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