *tree,
 410                          struct extent_state *state, unsigned *bits)
 411 {
 412         if (tree->ops && tree->ops->set_bit_hook)
 413                 tree->ops->set_bit_hook(tree->private_data, state, bits);
 414 }
 415
 416 static void clear_state_cb(struct extent_io_tree *tree,
 417                            struct extent_state *state, unsigned *bits)
 418 {
 419         if (tree->ops && tree->ops->clear_bit_hook)
 420                 tree->ops->clear_bit_hook(tree->private_data, state, bits);
 421 }
 422
 423 static void set_state_bits(struct extent_io_tree *tree,
 424                            struct extent_state *state, unsigned *bits,
 425                            struct extent_changeset *changeset);
 426
 427 /*
 428  * insert an extent_state struct into the tree.  'bits' are set on the
 429  * struct before it is inserted.
 430  *
 431  * This may return -EEXIST if the extent is already there, in which case the
 432  * state struct is freed.
 433  *
 434  * The tree lock is not taken internally.  This is a utility function and
 435  * probably isn't what you want to call (see set/clear_extent_bit).
 436  */
 437 static int insert_state(struct extent_io_tree *tree,
 438                         struct extent_state *state, u64 start, u64 end,
 439                         struct rb_node ***p,
 440                         struct rb_node **parent,
 441                         unsigned *bits, struct extent_changeset *changeset)
 442 {
 443         struct rb_node *node;
 444
 445         if (end < start)
 446                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 447                        end, start);
 448         state->start = start;
 449         state->end = end;
 450
 451         set_state_bits(tree, state, bits, changeset);
 452
 453         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 454         if (node) {
 455                 struct extent_state *found;
 456                 found = rb_entry(node, struct extent_state, rb_node);
 457                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
 458                        found->start, found->end, start, end);
 459                 return -EEXIST;
 460         }
 461         merge_state(tree, state);
 462         return 0;
 463 }
 464
 465 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 466                      u64 split)
 467 {
 468         if (tree->ops && tree->ops->split_extent_hook)
 469                 tree->ops->split_extent_hook(tree->private_data, orig, split);
 470 }
 471
 472 /*
 473  * split a given extent state struct in two, inserting the preallocated
 474  * struct 'prealloc' as the newly created second half.  'split' indicates an
 475  * offset inside 'orig' where it should be split.
 476  *
 477  * Before calling,
 478  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 479  * are two extent state structs in the tree:
 480  * prealloc: [orig->start, split - 1]
 481  * orig: [ split, orig->end ]
 482  *
 483  * The tree locks are not taken by this function. They need to be held
 484  * by the caller.
 485  */
 486 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 487                        struct extent_state *prealloc, u64 split)
 488 {
 489         struct rb_node *node;
 490
 491         split_cb(tree, orig, split);
 492
 493         prealloc->start = orig->start;
 494         prealloc->end = split - 1;
 495         prealloc->state = orig->state;
 496         orig->start = split;
 497
 498         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 499                            &prealloc->rb_node, NULL, NULL);
 500         if (node) {
 501                 free_extent_state(prealloc);
 502                 return -EEXIST;
 503         }
 504         return 0;
 505 }
 506
 507 static struct extent_state *next_state(struct extent_state *state)
 508 {
 509         struct rb_node *next = rb_next(&state->rb_node);
 510         if (next)
 511                 return rb_entry(next, struct extent_state, rb_node);
 512         else
 513                 return NULL;
 514 }
 515
 516 /*
 517  * utility function to clear some bits in an extent state struct.
 518  * it will optionally wake up any one waiting on this state (wake == 1).
 519  *
 520  * If no bits are set on the state struct after clearing things, the
 521  * struct is freed and removed from the tree
 522  */
 523 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 524                                             struct extent_state *state,
 525                                             unsigned *bits, int wake,
 526                                             struct extent_changeset *changeset)
 527 {
 528         struct extent_state *next;
 529         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 530
 531         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 532                 u64 range = state->end - state->start + 1;
 533                 WARN_ON(range > tree->dirty_bytes);
 534                 tree->dirty_bytes -= range;
 535         }
 536         clear_state_cb(tree, state, bits);
 537         add_extent_changeset(state, bits_to_clear, changeset, 0);
 538         state->state &= ~bits_to_clear;
 539         if (wake)
 540                 wake_up(&state->wq);
 541         if (state->state == 0) {
 542                 next = next_state(state);
 543                 if (extent_state_in_tree(state)) {
 544                         rb_erase(&state->rb_node, &tree->state);
 545                         RB_CLEAR_NODE(&state->rb_node);
 546                         free_extent_state(state);
 547                 } else {
 548                         WARN_ON(1);
 549                 }
 550         } else {
 551                 merge_state(tree, state);
 552                 next = next_state(state);
 553         }
 554         return next;
 555 }
 556
 557 static struct extent_state *
 558 alloc_extent_state_atomic(struct extent_state *prealloc)
 559 {
 560         if (!prealloc)
 561                 prealloc = alloc_extent_state(GFP_ATOMIC);
 562
 563         return prealloc;
 564 }
 565
 566 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 567 {
 568         btrfs_panic(tree_fs_info(tree), err,
 569                     "Locking error: Extent tree was modified by another thread while locked.");
 570 }
 571
 572 /*
 573  * clear some bits on a range in the tree.  This may require splitting
 574  * or inserting elements in the tree, so the gfp mask is used to
 575  * indicate which allocations or sleeping are allowed.
 576  *
 577  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 578  * the given range from the tree regardless of state (ie for truncate).
 579  *
 580  * the range [start, end] is inclusive.
 581  *
 582  * This takes the tree lock, and returns 0 on success and < 0 on error.
 583  */
 584 int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 585                               unsigned bits, int wake, int delete,
 586                               struct extent_state **cached_state,
 587                               gfp_t mask, struct extent_changeset *changeset)
 588 {
 589         struct extent_state *state;
 590         struct extent_state *cached;
 591         struct extent_state *prealloc = NULL;
 592         struct rb_node *node;
 593         u64 last_end;
 594         int err;
 595         int clear = 0;
 596
 597         btrfs_debug_check_extent_io_range(tree, start, end);
 598
 599         if (bits & EXTENT_DELALLOC)
 600                 bits |= EXTENT_NORESERVE;
 601
 602         if (delete)
 603                 bits |= ~EXTENT_CTLBITS;
 604         bits |= EXTENT_FIRST_DELALLOC;
 605
 606         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 607                 clear = 1;
 608 again:
 609         if (!prealloc && gfpflags_allow_blocking(mask)) {
 610                 /*
 611                  * Don't care for allocation failure here because we might end
 612                  * up not needing the pre-allocated extent state at all, which
 613                  * is the case if we only have in the tree extent states that
 614                  * cover our input range and don't cover too any other range.
 615                  * If we end up needing a new extent state we allocate it later.
 616                  */
 617                 prealloc = alloc_extent_state(mask);
 618         }
 619
 620         spin_lock(&tree->lock);
 621         if (cached_state) {
 622                 cached = *cached_state;
 623
 624                 if (clear) {
 625                         *cached_state = NULL;
 626                         cached_state = NULL;
 627                 }
 628
 629                 if (cached && extent_state_in_tree(cached) &&
 630                     cached->start <= start && cached->end > start) {
 631                         if (clear)
 632                                 refcount_dec(&cached->refs);
 633                         state = cached;
 634                         goto hit_next;
 635                 }
 636                 if (clear)
 637                         free_extent_state(cached);
 638         }
 639         /*
 640          * this search will find the extents that end after
 641          * our range starts
 642          */
 643         node = tree_search(tree, start);
 644         if (!node)
 645                 goto out;
 646         state = rb_entry(node, struct extent_state, rb_node);
 647 hit_next:
 648         if (state->start > end)
 649                 goto out;
 650         WARN_ON(state->end < start);
 651         last_end = state->end;
 652
 653         /* the state doesn't have the wanted bits, go ahead */
 654         if (!(state->state & bits)) {
 655                 state = next_state(state);
 656                 goto next;
 657         }
 658
 659         /*
 660          *     | ---- desired range ---- |
 661          *  | state | or
 662          *  | ------------- state -------------- |
 663          *
 664          * We need to split the extent we found, and may flip
 665          * bits on second half.
 666          *
 667          * If the extent we found extends past our range, we
 668          * just split and search again.  It'll get split again
 669          * the next time though.
 670          *
 671          * If the extent we found is inside our range, we clear
 672          * the desired bit on it.
 673          */
 674
 675         if (state->start < start) {
 676                 prealloc = alloc_extent_state_atomic(prealloc);
 677                 BUG_ON(!prealloc);
 678                 err = split_state(tree, state, prealloc, start);
 679                 if (err)
 680                         extent_io_tree_panic(tree, err);
 681
 682                 prealloc = NULL;
 683                 if (err)
 684                         goto out;
 685                 if (state->end <= end) {
 686                         state = clear_state_bit(tree, state, &bits, wake,
 687                                                 changeset);
 688                         goto next;
 689                 }
 690                 goto search_again;
 691         }
 692         /*
 693          * | ---- desired range ---- |
 694          *                        | state |
 695          * We need to split the extent, and clear the bit
 696          * on the first half
 697          */
 698         if (state->start <= end && state->end > end) {
 699                 prealloc = alloc_extent_state_atomic(prealloc);
 700                 BUG_ON(!prealloc);
 701                 err = split_state(tree, state, prealloc, end + 1);
 702                 if (err)
 703                         extent_io_tree_panic(tree, err);
 704
 705                 if (wake)
 706                         wake_up(&state->wq);
 707
 708                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 709
 710                 prealloc = NULL;
 711                 goto out;
 712         }
 713
 714         state = clear_state_bit(tree, state, &bits, wake, changeset);
 715 next:
 716         if (last_end == (u64)-1)
 717                 goto out;
 718         start = last_end + 1;
 719         if (start <= end && state && !need_resched())
 720                 goto hit_next;
 721
 722 search_again:
 723         if (start > end)
 724                 goto out;
 725         spin_unlock(&tree->lock);
 726         if (gfpflags_allow_blocking(mask))
 727                 cond_resched();
 728         goto again;
 729
 730 out:
 731         spin_unlock(&tree->lock);
 732         if (prealloc)
 733                 free_extent_state(prealloc);
 734
 735         return 0;
 736
 737 }
 738
 739 static void wait_on_state(struct extent_io_tree *tree,
 740                           struct extent_state *state)
 741                 __releases(tree->lock)
 742                 __acquires(tree->lock)
 743 {
 744         DEFINE_WAIT(wait);
 745         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 746         spin_unlock(&tree->lock);
 747         schedule();
 748         spin_lock(&tree->lock);
 749         finish_wait(&state->wq, &wait);
 750 }
 751
 752 /*
 753  * waits for one or more bits to clear on a range in the state tree.
 754  * The range [start, end] is inclusive.
 755  * The tree lock is taken by this function
 756  */
 757 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 758                             unsigned long bits)
 759 {
 760         struct extent_state *state;
 761         struct rb_node *node;
 762
 763         btrfs_debug_check_extent_io_range(tree, start, end);
 764
 765         spin_lock(&tree->lock);
 766 again:
 767         while (1) {
 768                 /*
 769                  * this search will find all the extents that end after
 770                  * our range starts
 771                  */
 772                 node = tree_search(tree, start);
 773 process_node:
 774                 if (!node)
 775                         break;
 776
 777                 state = rb_entry(node, struct extent_state, rb_node);
 778
 779                 if (state->start > end)
 780                         goto out;
 781
 782                 if (state->state & bits) {
 783                         start = state->start;
 784                         refcount_inc(&state->refs);
 785                         wait_on_state(tree, state);
 786                         free_extent_state(state);
 787                         goto again;
 788                 }
 789                 start = state->end + 1;
 790
 791                 if (start > end)
 792                         break;
 793
 794                 if (!cond_resched_lock(&tree->lock)) {
 795                         node = rb_next(node);
 796                         goto process_node;
 797                 }
 798         }
 799 out:
 800         spin_unlock(&tree->lock);
 801 }
 802
 803 static void set_state_bits(struct extent_io_tree *tree,
 804                            struct extent_state *state,
 805                            unsigned *bits, struct extent_changeset *changeset)
 806 {
 807         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 808
 809         set_state_cb(tree, state, bits);
 810         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 811                 u64 range = state->end - state->start + 1;
 812                 tree->dirty_bytes += range;
 813         }
 814         add_extent_changeset(state, bits_to_set, changeset, 1);
 815         state->state |= bits_to_set;
 816 }
 817
 818 static void cache_state_if_flags(struct extent_state *state,
 819                                  struct extent_state **cached_ptr,
 820                                  unsigned flags)
 821 {
 822         if (cached_ptr && !(*cached_ptr)) {
 823                 if (!flags || (state->state & flags)) {
 824                         *cached_ptr = state;
 825                         refcount_inc(&state->refs);
 826                 }
 827         }
 828 }
 829
 830 static void cache_state(struct extent_state *state,
 831                         struct extent_state **cached_ptr)
 832 {
 833         return cache_state_if_flags(state, cached_ptr,
 834                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 835 }
 836
 837 /*
 838  * set some bits on a range in the tree.  This may require allocations or
 839  * sleeping, so the gfp mask is used to indicate what is allowed.
 840  *
 841  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 842  * part of the range already has the desired bits set.  The start of the
 843  * existing range is returned in failed_start in this case.
 844  *
 845  * [start, end] is inclusive This takes the tree lock.
 846  */
 847
 848 static int __must_check
 849 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 850                  unsigned bits, unsigned exclusive_bits,
 851                  u64 *failed_start, struct extent_state **cached_state,
 852                  gfp_t mask, struct extent_changeset *changeset)
 853 {
 854         struct extent_state *state;
 855         struct extent_state *prealloc = NULL;
 856         struct rb_node *node;
 857         struct rb_node **p;
 858         struct rb_node *parent;
 859         int err = 0;
 860         u64 last_start;
 861         u64 last_end;
 862
 863         btrfs_debug_check_extent_io_range(tree, start, end);
 864
 865         bits |= EXTENT_FIRST_DELALLOC;
 866 again:
 867         if (!prealloc && gfpflags_allow_blocking(mask)) {
 868                 /*
 869                  * Don't care for allocation failure here because we might end
 870                  * up not needing the pre-allocated extent state at all, which
 871                  * is the case if we only have in the tree extent states that
 872                  * cover our input range and don't cover too any other range.
 873                  * If we end up needing a new extent state we allocate it later.
 874                  */
 875                 prealloc = alloc_extent_state(mask);
 876         }
 877
 878         spin_lock(&tree->lock);
 879         if (cached_state && *cached_state) {
 880                 state = *cached_state;
 881                 if (state->start <= start && state->end > start &&
 882                     extent_state_in_tree(state)) {
 883                         node = &state->rb_node;
 884                         goto hit_next;
 885                 }
 886         }
 887         /*
 888          * this search will find all the extents that end after
 889          * our range starts.
 890          */
 891         node = tree_search_for_insert(tree, start, &p, &parent);
 892         if (!node) {
 893                 prealloc = alloc_extent_state_atomic(prealloc);
 894                 BUG_ON(!prealloc);
 895                 err = insert_state(tree, prealloc, start, end,
 896                                    &p, &parent, &bits, changeset);
 897                 if (err)
 898                         extent_io_tree_panic(tree, err);
 899
 900                 cache_state(prealloc, cached_state);
 901                 prealloc = NULL;
 902                 goto out;
 903         }
 904         state = rb_entry(node, struct extent_state, rb_node);
 905 hit_next:
 906         last_start = state->start;
 907         last_end = state->end;
 908
 909         /*
 910          * | ---- desired range ---- |
 911          * | state |
 912          *
 913          * Just lock what we found and keep going
 914          */
 915         if (state->start == start && state->end <= end) {
 916                 if (state->state & exclusive_bits) {
 917                         *failed_start = state->start;
 918                         err = -EEXIST;
 919                         goto out;
 920                 }
 921
 922                 set_state_bits(tree, state, &bits, changeset);
 923                 cache_state(state, cached_state);
 924                 merge_state(tree, state);
 925                 if (last_end == (u64)-1)
 926                         goto out;
 927                 start = last_end + 1;
 928                 state = next_state(state);
 929                 if (start < end && state && state->start == start &&
 930                     !need_resched())
 931                         goto hit_next;
 932                 goto search_again;
 933         }
 934
 935         /*
 936          *     | ---- desired range ---- |
 937          * | state |
 938          *   or
 939          * | ------------- state -------------- |
 940          *
 941          * We need to split the extent we found, and may flip bits on
 942          * second half.
 943          *
 944          * If the extent we found extends past our
 945          * range, we just split and search again.  It'll get split
 946          * again the next time though.
 947          *
 948          * If the extent we found is inside our range, we set the
 949          * desired bit on it.
 950          */
 951         if (state->start < start) {
 952                 if (state->state & exclusive_bits) {
 953                         *failed_start = start;
 954                         err = -EEXIST;
 955                         goto out;
 956                 }
 957
 958                 prealloc = alloc_extent_state_atomic(prealloc);
 959                 BUG_ON(!prealloc);
 960                 err = split_state(tree, state, prealloc, start);
 961                 if (err)
 962                         extent_io_tree_panic(tree, err);
 963
 964                 prealloc = NULL;
 965                 if (err)
 966                         goto out;
 967                 if (state->end <= end) {
 968                         set_state_bits(tree, state, &bits, changeset);
 969                         cache_state(state, cached_state);
 970                         merge_state(tree, state);
 971                         if (last_end == (u64)-1)
 972                                 goto out;
 973                         start = last_end + 1;
 974                         state = next_state(state);
 975                         if (start < end && state && state->start == start &&
 976                             !need_resched())
 977                                 goto hit_next;
 978                 }
 979                 goto search_again;
 980         }
 981         /*
 982          * | ---- desired range ---- |
 983          *     | state | or               | state |
 984          *
 985          * There's a hole, we need to insert something in it and
 986          * ignore the extent we found.
 987          */
 988         if (state->start > start) {
 989                 u64 this_end;
 990                 if (end < last_start)
 991                         this_end = end;
 992                 else
 993                         this_end = last_start - 1;
 994
 995                 prealloc = alloc_extent_state_atomic(prealloc);
 996                 BUG_ON(!prealloc);
 997
 998                 /*
 999                  * Avoid to free 'prealloc' if it can be merged with
1000                  * the later extent.
1001                  */
1002                 err = insert_state(tree, prealloc, start, this_end,
1003                                    NULL, NULL, &bits, changeset);
1004                 if (err)
1005                         extent_io_tree_panic(tree, err);
1006
1007                 cache_state(prealloc, cached_state);
1008                 prealloc = NULL;
1009                 start = this_end + 1;
1010                 goto search_again;
1011         }
1012         /*
1013          * | ---- desired range ---- |
1014          *                        | state |
1015          * We need to split the extent, and set the bit
1016          * on the first half
1017          */
1018         if (state->start <= end && state->end > end) {
1019                 if (state->state & exclusive_bits) {
1020                         *failed_start = start;
1021                         err = -EEXIST;
1022                         goto out;
1023                 }
1024
1025                 prealloc = alloc_extent_state_atomic(prealloc);
1026                 BUG_ON(!prealloc);
1027                 err = split_state(tree, state, prealloc, end + 1);
1028                 if (err)
1029                         extent_io_tree_panic(tree, err);
1030
1031                 set_state_bits(tree, prealloc, &bits, changeset);
1032                 cache_state(prealloc, cached_state);
1033                 merge_state(tree, prealloc);
1034                 prealloc = NULL;
1035                 goto out;
1036         }
1037
1038 search_again:
1039         if (start > end)
1040                 goto out;
1041         spin_unlock(&tree->lock);
1042         if (gfpflags_allow_blocking(mask))
1043                 cond_resched();
1044         goto again;
1045
1046 out:
1047         spin_unlock(&tree->lock);
1048         if (prealloc)
1049                 free_extent_state(prealloc);
1050
1051         return err;
1052
1053 }
1054
1055 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1056                    unsigned bits, u64 * failed_start,
1057                    struct extent_state **cached_state, gfp_t mask)
1058 {
1059         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1060                                 cached_state, mask, NULL);
1061 }
1062
1063
1064 /**
1065  * convert_extent_bit - convert all bits in a given range from one bit to
1066  *                      another
1067  * @tree:       the io tree to search
1068  * @start:      the start offset in bytes
1069  * @end:        the end offset in bytes (inclusive)
1070  * @bits:       the bits to set in this range
1071  * @clear_bits: the bits to clear in this range
1072  * @cached_state:       state that we're going to cache
1073  *
1074  * This will go through and set bits for the given range.  If any states exist
1075  * already in this range they are set with the given bit and cleared of the
1076  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1077  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1078  * boundary bits like LOCK.
1079  *
1080  * All allocations are done with GFP_NOFS.
1081  */
1082 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1083                        unsigned bits, unsigned clear_bits,
1084                        struct extent_state **cached_state)
1085 {
1086         struct extent_state *state;
1087         struct extent_state *prealloc = NULL;
1088         struct rb_node *node;
1089         struct rb_node **p;
1090         struct rb_node *parent;
1091         int err = 0;
1092         u64 last_start;
1093         u64 last_end;
1094         bool first_iteration = true;
1095
1096         btrfs_debug_check_extent_io_range(tree, start, end);
1097
1098 again:
1099         if (!prealloc) {
1100                 /*
1101                  * Best effort, don't worry if extent state allocation fails
1102                  * here for the first iteration. We might have a cached state
1103                  * that matches exactly the target range, in which case no
1104                  * extent state allocations are needed. We'll only know this
1105                  * after locking the tree.
1106                  */
1107                 prealloc = alloc_extent_state(GFP_NOFS);
1108                 if (!prealloc && !first_iteration)
1109                         return -ENOMEM;
1110         }
1111
1112         spin_lock(&tree->lock);
1113         if (cached_state && *cached_state) {
1114                 state = *cached_state;
1115                 if (state->start <= start && state->end > start &&
1116                     extent_state_in_tree(state)) {
1117                         node = &state->rb_node;
1118                         goto hit_next;
1119                 }
1120         }
1121
1122         /*
1123          * this search will find all the extents that end after
1124          * our range starts.
1125          */
1126         node = tree_search_for_insert(tree, start, &p, &parent);
1127         if (!node) {
1128                 prealloc = alloc_extent_state_atomic(prealloc);
1129                 if (!prealloc) {
1130                         err = -ENOMEM;
1131                         goto out;
1132                 }
1133                 err = insert_state(tree, prealloc, start, end,
1134                                    &p, &parent, &bits, NULL);
1135                 if (err)
1136                         extent_io_tree_panic(tree, err);
1137                 cache_state(prealloc, cached_state);
1138                 prealloc = NULL;
1139                 goto out;
1140         }
1141         state = rb_entry(node, struct extent_state, rb_node);
1142 hit_next:
1143         last_start = state->start;
1144         last_end = state->end;
1145
1146         /*
1147          * | ---- desired range ---- |
1148          * | state |
1149          *
1150          * Just lock what we found and keep going
1151          */
1152         if (state->start == start && state->end <= end) {
1153                 set_state_bits(tree, state, &bits, NULL);
1154                 cache_state(state, cached_state);
1155                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1156                 if (last_end == (u64)-1)
1157                         goto out;
1158                 start = last_end + 1;
1159                 if (start < end && state && state->start == start &&
1160                     !need_resched())
1161                         goto hit_next;
1162                 goto search_again;
1163         }
1164
1165         /*
1166          *     | ---- desired range ---- |
1167          * | state |
1168          *   or
1169          * | ------------- state -------------- |
1170          *
1171          * We need to split the extent we found, and may flip bits on
1172          * second half.
1173          *
1174          * If the extent we found extends past our
1175          * range, we just split and search again.  It'll get split
1176          * again the next time though.
1177          *
1178          * If the extent we found is inside our range, we set the
1179          * desired bit on it.
1180          */
1181         if (state->start < start) {
1182                 prealloc = alloc_extent_state_atomic(prealloc);
1183                 if (!prealloc) {
1184                         err = -ENOMEM;
1185                         goto out;
1186                 }
1187                 err = split_state(tree, state, prealloc, start);
1188                 if (err)
1189                         extent_io_tree_panic(tree, err);
1190                 prealloc = NULL;
1191                 if (err)
1192                         goto out;
1193                 if (state->end <= end) {
1194                         set_state_bits(tree, state, &bits, NULL);
1195                         cache_state(state, cached_state);
1196                         state = clear_state_bit(tree, state, &clear_bits, 0,
1197                                                 NULL);
1198                         if (last_end == (u64)-1)
1199                                 goto out;
1200                         start = last_end + 1;
1201                         if (start < end && state && state->start == start &&
1202                             !need_resched())
1203                                 goto hit_next;
1204                 }
1205                 goto search_again;
1206         }
1207         /*
1208          * | ---- desired range ---- |
1209          *     | state | or               | state |
1210          *
1211          * There's a hole, we need to insert something in it and
1212          * ignore the extent we found.
1213          */
1214         if (state->start > start) {
1215                 u64 this_end;
1216                 if (end < last_start)
1217                         this_end = end;
1218                 else
1219                         this_end = last_start - 1;
1220
1221                 prealloc = alloc_extent_state_atomic(prealloc);
1222                 if (!prealloc) {
1223                         err = -ENOMEM;
1224                         goto out;
1225                 }
1226
1227                 /*
1228                  * Avoid to free 'prealloc' if it can be merged with
1229                  * the later extent.
1230                  */
1231                 err = insert_state(tree, prealloc, start, this_end,
1232                                    NULL, NULL, &bits, NULL);
1233                 if (err)
1234                         extent_io_tree_panic(tree, err);
1235                 cache_state(prealloc, cached_state);
1236                 prealloc = NULL;
1237                 start = this_end + 1;
1238                 goto search_again;
1239         }
1240         /*
1241          * | ---- desired range ---- |
1242          *                        | state |
1243          * We need to split the extent, and set the bit
1244          * on the first half
1245          */
1246         if (state->start <= end && state->end > end) {
1247                 prealloc = alloc_extent_state_atomic(prealloc);
1248                 if (!prealloc) {
1249                         err = -ENOMEM;
1250                         goto out;
1251                 }
1252
1253                 err = split_state(tree, state, prealloc, end + 1);
1254                 if (err)
1255                         extent_io_tree_panic(tree, err);
1256
1257                 set_state_bits(tree, prealloc, &bits, NULL);
1258                 cache_state(prealloc, cached_state);
1259                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1260                 prealloc = NULL;
1261                 goto out;
1262         }
1263
1264 search_again:
1265         if (start > end)
1266                 goto out;
1267         spin_unlock(&tree->lock);
1268         cond_resched();
1269         first_iteration = false;
1270         goto again;
1271
1272 out:
1273         spin_unlock(&tree->lock);
1274         if (prealloc)
1275                 free_extent_state(prealloc);
1276
1277         return err;
1278 }
1279
1280 /* wrappers around set/clear extent bit */
1281 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1282                            unsigned bits, struct extent_changeset *changeset)
1283 {
1284         /*
1285          * We don't support EXTENT_LOCKED yet, as current changeset will
1286          * record any bits changed, so for EXTENT_LOCKED case, it will
1287          * either fail with -EEXIST or changeset will record the whole
1288          * range.
1289          */
1290         BUG_ON(bits & EXTENT_LOCKED);
1291
1292         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1293                                 changeset);
1294 }
1295
1296 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1297                      unsigned bits, int wake, int delete,
1298                      struct extent_state **cached)
1299 {
1300         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1301                                   cached, GFP_NOFS, NULL);
1302 }
1303
1304 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1305                 unsigned bits, struct extent_changeset *changeset)
1306 {
1307         /*
1308          * Don't support EXTENT_LOCKED case, same reason as
1309          * set_record_extent_bits().
1310          */
1311         BUG_ON(bits & EXTENT_LOCKED);
1312
1313         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1314                                   changeset);
1315 }
1316
1317 /*
1318  * either insert or lock state struct between start and end use mask to tell
1319  * us if waiting is desired.
1320  */
1321 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1322                      struct extent_state **cached_state)
1323 {
1324         int err;
1325         u64 failed_start;
1326
1327         while (1) {
1328                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1329                                        EXTENT_LOCKED, &failed_start,
1330                                        cached_state, GFP_NOFS, NULL);
1331                 if (err == -EEXIST) {
1332                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1333                         start = failed_start;
1334                 } else
1335                         break;
1336                 WARN_ON(start > end);
1337         }
1338         return err;
1339 }
1340
1341 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1342 {
1343         int err;
1344         u64 failed_start;
1345
1346         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1347                                &failed_start, NULL, GFP_NOFS, NULL);
1348         if (err == -EEXIST) {
1349                 if (failed_start > start)
1350                         clear_extent_bit(tree, start, failed_start - 1,
1351                                          EXTENT_LOCKED, 1, 0, NULL);
1352                 return 0;
1353         }
1354         return 1;
1355 }
1356
1357 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1358 {
1359         unsigned long index = start >> PAGE_SHIFT;
1360         unsigned long end_index = end >> PAGE_SHIFT;
1361         struct page *page;
1362
1363         while (index <= end_index) {
1364                 page = find_get_page(inode->i_mapping, index);
1365                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1366                 clear_page_dirty_for_io(page);
1367                 put_page(page);
1368                 index++;
1369         }
1370 }
1371
1372 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1373 {
1374         unsigned long index = start >> PAGE_SHIFT;
1375         unsigned long end_index = end >> PAGE_SHIFT;
1376         struct page *page;
1377
1378         while (index <= end_index) {
1379                 page = find_get_page(inode->i_mapping, index);
1380                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1381                 __set_page_dirty_nobuffers(page);
1382                 account_page_redirty(page);
1383                 put_page(page);
1384                 index++;
1385         }
1386 }
1387
1388 /*
1389  * helper function to set both pages and extents in the tree writeback
1390  */
1391 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1392 {
1393         tree->ops->set_range_writeback(tree->private_data, start, end);
1394 }
1395
1396 /* find the first state struct with 'bits' set after 'start', and
1397  * return it.  tree->lock must be held.  NULL will returned if
1398  * nothing was found after 'start'
1399  */
1400 static struct extent_state *
1401 find_first_extent_bit_state(struct extent_io_tree *tree,
1402                             u64 start, unsigned bits)
1403 {
1404         struct rb_node *node;
1405         struct extent_state *state;
1406
1407         /*
1408          * this search will find all the extents that end after
1409          * our range starts.
1410          */
1411         node = tree_search(tree, start);
1412         if (!node)
1413                 goto out;
1414
1415         while (1) {
1416                 state = rb_entry(node, struct extent_state, rb_node);
1417                 if (state->end >= start && (state->state & bits))
1418                         return state;
1419
1420                 node = rb_next(node);
1421                 if (!node)
1422                         break;
1423         }
1424 out:
1425         return NULL;
1426 }
1427
1428 /*
1429  * find the first offset in the io tree with 'bits' set. zero is
1430  * returned if we find something, and *start_ret and *end_ret are
1431  * set to reflect the state struct that was found.
1432  *
1433  * If nothing was found, 1 is returned. If found something, return 0.
1434  */
1435 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1436                           u64 *start_ret, u64 *end_ret, unsigned bits,
1437                           struct extent_state **cached_state)
1438 {
1439         struct extent_state *state;
1440         struct rb_node *n;
1441         int ret = 1;
1442
1443         spin_lock(&tree->lock);
1444         if (cached_state && *cached_state) {
1445                 state = *cached_state;
1446                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1447                         n = rb_next(&state->rb_node);
1448                         while (n) {
1449                                 state = rb_entry(n, struct extent_state,
1450                                                  rb_node);
1451                                 if (state->state & bits)
1452                                         goto got_it;
1453                                 n = rb_next(n);
1454                         }
1455                         free_extent_state(*cached_state);
1456                         *cached_state = NULL;
1457                         goto out;
1458                 }
1459                 free_extent_state(*cached_state);
1460                 *cached_state = NULL;
1461         }
1462
1463         state = find_first_extent_bit_state(tree, start, bits);
1464 got_it:
1465         if (state) {
1466                 cache_state_if_flags(state, cached_state, 0);
1467                 *start_ret = state->start;
1468                 *end_ret = state->end;
1469                 ret = 0;
1470         }
1471 out:
1472         spin_unlock(&tree->lock);
1473         return ret;
1474 }
1475
1476 /*
1477  * find a contiguous range of bytes in the file marked as delalloc, not
1478  * more than 'max_bytes'.  start and end are used to return the range,
1479  *
1480  * 1 is returned if we find something, 0 if nothing was in the tree
1481  */
1482 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1483                                         u64 *start, u64 *end, u64 max_bytes,
1484                                         struct extent_state **cached_state)
1485 {
1486         struct rb_node *node;
1487         struct extent_state *state;
1488         u64 cur_start = *start;
1489         u64 found = 0;
1490         u64 total_bytes = 0;
1491
1492         spin_lock(&tree->lock);
1493
1494         /*
1495          * this search will find all the extents that end after
1496          * our range starts.
1497          */
1498         node = tree_search(tree, cur_start);
1499         if (!node) {
1500                 if (!found)
1501                         *end = (u64)-1;
1502                 goto out;
1503         }
1504
1505         while (1) {
1506                 state = rb_entry(node, struct extent_state, rb_node);
1507                 if (found && (state->start != cur_start ||
1508                               (state->state & EXTENT_BOUNDARY))) {
1509                         goto out;
1510                 }
1511                 if (!(state->state & EXTENT_DELALLOC)) {
1512                         if (!found)
1513                                 *end = state->end;
1514                         goto out;
1515                 }
1516                 if (!found) {
1517                         *start = state->start;
1518                         *cached_state = state;
1519                         refcount_inc(&state->refs);
1520                 }
1521                 found++;
1522                 *end = state->end;
1523                 cur_start = state->end + 1;
1524                 node = rb_next(node);
1525                 total_bytes += state->end - state->start + 1;
1526                 if (total_bytes >= max_bytes)
1527                         break;
1528                 if (!node)
1529                         break;
1530         }
1531 out:
1532         spin_unlock(&tree->lock);
1533         return found;
1534 }
1535
1536 static int __process_pages_contig(struct address_space *mapping,
1537                                   struct page *locked_page,
1538                                   pgoff_t start_index, pgoff_t end_index,
1539                                   unsigned long page_ops, pgoff_t *index_ret);
1540
1541 static noinline void __unlock_for_delalloc(struct inode *inode,
1542                                            struct page *locked_page,
1543                                            u64 start, u64 end)
1544 {
1545         unsigned long index = start >> PAGE_SHIFT;
1546         unsigned long end_index = end >> PAGE_SHIFT;
1547
1548         ASSERT(locked_page);
1549         if (index == locked_page->index && end_index == index)
1550                 return;
1551
1552         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1553                                PAGE_UNLOCK, NULL);
1554 }
1555
1556 static noinline int lock_delalloc_pages(struct inode *inode,
1557                                         struct page *locked_page,
1558                                         u64 delalloc_start,
1559                                         u64 delalloc_end)
1560 {
1561         unsigned long index = delalloc_start >> PAGE_SHIFT;
1562         unsigned long index_ret = index;
1563         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1564         int ret;
1565
1566         ASSERT(locked_page);
1567         if (index == locked_page->index && index == end_index)
1568                 return 0;
1569
1570         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1571                                      end_index, PAGE_LOCK, &index_ret);
1572         if (ret == -EAGAIN)
1573                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1574                                       (u64)index_ret << PAGE_SHIFT);
1575         return ret;
1576 }
1577
1578 /*
1579  * find a contiguous range of bytes in the file marked as delalloc, not
1580  * more than 'max_bytes'.  start and end are used to return the range,
1581  *
1582  * 1 is returned if we find something, 0 if nothing was in the tree
1583  */
1584 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1585                                     struct extent_io_tree *tree,
1586                                     struct page *locked_page, u64 *start,
1587                                     u64 *end, u64 max_bytes)
1588 {
1589         u64 delalloc_start;
1590         u64 delalloc_end;
1591         u64 found;
1592         struct extent_state *cached_state = NULL;
1593         int ret;
1594         int loops = 0;
1595
1596 again:
1597         /* step one, find a bunch of delalloc bytes starting at start */
1598         delalloc_start = *start;
1599         delalloc_end = 0;
1600         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1601                                     max_bytes, &cached_state);
1602         if (!found || delalloc_end <= *start) {
1603                 *start = delalloc_start;
1604                 *end = delalloc_end;
1605                 free_extent_state(cached_state);
1606                 return 0;
1607         }
1608
1609         /*
1610          * start comes from the offset of locked_page.  We have to lock
1611          * pages in order, so we can't process delalloc bytes before
1612          * locked_page
1613          */
1614         if (delalloc_start < *start)
1615                 delalloc_start = *start;
1616
1617         /*
1618          * make sure to limit the number of pages we try to lock down
1619          */
1620         if (delalloc_end + 1 - delalloc_start > max_bytes)
1621                 delalloc_end = delalloc_start + max_bytes - 1;
1622
1623         /* step two, lock all the pages after the page that has start */
1624         ret = lock_delalloc_pages(inode, locked_page,
1625                                   delalloc_start, delalloc_end);
1626         if (ret == -EAGAIN) {
1627                 /* some of the pages are gone, lets avoid looping by
1628                  * shortening the size of the delalloc range we're searching
1629                  */
1630                 free_extent_state(cached_state);
1631                 cached_state = NULL;
1632                 if (!loops) {
1633                         max_bytes = PAGE_SIZE;
1634                         loops = 1;
1635                         goto again;
1636                 } else {
1637                         found = 0;
1638                         goto out_failed;
1639                 }
1640         }
1641         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1642
1643         /* step three, lock the state bits for the whole range */
1644         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1645
1646         /* then test to make sure it is all still delalloc */
1647         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1648                              EXTENT_DELALLOC, 1, cached_state);
1649         if (!ret) {
1650                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1651                                      &cached_state, GFP_NOFS);
1652                 __unlock_for_delalloc(inode, locked_page,
1653                               delalloc_start, delalloc_end);
1654                 cond_resched();
1655                 goto again;
1656         }
1657         free_extent_state(cached_state);
1658         *start = delalloc_start;
1659         *end = delalloc_end;
1660 out_failed:
1661         return found;
1662 }
1663
1664 static int __process_pages_contig(struct address_space *mapping,
1665                                   struct page *locked_page,
1666                                   pgoff_t start_index, pgoff_t end_index,
1667                                   unsigned long page_ops, pgoff_t *index_ret)
1668 {
1669         unsigned long nr_pages = end_index - start_index + 1;
1670         unsigned long pages_locked = 0;
1671         pgoff_t index = start_index;
1672         struct page *pages[16];
1673         unsigned ret;
1674         int err = 0;
1675         int i;
1676
1677         if (page_ops & PAGE_LOCK) {
1678                 ASSERT(page_ops == PAGE_LOCK);
1679                 ASSERT(index_ret && *index_ret == start_index);
1680         }
1681
1682         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1683                 mapping_set_error(mapping, -EIO);
1684
1685         while (nr_pages > 0) {
1686                 ret = find_get_pages_contig(mapping, index,
1687                                      min_t(unsigned long,
1688                                      nr_pages, ARRAY_SIZE(pages)), pages);
1689                 if (ret == 0) {
1690                         /*
1691                          * Only if we're going to lock these pages,
1692                          * can we find nothing at @index.
1693                          */
1694                         ASSERT(page_ops & PAGE_LOCK);
1695                         err = -EAGAIN;
1696                         goto out;
1697                 }
1698
1699                 for (i = 0; i < ret; i++) {
1700                         if (page_ops & PAGE_SET_PRIVATE2)
1701                                 SetPagePrivate2(pages[i]);
1702
1703                         if (pages[i] == locked_page) {
1704                                 put_page(pages[i]);
1705                                 pages_locked++;
1706                                 continue;
1707                         }
1708                         if (page_ops & PAGE_CLEAR_DIRTY)
1709                                 clear_page_dirty_for_io(pages[i]);
1710                         if (page_ops & PAGE_SET_WRITEBACK)
1711                                 set_page_writeback(pages[i]);
1712                         if (page_ops & PAGE_SET_ERROR)
1713                                 SetPageError(pages[i]);
1714                         if (page_ops & PAGE_END_WRITEBACK)
1715                                 end_page_writeback(pages[i]);
1716                         if (page_ops & PAGE_UNLOCK)
1717                                 unlock_page(pages[i]);
1718                         if (page_ops & PAGE_LOCK) {
1719                                 lock_page(pages[i]);
1720                                 if (!PageDirty(pages[i]) ||
1721                                     pages[i]->mapping != mapping) {
1722                                         unlock_page(pages[i]);
1723                                         put_page(pages[i]);
1724                                         err = -EAGAIN;
1725                                         goto out;
1726                                 }
1727                         }
1728                         put_page(pages[i]);
1729                         pages_locked++;
1730                 }
1731                 nr_pages -= ret;
1732                 index += ret;
1733                 cond_resched();
1734         }
1735 out:
1736         if (err && index_ret)
1737                 *index_ret = start_index + pages_locked - 1;
1738         return err;
1739 }
1740
1741 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1742                                  u64 delalloc_end, struct page *locked_page,
1743                                  unsigned clear_bits,
1744                                  unsigned long page_ops)
1745 {
1746         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1747                          NULL);
1748
1749         __process_pages_contig(inode->i_mapping, locked_page,
1750                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1751                                page_ops, NULL);
1752 }
1753
1754 /*
1755  * count the number of bytes in the tree that have a given bit(s)
1756  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1757  * cached.  The total number found is returned.
1758  */
1759 u64 count_range_bits(struct extent_io_tree *tree,
1760                      u64 *start, u64 search_end, u64 max_bytes,
1761                      unsigned bits, int contig)
1762 {
1763         struct rb_node *node;
1764         struct extent_state *state;
1765         u64 cur_start = *start;
1766         u64 total_bytes = 0;
1767         u64 last = 0;
1768         int found = 0;
1769
1770         if (WARN_ON(search_end <= cur_start))
1771                 return 0;
1772
1773         spin_lock(&tree->lock);
1774         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1775                 total_bytes = tree->dirty_bytes;
1776                 goto out;
1777         }
1778         /*
1779          * this search will find all the extents that end after
1780          * our range starts.
1781          */
1782         node = tree_search(tree, cur_start);
1783         if (!node)
1784                 goto out;
1785
1786         while (1) {
1787                 state = rb_entry(node, struct extent_state, rb_node);
1788                 if (state->start > search_end)
1789                         break;
1790                 if (contig && found && state->start > last + 1)
1791                         break;
1792                 if (state->end >= cur_start && (state->state & bits) == bits) {
1793                         total_bytes += min(search_end, state->end) + 1 -
1794                                        max(cur_start, state->start);
1795                         if (total_bytes >= max_bytes)
1796                                 break;
1797                         if (!found) {
1798                                 *start = max(cur_start, state->start);
1799                                 found = 1;
1800                         }
1801                         last = state->end;
1802                 } else if (contig && found) {
1803                         break;
1804                 }
1805                 node = rb_next(node);
1806                 if (!node)
1807                         break;
1808         }
1809 out:
1810         spin_unlock(&tree->lock);
1811         return total_bytes;
1812 }
1813
1814 /*
1815  * set the private field for a given byte offset in the tree.  If there isn't
1816  * an extent_state there already, this does nothing.
1817  */
1818 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1819                 struct io_failure_record *failrec)
1820 {
1821         struct rb_node *node;
1822         struct extent_state *state;
1823         int ret = 0;
1824
1825         spin_lock(&tree->lock);
1826         /*
1827          * this search will find all the extents that end after
1828          * our range starts.
1829          */
1830         node = tree_search(tree, start);
1831         if (!node) {
1832                 ret = -ENOENT;
1833                 goto out;
1834         }
1835         state = rb_entry(node, struct extent_state, rb_node);
1836         if (state->start != start) {
1837                 ret = -ENOENT;
1838                 goto out;
1839         }
1840         state->failrec = failrec;
1841 out:
1842         spin_unlock(&tree->lock);
1843         return ret;
1844 }
1845
1846 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1847                 struct io_failure_record **failrec)
1848 {
1849         struct rb_node *node;
1850         struct extent_state *state;
1851         int ret = 0;
1852
1853         spin_lock(&tree->lock);
1854         /*
1855          * this search will find all the extents that end after
1856          * our range starts.
1857          */
1858         node = tree_search(tree, start);
1859         if (!node) {
1860                 ret = -ENOENT;
1861                 goto out;
1862         }
1863         state = rb_entry(node, struct extent_state, rb_node);
1864         if (state->start != start) {
1865                 ret = -ENOENT;
1866                 goto out;
1867         }
1868         *failrec = state->failrec;
1869 out:
1870         spin_unlock(&tree->lock);
1871         return ret;
1872 }
1873
1874 /*
1875  * searches a range in the state tree for a given mask.
1876  * If 'filled' == 1, this returns 1 only if every extent in the tree
1877  * has the bits set.  Otherwise, 1 is returned if any bit in the
1878  * range is found set.
1879  */
1880 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1881                    unsigned bits, int filled, struct extent_state *cached)
1882 {
1883         struct extent_state *state = NULL;
1884         struct rb_node *node;
1885         int bitset = 0;
1886
1887         spin_lock(&tree->lock);
1888         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1889             cached->end > start)
1890                 node = &cached->rb_node;
1891         else
1892                 node = tree_search(tree, start);
1893         while (node && start <= end) {
1894                 state = rb_entry(node, struct extent_state, rb_node);
1895
1896                 if (filled && state->start > start) {
1897                         bitset = 0;
1898                         break;
1899                 }
1900
1901                 if (state->start > end)
1902                         break;
1903
1904                 if (state->state & bits) {
1905                         bitset = 1;
1906                         if (!filled)
1907                                 break;
1908                 } else if (filled) {
1909                         bitset = 0;
1910                         break;
1911                 }
1912
1913                 if (state->end == (u64)-1)
1914                         break;
1915
1916                 start = state->end + 1;
1917                 if (start > end)
1918                         break;
1919                 node = rb_next(node);
1920                 if (!node) {
1921                         if (filled)
1922                                 bitset = 0;
1923                         break;
1924                 }
1925         }
1926         spin_unlock(&tree->lock);
1927         return bitset;
1928 }
1929
1930 /*
1931  * helper function to set a given page up to date if all the
1932  * extents in the tree for that page are up to date
1933  */
1934 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1935 {
1936         u64 start = page_offset(page);
1937         u64 end = start + PAGE_SIZE - 1;
1938         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1939                 SetPageUptodate(page);
1940 }
1941
1942 int free_io_failure(struct extent_io_tree *failure_tree,
1943                     struct extent_io_tree *io_tree,
1944                     struct io_failure_record *rec)
1945 {
1946         int ret;
1947         int err = 0;
1948
1949         set_state_failrec(failure_tree, rec->start, NULL);
1950         ret = clear_extent_bits(failure_tree, rec->start,
1951                                 rec->start + rec->len - 1,
1952                                 EXTENT_LOCKED | EXTENT_DIRTY);
1953         if (ret)
1954                 err = ret;
1955
1956         ret = clear_extent_bits(io_tree, rec->start,
1957                                 rec->start + rec->len - 1,
1958                                 EXTENT_DAMAGED);
1959         if (ret && !err)
1960                 err = ret;
1961
1962         kfree(rec);
1963         return err;
1964 }
1965
1966 /*
1967  * this bypasses the standard btrfs submit functions deliberately, as
1968  * the standard behavior is to write all copies in a raid setup. here we only
1969  * want to write the one bad copy. so we do the mapping for ourselves and issue
1970  * submit_bio directly.
1971  * to avoid any synchronization issues, wait for the data after writing, which
1972  * actually prevents the read that triggered the error from finishing.
1973  * currently, there can be no more than two copies of every data bit. thus,
1974  * exactly one rewrite is required.
1975  */
1976 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
1977                       u64 length, u64 logical, struct page *page,
1978                       unsigned int pg_offset, int mirror_num)
1979 {
1980         struct bio *bio;
1981         struct btrfs_device *dev;
1982         u64 map_length = 0;
1983         u64 sector;
1984         struct btrfs_bio *bbio = NULL;
1985         int ret;
1986
1987         ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
1988         BUG_ON(!mirror_num);
1989
1990         bio = btrfs_io_bio_alloc(1);
1991         bio->bi_iter.bi_size = 0;
1992         map_length = length;
1993
1994         /*
1995          * Avoid races with device replace and make sure our bbio has devices
1996          * associated to its stripes that don't go away while we are doing the
1997          * read repair operation.
1998          */
1999         btrfs_bio_counter_inc_blocked(fs_info);
2000         if (btrfs_is_parity_mirror(fs_info, logical, length)) {
2001                 /*
2002                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2003                  * to update all raid stripes, but here we just want to correct
2004                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2005                  * stripe's dev and sector.
2006                  */
2007                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2008                                       &map_length, &bbio, 0);
2009                 if (ret) {
2010                         btrfs_bio_counter_dec(fs_info);
2011                         bio_put(bio);
2012                         return -EIO;
2013                 }
2014                 ASSERT(bbio->mirror_num == 1);
2015         } else {
2016                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2017                                       &map_length, &bbio, mirror_num);
2018                 if (ret) {
2019                         btrfs_bio_counter_dec(fs_info);
2020                         bio_put(bio);
2021                         return -EIO;
2022                 }
2023                 BUG_ON(mirror_num != bbio->mirror_num);
2024         }
2025
2026         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2027         bio->bi_iter.bi_sector = sector;
2028         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2029         btrfs_put_bbio(bbio);
2030         if (!dev || !dev->bdev || !dev->writeable) {
2031                 btrfs_bio_counter_dec(fs_info);
2032                 bio_put(bio);
2033                 return -EIO;
2034         }
2035         bio_set_dev(bio, dev->bdev);
2036         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2037         bio_add_page(bio, page, length, pg_offset);
2038
2039         if (btrfsic_submit_bio_wait(bio)) {
2040                 /* try to remap that extent elsewhere? */
2041                 btrfs_bio_counter_dec(fs_info);
2042                 bio_put(bio);
2043                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2044                 return -EIO;
2045         }
2046
2047         btrfs_info_rl_in_rcu(fs_info,
2048                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2049                                   ino, start,
2050                                   rcu_str_deref(dev->name), sector);
2051         btrfs_bio_counter_dec(fs_info);
2052         bio_put(bio);
2053         return 0;
2054 }
2055
2056 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2057                          struct extent_buffer *eb, int mirror_num)
2058 {
2059         u64 start = eb->start;
2060         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2061         int ret = 0;
2062
2063         if (sb_rdonly(fs_info->sb))
2064                 return -EROFS;
2065
2066         for (i = 0; i < num_pages; i++) {
2067                 struct page *p = eb->pages[i];
2068
2069                 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2070                                         start - page_offset(p), mirror_num);
2071                 if (ret)
2072                         break;
2073                 start += PAGE_SIZE;
2074         }
2075
2076         return ret;
2077 }
2078
2079 /*
2080  * each time an IO finishes, we do a fast check in the IO failure tree
2081  * to see if we need to process or clean up an io_failure_record
2082  */
2083 int clean_io_failure(struct btrfs_fs_info *fs_info,
2084                      struct extent_io_tree *failure_tree,
2085                      struct extent_io_tree *io_tree, u64 start,
2086                      struct page *page, u64 ino, unsigned int pg_offset)
2087 {
2088         u64 private;
2089         struct io_failure_record *failrec;
2090         struct extent_state *state;
2091         int num_copies;
2092         int ret;
2093
2094         private = 0;
2095         ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2096                                EXTENT_DIRTY, 0);
2097         if (!ret)
2098                 return 0;
2099
2100         ret = get_state_failrec(failure_tree, start, &failrec);
2101         if (ret)
2102                 return 0;
2103
2104         BUG_ON(!failrec->this_mirror);
2105
2106         if (failrec->in_validation) {
2107                 /* there was no real error, just free the record */
2108                 btrfs_debug(fs_info,
2109                         "clean_io_failure: freeing dummy error at %llu",
2110                         failrec->start);
2111                 goto out;
2112         }
2113         if (sb_rdonly(fs_info->sb))
2114                 goto out;
2115
2116         spin_lock(&io_tree->lock);
2117         state = find_first_extent_bit_state(io_tree,
2118                                             failrec->start,
2119                                             EXTENT_LOCKED);
2120         spin_unlock(&io_tree->lock);
2121
2122         if (state && state->start <= failrec->start &&
2123             state->end >= failrec->start + failrec->len - 1) {
2124                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2125                                               failrec->len);
2126                 if (num_copies > 1)  {
2127                         repair_io_failure(fs_info, ino, start, failrec->len,
2128                                           failrec->logical, page, pg_offset,
2129                                           failrec->failed_mirror);
2130                 }
2131         }
2132
2133 out:
2134         free_io_failure(failure_tree, io_tree, failrec);
2135
2136         return 0;
2137 }
2138
2139 /*
2140  * Can be called when
2141  * - hold extent lock
2142  * - under ordered extent
2143  * - the inode is freeing
2144  */
2145 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2146 {
2147         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2148         struct io_failure_record *failrec;
2149         struct extent_state *state, *next;
2150
2151         if (RB_EMPTY_ROOT(&failure_tree->state))
2152                 return;
2153
2154         spin_lock(&failure_tree->lock);
2155         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2156         while (state) {
2157                 if (state->start > end)
2158                         break;
2159
2160                 ASSERT(state->end <= end);
2161
2162                 next = next_state(state);
2163
2164                 failrec = state->failrec;
2165                 free_extent_state(state);
2166                 kfree(failrec);
2167
2168                 state = next;
2169         }
2170         spin_unlock(&failure_tree->lock);
2171 }
2172
2173 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2174                 struct io_failure_record **failrec_ret)
2175 {
2176         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2177         struct io_failure_record *failrec;
2178         struct extent_map *em;
2179         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2180         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2181         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2182         int ret;
2183         u64 logical;
2184
2185         ret = get_state_failrec(failure_tree, start, &failrec);
2186         if (ret) {
2187                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2188                 if (!failrec)
2189                         return -ENOMEM;
2190
2191                 failrec->start = start;
2192                 failrec->len = end - start + 1;
2193                 failrec->this_mirror = 0;
2194                 failrec->bio_flags = 0;
2195                 failrec->in_validation = 0;
2196
2197                 read_lock(&em_tree->lock);
2198                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2199                 if (!em) {
2200                         read_unlock(&em_tree->lock);
2201                         kfree(failrec);
2202                         return -EIO;
2203                 }
2204
2205                 if (em->start > start || em->start + em->len <= start) {
2206                         free_extent_map(em);
2207                         em = NULL;
2208                 }
2209                 read_unlock(&em_tree->lock);
2210                 if (!em) {
2211                         kfree(failrec);
2212                         return -EIO;
2213                 }
2214
2215                 logical = start - em->start;
2216                 logical = em->block_start + logical;
2217                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2218                         logical = em->block_start;
2219                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2220                         extent_set_compress_type(&failrec->bio_flags,
2221                                                  em->compress_type);
2222                 }
2223
2224                 btrfs_debug(fs_info,
2225                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2226                         logical, start, failrec->len);
2227
2228                 failrec->logical = logical;
2229                 free_extent_map(em);
2230
2231                 /* set the bits in the private failure tree */
2232                 ret = set_extent_bits(failure_tree, start, end,
2233                                         EXTENT_LOCKED | EXTENT_DIRTY);
2234                 if (ret >= 0)
2235                         ret = set_state_failrec(failure_tree, start, failrec);
2236                 /* set the bits in the inode's tree */
2237                 if (ret >= 0)
2238                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2239                 if (ret < 0) {
2240                         kfree(failrec);
2241                         return ret;
2242                 }
2243         } else {
2244                 btrfs_debug(fs_info,
2245                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2246                         failrec->logical, failrec->start, failrec->len,
2247                         failrec->in_validation);
2248                 /*
2249                  * when data can be on disk more than twice, add to failrec here
2250                  * (e.g. with a list for failed_mirror) to make
2251                  * clean_io_failure() clean all those errors at once.
2252                  */
2253         }
2254
2255         *failrec_ret = failrec;
2256
2257         return 0;
2258 }
2259
2260 bool btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2261                            struct io_failure_record *failrec, int failed_mirror)
2262 {
2263         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2264         int num_copies;
2265
2266         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2267         if (num_copies == 1) {
2268                 /*
2269                  * we only have a single copy of the data, so don't bother with
2270                  * all the retry and error correction code that follows. no
2271                  * matter what the error is, it is very likely to persist.
2272                  */
2273                 btrfs_debug(fs_info,
2274                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2275                         num_copies, failrec->this_mirror, failed_mirror);
2276                 return false;
2277         }
2278
2279         /*
2280          * there are two premises:
2281          *      a) deliver good data to the caller
2282          *      b) correct the bad sectors on disk
2283          */
2284         if (failed_bio->bi_vcnt > 1) {
2285                 /*
2286                  * to fulfill b), we need to know the exact failing sectors, as
2287                  * we don't want to rewrite any more than the failed ones. thus,
2288                  * we need separate read requests for the failed bio
2289                  *
2290                  * if the following BUG_ON triggers, our validation request got
2291                  * merged. we need separate requests for our algorithm to work.
2292                  */
2293                 BUG_ON(failrec->in_validation);
2294                 failrec->in_validation = 1;
2295                 failrec->this_mirror = failed_mirror;
2296         } else {
2297                 /*
2298                  * we're ready to fulfill a) and b) alongside. get a good copy
2299                  * of the failed sector and if we succeed, we have setup
2300                  * everything for repair_io_failure to do the rest for us.
2301                  */
2302                 if (failrec->in_validation) {
2303                         BUG_ON(failrec->this_mirror != failed_mirror);
2304                         failrec->in_validation = 0;
2305                         failrec->this_mirror = 0;
2306                 }
2307                 failrec->failed_mirror = failed_mirror;
2308                 failrec->this_mirror++;
2309                 if (failrec->this_mirror == failed_mirror)
2310                         failrec->this_mirror++;
2311         }
2312
2313         if (failrec->this_mirror > num_copies) {
2314                 btrfs_debug(fs_info,
2315                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2316                         num_copies, failrec->this_mirror, failed_mirror);
2317                 return false;
2318         }
2319
2320         return true;
2321 }
2322
2323
2324 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2325                                     struct io_failure_record *failrec,
2326                                     struct page *page, int pg_offset, int icsum,
2327                                     bio_end_io_t *endio_func, void *data)
2328 {
2329         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2330         struct bio *bio;
2331         struct btrfs_io_bio *btrfs_failed_bio;
2332         struct btrfs_io_bio *btrfs_bio;
2333
2334         bio = btrfs_io_bio_alloc(1);
2335         bio->bi_end_io = endio_func;
2336         bio->bi_iter.bi_sector = failrec->logical >> 9;
2337         bio_set_dev(bio, fs_info->fs_devices->latest_bdev);
2338         bio->bi_iter.bi_size = 0;
2339         bio->bi_private = data;
2340
2341         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2342         if (btrfs_failed_bio->csum) {
2343                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2344
2345                 btrfs_bio = btrfs_io_bio(bio);
2346                 btrfs_bio->csum = btrfs_bio->csum_inline;
2347                 icsum *= csum_size;
2348                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2349                        csum_size);
2350         }
2351
2352         bio_add_page(bio, page, failrec->len, pg_offset);
2353
2354         return bio;
2355 }
2356
2357 /*
2358  * this is a generic handler for readpage errors (default
2359  * readpage_io_failed_hook). if other copies exist, read those and write back
2360  * good data to the failed position. does not investigate in remapping the
2361  * failed extent elsewhere, hoping the device will be smart enough to do this as
2362  * needed
2363  */
2364
2365 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2366                               struct page *page, u64 start, u64 end,
2367                               int failed_mirror)
2368 {
2369         struct io_failure_record *failrec;
2370         struct inode *inode = page->mapping->host;
2371         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2372         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2373         struct bio *bio;
2374         int read_mode = 0;
2375         blk_status_t status;
2376         int ret;
2377
2378         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2379
2380         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2381         if (ret)
2382                 return ret;
2383
2384         if (!btrfs_check_repairable(inode, failed_bio, failrec,
2385                                     failed_mirror)) {
2386                 free_io_failure(failure_tree, tree, failrec);
2387                 return -EIO;
2388         }
2389
2390         if (failed_bio->bi_vcnt > 1)
2391                 read_mode |= REQ_FAILFAST_DEV;
2392
2393         phy_offset >>= inode->i_sb->s_blocksize_bits;
2394         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2395                                       start - page_offset(page),
2396                                       (int)phy_offset, failed_bio->bi_end_io,
2397                                       NULL);
2398         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2399
2400         btrfs_debug(btrfs_sb(inode->i_sb),
2401                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2402                 read_mode, failrec->this_mirror, failrec->in_validation);
2403
2404         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2405                                          failrec->bio_flags, 0);
2406         if (status) {
2407                 free_io_failure(failure_tree, tree, failrec);
2408                 bio_put(bio);
2409                 ret = blk_status_to_errno(status);
2410         }
2411
2412         return ret;
2413 }
2414
2415 /* lots and lots of room for performance fixes in the end_bio funcs */
2416
2417 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2418 {
2419         int uptodate = (err == 0);
2420         struct extent_io_tree *tree;
2421         int ret = 0;
2422
2423         tree = &BTRFS_I(page->mapping->host)->io_tree;
2424
2425         if (tree->ops && tree->ops->writepage_end_io_hook)
2426                 tree->ops->writepage_end_io_hook(page, start, end, NULL,
2427                                 uptodate);
2428
2429         if (!uptodate) {
2430                 ClearPageUptodate(page);
2431                 SetPageError(page);
2432                 ret = err < 0 ? err : -EIO;
2433                 mapping_set_error(page->mapping, ret);
2434         }
2435 }
2436
2437 /*
2438  * after a writepage IO is done, we need to:
2439  * clear the uptodate bits on error
2440  * clear the writeback bits in the extent tree for this IO
2441  * end_page_writeback if the page has no more pending IO
2442  *
2443  * Scheduling is not allowed, so the extent state tree is expected
2444  * to have one and only one object corresponding to this IO.
2445  */
2446 static void end_bio_extent_writepage(struct bio *bio)
2447 {
2448         int error = blk_status_to_errno(bio->bi_status);
2449         struct bio_vec *bvec;
2450         u64 start;
2451         u64 end;
2452         int i;
2453
2454         ASSERT(!bio_flagged(bio, BIO_CLONED));
2455         bio_for_each_segment_all(bvec, bio, i) {
2456                 struct page *page = bvec->bv_page;
2457                 struct inode *inode = page->mapping->host;
2458                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2459
2460                 /* We always issue full-page reads, but if some block
2461                  * in a page fails to read, blk_update_request() will
2462                  * advance bv_offset and adjust bv_len to compensate.
2463                  * Print a warning for nonzero offsets, and an error
2464                  * if they don't add up to a full page.  */
2465                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2466                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2467                                 btrfs_err(fs_info,
2468                                    "partial page write in btrfs with offset %u and length %u",
2469                                         bvec->bv_offset, bvec->bv_len);
2470                         else
2471                                 btrfs_info(fs_info,
2472                                    "incomplete page write in btrfs with offset %u and length %u",
2473                                         bvec->bv_offset, bvec->bv_len);
2474                 }
2475
2476                 start = page_offset(page);
2477                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2478
2479                 end_extent_writepage(page, error, start, end);
2480                 end_page_writeback(page);
2481         }
2482
2483         bio_put(bio);
2484 }
2485
2486 static void
2487 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2488                               int uptodate)
2489 {
2490         struct extent_state *cached = NULL;
2491         u64 end = start + len - 1;
2492
2493         if (uptodate && tree->track_uptodate)
2494                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2495         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2496 }
2497
2498 /*
2499  * after a readpage IO is done, we need to:
2500  * clear the uptodate bits on error
2501  * set the uptodate bits if things worked
2502  * set the page up to date if all extents in the tree are uptodate
2503  * clear the lock bit in the extent tree
2504  * unlock the page if there are no other extents locked for it
2505  *
2506  * Scheduling is not allowed, so the extent state tree is expected
2507  * to have one and only one object corresponding to this IO.
2508  */
2509 static void end_bio_extent_readpage(struct bio *bio)
2510 {
2511         struct bio_vec *bvec;
2512         int uptodate = !bio->bi_status;
2513         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2514         struct extent_io_tree *tree, *failure_tree;
2515         u64 offset = 0;
2516         u64 start;
2517         u64 end;
2518         u64 len;
2519         u64 extent_start = 0;
2520         u64 extent_len = 0;
2521         int mirror;
2522         int ret;
2523         int i;
2524
2525         ASSERT(!bio_flagged(bio, BIO_CLONED));
2526         bio_for_each_segment_all(bvec, bio, i) {
2527                 struct page *page = bvec->bv_page;
2528                 struct inode *inode = page->mapping->host;
2529                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2530
2531                 btrfs_debug(fs_info,
2532                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2533                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2534                         io_bio->mirror_num);
2535                 tree = &BTRFS_I(inode)->io_tree;
2536                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2537
2538                 /* We always issue full-page reads, but if some block
2539                  * in a page fails to read, blk_update_request() will
2540                  * advance bv_offset and adjust bv_len to compensate.
2541                  * Print a warning for nonzero offsets, and an error
2542                  * if they don't add up to a full page.  */
2543                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2544                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2545                                 btrfs_err(fs_info,
2546                                         "partial page read in btrfs with offset %u and length %u",
2547                                         bvec->bv_offset, bvec->bv_len);
2548                         else
2549                                 btrfs_info(fs_info,
2550                                         "incomplete page read in btrfs with offset %u and length %u",
2551                                         bvec->bv_offset, bvec->bv_len);
2552                 }
2553
2554                 start = page_offset(page);
2555                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2556                 len = bvec->bv_len;
2557
2558                 mirror = io_bio->mirror_num;
2559                 if (likely(uptodate && tree->ops)) {
2560                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2561                                                               page, start, end,
2562                                                               mirror);
2563                         if (ret)
2564                                 uptodate = 0;
2565                         else
2566                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2567                                                  failure_tree, tree, start,
2568                                                  page,
2569                                                  btrfs_ino(BTRFS_I(inode)), 0);
2570                 }
2571
2572                 if (likely(uptodate))
2573                         goto readpage_ok;
2574
2575                 if (tree->ops) {
2576                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2577                         if (ret == -EAGAIN) {
2578                                 /*
2579                                  * Data inode's readpage_io_failed_hook() always
2580                                  * returns -EAGAIN.
2581                                  *
2582                                  * The generic bio_readpage_error handles errors
2583                                  * the following way: If possible, new read
2584                                  * requests are created and submitted and will
2585                                  * end up in end_bio_extent_readpage as well (if
2586                                  * we're lucky, not in the !uptodate case). In
2587                                  * that case it returns 0 and we just go on with
2588                                  * the next page in our bio. If it can't handle
2589                                  * the error it will return -EIO and we remain
2590                                  * responsible for that page.
2591                                  */
2592                                 ret = bio_readpage_error(bio, offset, page,
2593                                                          start, end, mirror);
2594                                 if (ret == 0) {
2595                                         uptodate = !bio->bi_status;
2596                                         offset += len;
2597                                         continue;
2598                                 }
2599                         }
2600
2601                         /*
2602                          * metadata's readpage_io_failed_hook() always returns
2603                          * -EIO and fixes nothing.  -EIO is also returned if
2604                          * data inode error could not be fixed.
2605                          */
2606                         ASSERT(ret == -EIO);
2607                 }
2608 readpage_ok:
2609                 if (likely(uptodate)) {
2610                         loff_t i_size = i_size_read(inode);
2611                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2612                         unsigned off;
2613
2614                         /* Zero out the end if this page straddles i_size */
2615                         off = i_size & (PAGE_SIZE-1);
2616                         if (page->index == end_index && off)
2617                                 zero_user_segment(page, off, PAGE_SIZE);
2618                         SetPageUptodate(page);
2619                 } else {
2620                         ClearPageUptodate(page);
2621                         SetPageError(page);
2622                 }
2623                 unlock_page(page);
2624                 offset += len;
2625
2626                 if (unlikely(!uptodate)) {
2627                         if (extent_len) {
2628                                 endio_readpage_release_extent(tree,
2629                                                               extent_start,
2630                                                               extent_len, 1);
2631                                 extent_start = 0;
2632                                 extent_len = 0;
2633                         }
2634                         endio_readpage_release_extent(tree, start,
2635                                                       end - start + 1, 0);
2636                 } else if (!extent_len) {
2637                         extent_start = start;
2638                         extent_len = end + 1 - start;
2639                 } else if (extent_start + extent_len == start) {
2640                         extent_len += end + 1 - start;
2641                 } else {
2642                         endio_readpage_release_extent(tree, extent_start,
2643                                                       extent_len, uptodate);
2644                         extent_start = start;
2645                         extent_len = end + 1 - start;
2646                 }
2647         }
2648
2649         if (extent_len)
2650                 endio_readpage_release_extent(tree, extent_start, extent_len,
2651                                               uptodate);
2652         if (io_bio->end_io)
2653                 io_bio->end_io(io_bio, blk_status_to_errno(bio->bi_status));
2654         bio_put(bio);
2655 }
2656
2657 /*
2658  * Initialize the members up to but not including 'bio'. Use after allocating a
2659  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2660  * 'bio' because use of __GFP_ZERO is not supported.
2661  */
2662 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2663 {
2664         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2665 }
2666
2667 /*
2668  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2669  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2670  * for the appropriate container_of magic
2671  */
2672 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte)
2673 {
2674         struct bio *bio;
2675
2676         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, btrfs_bioset);
2677         bio_set_dev(bio, bdev);
2678         bio->bi_iter.bi_sector = first_byte >> 9;
2679         btrfs_io_bio_init(btrfs_io_bio(bio));
2680         return bio;
2681 }
2682
2683 struct bio *btrfs_bio_clone(struct bio *bio)
2684 {
2685         struct btrfs_io_bio *btrfs_bio;
2686         struct bio *new;
2687
2688         /* Bio allocation backed by a bioset does not fail */
2689         new = bio_clone_fast(bio, GFP_NOFS, btrfs_bioset);
2690         btrfs_bio = btrfs_io_bio(new);
2691         btrfs_io_bio_init(btrfs_bio);
2692         btrfs_bio->iter = bio->bi_iter;
2693         return new;
2694 }
2695
2696 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2697 {
2698         struct bio *bio;
2699
2700         /* Bio allocation backed by a bioset does not fail */
2701         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, btrfs_bioset);
2702         btrfs_io_bio_init(btrfs_io_bio(bio));
2703         return bio;
2704 }
2705
2706 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2707 {
2708         struct bio *bio;
2709         struct btrfs_io_bio *btrfs_bio;
2710
2711         /* this will never fail when it's backed by a bioset */
2712         bio = bio_clone_fast(orig, GFP_NOFS, btrfs_bioset);
2713         ASSERT(bio);
2714
2715         btrfs_bio = btrfs_io_bio(bio);
2716         btrfs_io_bio_init(btrfs_bio);
2717
2718         bio_trim(bio, offset >> 9, size >> 9);
2719         btrfs_bio->iter = bio->bi_iter;
2720         return bio;
2721 }
2722
2723 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2724                                        unsigned long bio_flags)
2725 {
2726         blk_status_t ret = 0;
2727         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2728         struct page *page = bvec->bv_page;
2729         struct extent_io_tree *tree = bio->bi_private;
2730         u64 start;
2731
2732         start = page_offset(page) + bvec->bv_offset;
2733
2734         bio->bi_private = NULL;
2735         bio_get(bio);
2736
2737         if (tree->ops)
2738                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
2739                                            mirror_num, bio_flags, start);
2740         else
2741                 btrfsic_submit_bio(bio);
2742
2743         bio_put(bio);
2744         return blk_status_to_errno(ret);
2745 }
2746
2747 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2748                      unsigned long offset, size_t size, struct bio *bio,
2749                      unsigned long bio_flags)
2750 {
2751         int ret = 0;
2752         if (tree->ops)
2753                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2754                                                 bio_flags);
2755         return ret;
2756
2757 }
2758
2759 /*
2760  * @opf:        bio REQ_OP_* and REQ_* flags as one value
2761  */
2762 static int submit_extent_page(unsigned int opf, struct extent_io_tree *tree,
2763                               struct writeback_control *wbc,
2764                               struct page *page, u64 offset,
2765                               size_t size, unsigned long pg_offset,
2766                               struct block_device *bdev,
2767                               struct bio **bio_ret,
2768                               bio_end_io_t end_io_func,
2769                               int mirror_num,
2770                               unsigned long prev_bio_flags,
2771                               unsigned long bio_flags,
2772                               bool force_bio_submit)
2773 {
2774         int ret = 0;
2775         struct bio *bio;
2776         int contig = 0;
2777         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2778         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2779         sector_t sector = offset >> 9;
2780
2781         if (bio_ret && *bio_ret) {
2782                 bio = *bio_ret;
2783                 if (old_compressed)
2784                         contig = bio->bi_iter.bi_sector == sector;
2785                 else
2786                         contig = bio_end_sector(bio) == sector;
2787
2788                 if (prev_bio_flags != bio_flags || !contig ||
2789                     force_bio_submit ||
2790                     merge_bio(tree, page, pg_offset, page_size, bio, bio_flags) ||
2791                     bio_add_page(bio, page, page_size, pg_offset) < page_size) {
2792                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2793                         if (ret < 0) {
2794                                 *bio_ret = NULL;
2795                                 return ret;
2796                         }
2797                         bio = NULL;
2798                 } else {
2799                         if (wbc)
2800                                 wbc_account_io(wbc, page, page_size);
2801                         return 0;
2802                 }
2803         }
2804
2805         bio = btrfs_bio_alloc(bdev, offset);
2806         bio_add_page(bio, page, page_size, pg_offset);
2807         bio->bi_end_io = end_io_func;
2808         bio->bi_private = tree;
2809         bio->bi_write_hint = page->mapping->host->i_write_hint;
2810         bio->bi_opf = opf;
2811         if (wbc) {
2812                 wbc_init_bio(wbc, bio);
2813                 wbc_account_io(wbc, page, page_size);
2814         }
2815
2816         if (bio_ret)
2817                 *bio_ret = bio;
2818         else
2819                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2820
2821         return ret;
2822 }
2823
2824 static void attach_extent_buffer_page(struct extent_buffer *eb,
2825                                       struct page *page)
2826 {
2827         if (!PagePrivate(page)) {
2828                 SetPagePrivate(page);
2829                 get_page(page);
2830                 set_page_private(page, (unsigned long)eb);
2831         } else {
2832                 WARN_ON(page->private != (unsigned long)eb);
2833         }
2834 }
2835
2836 void set_page_extent_mapped(struct page *page)
2837 {
2838         if (!PagePrivate(page)) {
2839                 SetPagePrivate(page);
2840                 get_page(page);
2841                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2842         }
2843 }
2844
2845 static struct extent_map *
2846 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2847                  u64 start, u64 len, get_extent_t *get_extent,
2848                  struct extent_map **em_cached)
2849 {
2850         struct extent_map *em;
2851
2852         if (em_cached && *em_cached) {
2853                 em = *em_cached;
2854                 if (extent_map_in_tree(em) && start >= em->start &&
2855                     start < extent_map_end(em)) {
2856                         refcount_inc(&em->refs);
2857                         return em;
2858                 }
2859
2860                 free_extent_map(em);
2861                 *em_cached = NULL;
2862         }
2863
2864         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2865         if (em_cached && !IS_ERR_OR_NULL(em)) {
2866                 BUG_ON(*em_cached);
2867                 refcount_inc(&em->refs);
2868                 *em_cached = em;
2869         }
2870         return em;
2871 }
2872 /*
2873  * basic readpage implementation.  Locked extent state structs are inserted
2874  * into the tree that are removed when the IO is done (by the end_io
2875  * handlers)
2876  * XXX JDM: This needs looking at to ensure proper page locking
2877  * return 0 on success, otherwise return error
2878  */
2879 static int __do_readpage(struct extent_io_tree *tree,
2880                          struct page *page,
2881                          get_extent_t *get_extent,
2882                          struct extent_map **em_cached,
2883                          struct bio **bio, int mirror_num,
2884                          unsigned long *bio_flags, unsigned int read_flags,
2885                          u64 *prev_em_start)
2886 {
2887         struct inode *inode = page->mapping->host;
2888         u64 start = page_offset(page);
2889         u64 page_end = start + PAGE_SIZE - 1;
2890         u64 end;
2891         u64 cur = start;
2892         u64 extent_offset;
2893         u64 last_byte = i_size_read(inode);
2894         u64 block_start;
2895         u64 cur_end;
2896         struct extent_map *em;
2897         struct block_device *bdev;
2898         int ret = 0;
2899         int nr = 0;
2900         size_t pg_offset = 0;
2901         size_t iosize;
2902         size_t disk_io_size;
2903         size_t blocksize = inode->i_sb->s_blocksize;
2904         unsigned long this_bio_flag = 0;
2905
2906         set_page_extent_mapped(page);
2907
2908         end = page_end;
2909         if (!PageUptodate(page)) {
2910                 if (cleancache_get_page(page) == 0) {
2911                         BUG_ON(blocksize != PAGE_SIZE);
2912                         unlock_extent(tree, start, end);
2913                         goto out;
2914                 }
2915         }
2916
2917         if (page->index == last_byte >> PAGE_SHIFT) {
2918                 char *userpage;
2919                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2920
2921                 if (zero_offset) {
2922                         iosize = PAGE_SIZE - zero_offset;
2923                         userpage = kmap_atomic(page);
2924                         memset(userpage + zero_offset, 0, iosize);
2925                         flush_dcache_page(page);
2926                         kunmap_atomic(userpage);
2927                 }
2928         }
2929         while (cur <= end) {
2930                 bool force_bio_submit = false;
2931                 u64 offset;
2932
2933                 if (cur >= last_byte) {
2934                         char *userpage;
2935                         struct extent_state *cached = NULL;
2936
2937                         iosize = PAGE_SIZE - pg_offset;
2938                         userpage = kmap_atomic(page);
2939                         memset(userpage + pg_offset, 0, iosize);
2940                         flush_dcache_page(page);
2941                         kunmap_atomic(userpage);
2942                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2943                                             &cached, GFP_NOFS);
2944                         unlock_extent_cached(tree, cur,
2945                                              cur + iosize - 1,
2946                                              &cached, GFP_NOFS);
2947                         break;
2948                 }
2949                 em = __get_extent_map(inode, page, pg_offset, cur,
2950                                       end - cur + 1, get_extent, em_cached);
2951                 if (IS_ERR_OR_NULL(em)) {
2952                         SetPageError(page);
2953                         unlock_extent(tree, cur, end);
2954                         break;
2955                 }
2956                 extent_offset = cur - em->start;
2957                 BUG_ON(extent_map_end(em) <= cur);
2958                 BUG_ON(end < cur);
2959
2960                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2961                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2962                         extent_set_compress_type(&this_bio_flag,
2963                                                  em->compress_type);
2964                 }
2965
2966                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2967                 cur_end = min(extent_map_end(em) - 1, end);
2968                 iosize = ALIGN(iosize, blocksize);
2969                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2970                         disk_io_size = em->block_len;
2971                         offset = em->block_start;
2972                 } else {
2973                         offset = em->block_start + extent_offset;
2974                         disk_io_size = iosize;
2975                 }
2976                 bdev = em->bdev;
2977                 block_start = em->block_start;
2978                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2979                         block_start = EXTENT_MAP_HOLE;
2980
2981                 /*
2982                  * If we have a file range that points to a compressed extent
2983                  * and it's followed by a consecutive file range that points to
2984                  * to the same compressed extent (possibly with a different
2985                  * offset and/or length, so it either points to the whole extent
2986                  * or only part of it), we must make sure we do not submit a
2987                  * single bio to populate the pages for the 2 ranges because
2988                  * this makes the compressed extent read zero out the pages
2989                  * belonging to the 2nd range. Imagine the following scenario:
2990                  *
2991                  *  File layout
2992                  *  [0 - 8K]                     [8K - 24K]
2993                  *    |                               |
2994                  *    |                               |
2995                  * points to extent X,         points to extent X,
2996                  * offset 4K, length of 8K     offset 0, length 16K
2997                  *
2998                  * [extent X, compressed length = 4K uncompressed length = 16K]
2999                  *
3000                  * If the bio to read the compressed extent covers both ranges,
3001                  * it will decompress extent X into the pages belonging to the
3002                  * first range and then it will stop, zeroing out the remaining
3003                  * pages that belong to the other range that points to extent X.
3004                  * So here we make sure we submit 2 bios, one for the first
3005                  * range and another one for the third range. Both will target
3006                  * the same physical extent from disk, but we can't currently
3007                  * make the compressed bio endio callback populate the pages
3008                  * for both ranges because each compressed bio is tightly
3009                  * coupled with a single extent map, and each range can have
3010                  * an extent map with a different offset value relative to the
3011                  * uncompressed data of our extent and different lengths. This
3012                  * is a corner case so we prioritize correctness over
3013                  * non-optimal behavior (submitting 2 bios for the same extent).
3014                  */
3015                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3016                     prev_em_start && *prev_em_start != (u64)-1 &&
3017                     *prev_em_start != em->orig_start)
3018                         force_bio_submit = true;
3019
3020                 if (prev_em_start)
3021                         *prev_em_start = em->orig_start;
3022
3023                 free_extent_map(em);
3024                 em = NULL;
3025
3026                 /* we've found a hole, just zero and go on */
3027                 if (block_start == EXTENT_MAP_HOLE) {
3028                         char *userpage;
3029                         struct extent_state *cached = NULL;
3030
3031                         userpage = kmap_atomic(page);
3032                         memset(userpage + pg_offset, 0, iosize);
3033                         flush_dcache_page(page);
3034                         kunmap_atomic(userpage);
3035
3036                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3037                                             &cached, GFP_NOFS);
3038                         unlock_extent_cached(tree, cur,
3039                                              cur + iosize - 1,
3040                                              &cached, GFP_NOFS);
3041                         cur = cur + iosize;
3042                         pg_offset += iosize;
3043                         continue;
3044                 }
3045                 /* the get_extent function already copied into the page */
3046                 if (test_range_bit(tree, cur, cur_end,
3047                                    EXTENT_UPTODATE, 1, NULL)) {
3048                         check_page_uptodate(tree, page);
3049                         unlock_extent(tree, cur, cur + iosize - 1);
3050                         cur = cur + iosize;
3051                         pg_offset += iosize;
3052                         continue;
3053                 }
3054                 /* we have an inline extent but it didn't get marked up
3055                  * to date.  Error out
3056                  */
3057                 if (block_start == EXTENT_MAP_INLINE) {
3058                         SetPageError(page);
3059                         unlock_extent(tree, cur, cur + iosize - 1);
3060                         cur = cur + iosize;
3061                         pg_offset += iosize;
3062                         continue;
3063                 }
3064
3065                 ret = submit_extent_page(REQ_OP_READ | read_flags, tree, NULL,
3066                                          page, offset, disk_io_size,
3067                                          pg_offset, bdev, bio,
3068                                          end_bio_extent_readpage, mirror_num,
3069                                          *bio_flags,
3070                                          this_bio_flag,
3071                                          force_bio_submit);
3072                 if (!ret) {
3073                         nr++;
3074                         *bio_flags = this_bio_flag;
3075                 } else {
3076                         SetPageError(page);
3077                         unlock_extent(tree, cur, cur + iosize - 1);
3078                         goto out;
3079                 }
3080                 cur = cur + iosize;
3081                 pg_offset += iosize;
3082         }
3083 out:
3084         if (!nr) {
3085                 if (!PageError(page))
3086                         SetPageUptodate(page);
3087                 unlock_page(page);
3088         }
3089         return ret;
3090 }
3091
3092 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3093                                              struct page *pages[], int nr_pages,
3094                                              u64 start, u64 end,
3095                                              get_extent_t *get_extent,
3096                                              struct extent_map **em_cached,
3097                                              struct bio **bio,
3098                                              unsigned long *bio_flags,
3099                                              u64 *prev_em_start)
3100 {
3101         struct inode *inode;
3102         struct btrfs_ordered_extent *ordered;
3103         int index;
3104
3105         inode = pages[0]->mapping->host;
3106         while (1) {
3107                 lock_extent(tree, start, end);
3108                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3109                                                      end - start + 1);
3110                 if (!ordered)
3111                         break;
3112                 unlock_extent(tree, start, end);
3113                 btrfs_start_ordered_extent(inode, ordered, 1);
3114                 btrfs_put_ordered_extent(ordered);
3115         }
3116
3117         for (index = 0; index < nr_pages; index++) {
3118                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3119                               0, bio_flags, 0, prev_em_start);
3120                 put_page(pages[index]);
3121         }
3122 }