fs/btrfs/extent_io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include <linux/bitops.h>
   4 #include <linux/slab.h>
   5 #include <linux/bio.h>
   6 #include <linux/mm.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/page-flags.h>
   9 #include <linux/spinlock.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/swap.h>
  12 #include <linux/writeback.h>
  13 #include <linux/pagevec.h>
  14 #include <linux/prefetch.h>
  15 #include <linux/cleancache.h>
  16 #include "extent_io.h"
  17 #include "extent_map.h"
  18 #include "ctree.h"
  19 #include "btrfs_inode.h"
  20 #include "volumes.h"
  21 #include "check-integrity.h"
  22 #include "locking.h"
  23 #include "rcu-string.h"
  24 #include "backref.h"
  25 #include "disk-io.h"
  26
  27 static struct kmem_cache *extent_state_cache;
  28 static struct kmem_cache *extent_buffer_cache;
  29 static struct bio_set btrfs_bioset;
  30
  31 static inline bool extent_state_in_tree(const struct extent_state *state)
  32 {
  33         return !RB_EMPTY_NODE(&state->rb_node);
  34 }
  35
  36 #ifdef CONFIG_BTRFS_DEBUG
  37 static LIST_HEAD(buffers);
  38 static LIST_HEAD(states);
  39
  40 static DEFINE_SPINLOCK(leak_lock);
  41
  42 static inline
  43 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  44 {
  45         unsigned long flags;
  46
  47         spin_lock_irqsave(&leak_lock, flags);
  48         list_add(new, head);
  49         spin_unlock_irqrestore(&leak_lock, flags);
  50 }
  51
  52 static inline
  53 void btrfs_leak_debug_del(struct list_head *entry)
  54 {
  55         unsigned long flags;
  56
  57         spin_lock_irqsave(&leak_lock, flags);
  58         list_del(entry);
  59         spin_unlock_irqrestore(&leak_lock, flags);
  60 }
  61
  62 static inline
  63 void btrfs_leak_debug_check(void)
  64 {
  65         struct extent_state *state;
  66         struct extent_buffer *eb;
  67
  68         while (!list_empty(&states)) {
  69                 state = list_entry(states.next, struct extent_state, leak_list);
  70                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  71                        state->start, state->end, state->state,
  72                        extent_state_in_tree(state),
  73                        refcount_read(&state->refs));
  74                 list_del(&state->leak_list);
  75                 kmem_cache_free(extent_state_cache, state);
  76         }
  77
  78         while (!list_empty(&buffers)) {
  79                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  80                 pr_err("BTRFS: buffer leak start %llu len %lu refs %d bflags %lu\n",
  81                        eb->start, eb->len, atomic_read(&eb->refs), eb->bflags);
  82                 list_del(&eb->leak_list);
  83                 kmem_cache_free(extent_buffer_cache, eb);
  84         }
  85 }
  86
  87 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  88         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  89 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  90                 struct extent_io_tree *tree, u64 start, u64 end)
  91 {
  92         struct inode *inode = tree->private_data;
  93         u64 isize;
  94
  95         if (!inode || !is_data_inode(inode))
  96                 return;
  97
  98         isize = i_size_read(inode);
  99         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
 100                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
 101                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
 102                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
 103         }
 104 }
 105 #else
 106 #define btrfs_leak_debug_add(new, head) do {} while (0)
 107 #define btrfs_leak_debug_del(entry)     do {} while (0)
 108 #define btrfs_leak_debug_check()        do {} while (0)
 109 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 110 #endif
 111
 112 #define BUFFER_LRU_MAX 64
 113
 114 struct tree_entry {
 115         u64 start;
 116         u64 end;
 117         struct rb_node rb_node;
 118 };
 119
 120 struct extent_page_data {
 121         struct bio *bio;
 122         struct extent_io_tree *tree;
 123         /* tells writepage not to lock the state bits for this range
 124          * it still does the unlocking
 125          */
 126         unsigned int extent_locked:1;
 127
 128         /* tells the submit_bio code to use REQ_SYNC */
 129         unsigned int sync_io:1;
 130 };
 131
 132 static int add_extent_changeset(struct extent_state *state, unsigned bits,
 133                                  struct extent_changeset *changeset,
 134                                  int set)
 135 {
 136         int ret;
 137
 138         if (!changeset)
 139                 return 0;
 140         if (set && (state->state & bits) == bits)
 141                 return 0;
 142         if (!set && (state->state & bits) == 0)
 143                 return 0;
 144         changeset->bytes_changed += state->end - state->start + 1;
 145         ret = ulist_add(&changeset->range_changed, state->start, state->end,
 146                         GFP_ATOMIC);
 147         return ret;
 148 }
 149
 150 static void flush_write_bio(struct extent_page_data *epd);
 151
 152 int __init extent_io_init(void)
 153 {
 154         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 155                         sizeof(struct extent_state), 0,
 156                         SLAB_MEM_SPREAD, NULL);
 157         if (!extent_state_cache)
 158                 return -ENOMEM;
 159
 160         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 161                         sizeof(struct extent_buffer), 0,
 162                         SLAB_MEM_SPREAD, NULL);
 163         if (!extent_buffer_cache)
 164                 goto free_state_cache;
 165
 166         if (bioset_init(&btrfs_bioset, BIO_POOL_SIZE,
 167                         offsetof(struct btrfs_io_bio, bio),
 168                         BIOSET_NEED_BVECS))
 169                 goto free_buffer_cache;
 170
 171         if (bioset_integrity_create(&btrfs_bioset, BIO_POOL_SIZE))
 172                 goto free_bioset;
 173
 174         return 0;
 175
 176 free_bioset:
 177         bioset_exit(&btrfs_bioset);
 178
 179 free_buffer_cache:
 180         kmem_cache_destroy(extent_buffer_cache);
 181         extent_buffer_cache = NULL;
 182
 183 free_state_cache:
 184         kmem_cache_destroy(extent_state_cache);
 185         extent_state_cache = NULL;
 186         return -ENOMEM;
 187 }
 188
 189 void __cold extent_io_exit(void)
 190 {
 191         btrfs_leak_debug_check();
 192
 193         /*
 194          * Make sure all delayed rcu free are flushed before we
 195          * destroy caches.
 196          */
 197         rcu_barrier();
 198         kmem_cache_destroy(extent_state_cache);
 199         kmem_cache_destroy(extent_buffer_cache);
 200         bioset_exit(&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 /*
 357  * utility function to look for merge candidates inside a given range.
 358  * Any extents with matching state are merged together into a single
 359  * extent in the tree.  Extents with EXTENT_IO in their state field
 360  * are not merged because the end_io handlers need to be able to do
 361  * operations on them without sleeping (or doing allocations/splits).
 362  *
 363  * This should be called with the tree lock held.
 364  */
 365 static void merge_state(struct extent_io_tree *tree,
 366                         struct extent_state *state)
 367 {
 368         struct extent_state *other;
 369         struct rb_node *other_node;
 370
 371         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 372                 return;
 373
 374         other_node = rb_prev(&state->rb_node);
 375         if (other_node) {
 376                 other = rb_entry(other_node, struct extent_state, rb_node);
 377                 if (other->end == state->start - 1 &&
 378                     other->state == state->state) {
 379                         if (tree->private_data &&
 380                             is_data_inode(tree->private_data))
 381                                 btrfs_merge_delalloc_extent(tree->private_data,
 382                                                             state, other);
 383                         state->start = other->start;
 384                         rb_erase(&other->rb_node, &tree->state);
 385                         RB_CLEAR_NODE(&other->rb_node);
 386                         free_extent_state(other);
 387                 }
 388         }
 389         other_node = rb_next(&state->rb_node);
 390         if (other_node) {
 391                 other = rb_entry(other_node, struct extent_state, rb_node);
 392                 if (other->start == state->end + 1 &&
 393                     other->state == state->state) {
 394                         if (tree->private_data &&
 395                             is_data_inode(tree->private_data))
 396                                 btrfs_merge_delalloc_extent(tree->private_data,
 397                                                             state, other);
 398                         state->end = other->end;
 399                         rb_erase(&other->rb_node, &tree->state);
 400                         RB_CLEAR_NODE(&other->rb_node);
 401                         free_extent_state(other);
 402                 }
 403         }
 404 }
 405
 406 static void set_state_bits(struct extent_io_tree *tree,
 407                            struct extent_state *state, unsigned *bits,
 408                            struct extent_changeset *changeset);
 409
 410 /*
 411  * insert an extent_state struct into the tree.  'bits' are set on the
 412  * struct before it is inserted.
 413  *
 414  * This may return -EEXIST if the extent is already there, in which case the
 415  * state struct is freed.
 416  *
 417  * The tree lock is not taken internally.  This is a utility function and
 418  * probably isn't what you want to call (see set/clear_extent_bit).
 419  */
 420 static int insert_state(struct extent_io_tree *tree,
 421                         struct extent_state *state, u64 start, u64 end,
 422                         struct rb_node ***p,
 423                         struct rb_node **parent,
 424                         unsigned *bits, struct extent_changeset *changeset)
 425 {
 426         struct rb_node *node;
 427
 428         if (end < start)
 429                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 430                        end, start);
 431         state->start = start;
 432         state->end = end;
 433
 434         set_state_bits(tree, state, bits, changeset);
 435
 436         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 437         if (node) {
 438                 struct extent_state *found;
 439                 found = rb_entry(node, struct extent_state, rb_node);
 440                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
 441                        found->start, found->end, start, end);
 442                 return -EEXIST;
 443         }
 444         merge_state(tree, state);
 445         return 0;
 446 }
 447
 448 /*
 449  * split a given extent state struct in two, inserting the preallocated
 450  * struct 'prealloc' as the newly created second half.  'split' indicates an
 451  * offset inside 'orig' where it should be split.
 452  *
 453  * Before calling,
 454  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 455  * are two extent state structs in the tree:
 456  * prealloc: [orig->start, split - 1]
 457  * orig: [ split, orig->end ]
 458  *
 459  * The tree locks are not taken by this function. They need to be held
 460  * by the caller.
 461  */
 462 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 463                        struct extent_state *prealloc, u64 split)
 464 {
 465         struct rb_node *node;
 466
 467         if (tree->private_data && is_data_inode(tree->private_data))
 468                 btrfs_split_delalloc_extent(tree->private_data, orig, split);
 469
 470         prealloc->start = orig->start;
 471         prealloc->end = split - 1;
 472         prealloc->state = orig->state;
 473         orig->start = split;
 474
 475         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 476                            &prealloc->rb_node, NULL, NULL);
 477         if (node) {
 478                 free_extent_state(prealloc);
 479                 return -EEXIST;
 480         }
 481         return 0;
 482 }
 483
 484 static struct extent_state *next_state(struct extent_state *state)
 485 {
 486         struct rb_node *next = rb_next(&state->rb_node);
 487         if (next)
 488                 return rb_entry(next, struct extent_state, rb_node);
 489         else
 490                 return NULL;
 491 }
 492
 493 /*
 494  * utility function to clear some bits in an extent state struct.
 495  * it will optionally wake up any one waiting on this state (wake == 1).
 496  *
 497  * If no bits are set on the state struct after clearing things, the
 498  * struct is freed and removed from the tree
 499  */
 500 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 501                                             struct extent_state *state,
 502                                             unsigned *bits, int wake,
 503                                             struct extent_changeset *changeset)
 504 {
 505         struct extent_state *next;
 506         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 507         int ret;
 508
 509         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 510                 u64 range = state->end - state->start + 1;
 511                 WARN_ON(range > tree->dirty_bytes);
 512                 tree->dirty_bytes -= range;
 513         }
 514
 515         if (tree->private_data && is_data_inode(tree->private_data))
 516                 btrfs_clear_delalloc_extent(tree->private_data, state, bits);
 517
 518         ret = add_extent_changeset(state, bits_to_clear, changeset, 0);
 519         BUG_ON(ret < 0);
 520         state->state &= ~bits_to_clear;
 521         if (wake)
 522                 wake_up(&state->wq);
 523         if (state->state == 0) {
 524                 next = next_state(state);
 525                 if (extent_state_in_tree(state)) {
 526                         rb_erase(&state->rb_node, &tree->state);
 527                         RB_CLEAR_NODE(&state->rb_node);
 528                         free_extent_state(state);
 529                 } else {
 530                         WARN_ON(1);
 531                 }
 532         } else {
 533                 merge_state(tree, state);
 534                 next = next_state(state);
 535         }
 536         return next;
 537 }
 538
 539 static struct extent_state *
 540 alloc_extent_state_atomic(struct extent_state *prealloc)
 541 {
 542         if (!prealloc)
 543                 prealloc = alloc_extent_state(GFP_ATOMIC);
 544
 545         return prealloc;
 546 }
 547
 548 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 549 {
 550         struct inode *inode = tree->private_data;
 551
 552         btrfs_panic(btrfs_sb(inode->i_sb), err,
 553         "locking error: extent tree was modified by another thread while locked");
 554 }
 555
 556 /*
 557  * clear some bits on a range in the tree.  This may require splitting
 558  * or inserting elements in the tree, so the gfp mask is used to
 559  * indicate which allocations or sleeping are allowed.
 560  *
 561  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 562  * the given range from the tree regardless of state (ie for truncate).
 563  *
 564  * the range [start, end] is inclusive.
 565  *
 566  * This takes the tree lock, and returns 0 on success and < 0 on error.
 567  */
 568 int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 569                               unsigned bits, int wake, int delete,
 570                               struct extent_state **cached_state,
 571                               gfp_t mask, struct extent_changeset *changeset)
 572 {
 573         struct extent_state *state;
 574         struct extent_state *cached;
 575         struct extent_state *prealloc = NULL;
 576         struct rb_node *node;
 577         u64 last_end;
 578         int err;
 579         int clear = 0;
 580
 581         btrfs_debug_check_extent_io_range(tree, start, end);
 582
 583         if (bits & EXTENT_DELALLOC)
 584                 bits |= EXTENT_NORESERVE;
 585
 586         if (delete)
 587                 bits |= ~EXTENT_CTLBITS;
 588         bits |= EXTENT_FIRST_DELALLOC;
 589
 590         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 591                 clear = 1;
 592 again:
 593         if (!prealloc && gfpflags_allow_blocking(mask)) {
 594                 /*
 595                  * Don't care for allocation failure here because we might end
 596                  * up not needing the pre-allocated extent state at all, which
 597                  * is the case if we only have in the tree extent states that
 598                  * cover our input range and don't cover too any other range.
 599                  * If we end up needing a new extent state we allocate it later.
 600                  */
 601                 prealloc = alloc_extent_state(mask);
 602         }
 603
 604         spin_lock(&tree->lock);
 605         if (cached_state) {
 606                 cached = *cached_state;
 607
 608                 if (clear) {
 609                         *cached_state = NULL;
 610                         cached_state = NULL;
 611                 }
 612
 613                 if (cached && extent_state_in_tree(cached) &&
 614                     cached->start <= start && cached->end > start) {
 615                         if (clear)
 616                                 refcount_dec(&cached->refs);
 617                         state = cached;
 618                         goto hit_next;
 619                 }
 620                 if (clear)
 621                         free_extent_state(cached);
 622         }
 623         /*
 624          * this search will find the extents that end after
 625          * our range starts
 626          */
 627         node = tree_search(tree, start);
 628         if (!node)
 629                 goto out;
 630         state = rb_entry(node, struct extent_state, rb_node);
 631 hit_next:
 632         if (state->start > end)
 633                 goto out;
 634         WARN_ON(state->end < start);
 635         last_end = state->end;
 636
 637         /* the state doesn't have the wanted bits, go ahead */
 638         if (!(state->state & bits)) {
 639                 state = next_state(state);
 640                 goto next;
 641         }
 642
 643         /*
 644          *     | ---- desired range ---- |
 645          *  | state | or
 646          *  | ------------- state -------------- |
 647          *
 648          * We need to split the extent we found, and may flip
 649          * bits on second half.
 650          *
 651          * If the extent we found extends past our range, we
 652          * just split and search again.  It'll get split again
 653          * the next time though.
 654          *
 655          * If the extent we found is inside our range, we clear
 656          * the desired bit on it.
 657          */
 658
 659         if (state->start < start) {
 660                 prealloc = alloc_extent_state_atomic(prealloc);
 661                 BUG_ON(!prealloc);
 662                 err = split_state(tree, state, prealloc, start);
 663                 if (err)
 664                         extent_io_tree_panic(tree, err);
 665
 666                 prealloc = NULL;
 667                 if (err)
 668                         goto out;
 669                 if (state->end <= end) {
 670                         state = clear_state_bit(tree, state, &bits, wake,
 671                                                 changeset);
 672                         goto next;
 673                 }
 674                 goto search_again;
 675         }
 676         /*
 677          * | ---- desired range ---- |
 678          *                        | state |
 679          * We need to split the extent, and clear the bit
 680          * on the first half
 681          */
 682         if (state->start <= end && state->end > end) {
 683                 prealloc = alloc_extent_state_atomic(prealloc);
 684                 BUG_ON(!prealloc);
 685                 err = split_state(tree, state, prealloc, end + 1);
 686                 if (err)
 687                         extent_io_tree_panic(tree, err);
 688
 689                 if (wake)
 690                         wake_up(&state->wq);
 691
 692                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 693
 694                 prealloc = NULL;
 695                 goto out;
 696         }
 697
 698         state = clear_state_bit(tree, state, &bits, wake, changeset);
 699 next:
 700         if (last_end == (u64)-1)
 701                 goto out;
 702         start = last_end + 1;
 703         if (start <= end && state && !need_resched())
 704                 goto hit_next;
 705
 706 search_again:
 707         if (start > end)
 708                 goto out;
 709         spin_unlock(&tree->lock);
 710         if (gfpflags_allow_blocking(mask))
 711                 cond_resched();
 712         goto again;
 713
 714 out:
 715         spin_unlock(&tree->lock);
 716         if (prealloc)
 717                 free_extent_state(prealloc);
 718
 719         return 0;
 720
 721 }
 722
 723 static void wait_on_state(struct extent_io_tree *tree,
 724                           struct extent_state *state)
 725                 __releases(tree->lock)
 726                 __acquires(tree->lock)
 727 {
 728         DEFINE_WAIT(wait);
 729         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 730         spin_unlock(&tree->lock);
 731         schedule();
 732         spin_lock(&tree->lock);
 733         finish_wait(&state->wq, &wait);
 734 }
 735
 736 /*
 737  * waits for one or more bits to clear on a range in the state tree.
 738  * The range [start, end] is inclusive.
 739  * The tree lock is taken by this function
 740  */
 741 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 742                             unsigned long bits)
 743 {
 744         struct extent_state *state;
 745         struct rb_node *node;
 746
 747         btrfs_debug_check_extent_io_range(tree, start, end);
 748
 749         spin_lock(&tree->lock);
 750 again:
 751         while (1) {
 752                 /*
 753                  * this search will find all the extents that end after
 754                  * our range starts
 755                  */
 756                 node = tree_search(tree, start);
 757 process_node:
 758                 if (!node)
 759                         break;
 760
 761                 state = rb_entry(node, struct extent_state, rb_node);
 762
 763                 if (state->start > end)
 764                         goto out;
 765
 766                 if (state->state & bits) {
 767                         start = state->start;
 768                         refcount_inc(&state->refs);
 769                         wait_on_state(tree, state);
 770                         free_extent_state(state);
 771                         goto again;
 772                 }
 773                 start = state->end + 1;
 774
 775                 if (start > end)
 776                         break;
 777
 778                 if (!cond_resched_lock(&tree->lock)) {
 779                         node = rb_next(node);
 780                         goto process_node;
 781                 }
 782         }
 783 out:
 784         spin_unlock(&tree->lock);
 785 }
 786
 787 static void set_state_bits(struct extent_io_tree *tree,
 788                            struct extent_state *state,
 789                            unsigned *bits, struct extent_changeset *changeset)
 790 {
 791         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 792         int ret;
 793
 794         if (tree->private_data && is_data_inode(tree->private_data))
 795                 btrfs_set_delalloc_extent(tree->private_data, state, bits);
 796
 797         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 798                 u64 range = state->end - state->start + 1;
 799                 tree->dirty_bytes += range;
 800         }
 801         ret = add_extent_changeset(state, bits_to_set, changeset, 1);
 802         BUG_ON(ret < 0);
 803         state->state |= bits_to_set;
 804 }
 805
 806 static void cache_state_if_flags(struct extent_state *state,
 807                                  struct extent_state **cached_ptr,
 808                                  unsigned flags)
 809 {
 810         if (cached_ptr && !(*cached_ptr)) {
 811                 if (!flags || (state->state & flags)) {
 812                         *cached_ptr = state;
 813                         refcount_inc(&state->refs);
 814                 }
 815         }
 816 }
 817
 818 static void cache_state(struct extent_state *state,
 819                         struct extent_state **cached_ptr)
 820 {
 821         return cache_state_if_flags(state, cached_ptr,
 822                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 823 }
 824
 825 /*
 826  * set some bits on a range in the tree.  This may require allocations or
 827  * sleeping, so the gfp mask is used to indicate what is allowed.
 828  *
 829  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 830  * part of the range already has the desired bits set.  The start of the
 831  * existing range is returned in failed_start in this case.
 832  *
 833  * [start, end] is inclusive This takes the tree lock.
 834  */
 835
 836 static int __must_check
 837 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 838                  unsigned bits, unsigned exclusive_bits,
 839                  u64 *failed_start, struct extent_state **cached_state,
 840                  gfp_t mask, struct extent_changeset *changeset)
 841 {
 842         struct extent_state *state;
 843         struct extent_state *prealloc = NULL;
 844         struct rb_node *node;
 845         struct rb_node **p;
 846         struct rb_node *parent;
 847         int err = 0;
 848         u64 last_start;
 849         u64 last_end;
 850
 851         btrfs_debug_check_extent_io_range(tree, start, end);
 852
 853         bits |= EXTENT_FIRST_DELALLOC;
 854 again:
 855         if (!prealloc && gfpflags_allow_blocking(mask)) {
 856                 /*
 857                  * Don't care for allocation failure here because we might end
 858                  * up not needing the pre-allocated extent state at all, which
 859                  * is the case if we only have in the tree extent states that
 860                  * cover our input range and don't cover too any other range.
 861                  * If we end up needing a new extent state we allocate it later.
 862                  */
 863                 prealloc = alloc_extent_state(mask);
 864         }
 865
 866         spin_lock(&tree->lock);
 867         if (cached_state && *cached_state) {
 868                 state = *cached_state;
 869                 if (state->start <= start && state->end > start &&
 870                     extent_state_in_tree(state)) {
 871                         node = &state->rb_node;
 872                         goto hit_next;
 873                 }
 874         }
 875         /*
 876          * this search will find all the extents that end after
 877          * our range starts.
 878          */
 879         node = tree_search_for_insert(tree, start, &p, &parent);
 880         if (!node) {
 881                 prealloc = alloc_extent_state_atomic(prealloc);
 882                 BUG_ON(!prealloc);
 883                 err = insert_state(tree, prealloc, start, end,
 884                                    &p, &parent, &bits, changeset);
 885                 if (err)
 886                         extent_io_tree_panic(tree, err);
 887
 888                 cache_state(prealloc, cached_state);
 889                 prealloc = NULL;
 890                 goto out;
 891         }
 892         state = rb_entry(node, struct extent_state, rb_node);
 893 hit_next:
 894         last_start = state->start;
 895         last_end = state->end;
 896
 897         /*
 898          * | ---- desired range ---- |
 899          * | state |
 900          *
 901          * Just lock what we found and keep going
 902          */
 903         if (state->start == start && state->end <= end) {
 904                 if (state->state & exclusive_bits) {
 905                         *failed_start = state->start;
 906                         err = -EEXIST;
 907                         goto out;
 908                 }
 909
 910                 set_state_bits(tree, state, &bits, changeset);
 911                 cache_state(state, cached_state);
 912                 merge_state(tree, state);
 913                 if (last_end == (u64)-1)
 914                         goto out;
 915                 start = last_end + 1;
 916                 state = next_state(state);
 917                 if (start < end && state && state->start == start &&
 918                     !need_resched())
 919                         goto hit_next;
 920                 goto search_again;
 921         }
 922
 923         /*
 924          *     | ---- desired range ---- |
 925          * | state |
 926          *   or
 927          * | ------------- state -------------- |
 928          *
 929          * We need to split the extent we found, and may flip bits on
 930          * second half.
 931          *
 932          * If the extent we found extends past our
 933          * range, we just split and search again.  It'll get split
 934          * again the next time though.
 935          *
 936          * If the extent we found is inside our range, we set the
 937          * desired bit on it.
 938          */
 939         if (state->start < start) {
 940                 if (state->state & exclusive_bits) {
 941                         *failed_start = start;
 942                         err = -EEXIST;
 943                         goto out;
 944                 }
 945
 946                 prealloc = alloc_extent_state_atomic(prealloc);
 947                 BUG_ON(!prealloc);
 948                 err = split_state(tree, state, prealloc, start);
 949                 if (err)
 950                         extent_io_tree_panic(tree, err);
 951
 952                 prealloc = NULL;
 953                 if (err)
 954                         goto out;
 955                 if (state->end <= end) {
 956                         set_state_bits(tree, state, &bits, changeset);
 957                         cache_state(state, cached_state);
 958                         merge_state(tree, state);
 959                         if (last_end == (u64)-1)
 960                                 goto out;
 961                         start = last_end + 1;
 962                         state = next_state(state);
 963                         if (start < end && state && state->start == start &&
 964                             !need_resched())
 965                                 goto hit_next;
 966                 }
 967                 goto search_again;
 968         }
 969         /*
 970          * | ---- desired range ---- |
 971          *     | state | or               | state |
 972          *
 973          * There's a hole, we need to insert something in it and
 974          * ignore the extent we found.
 975          */
 976         if (state->start > start) {
 977                 u64 this_end;
 978                 if (end < last_start)
 979                         this_end = end;
 980                 else
 981                         this_end = last_start - 1;
 982
 983                 prealloc = alloc_extent_state_atomic(prealloc);
 984                 BUG_ON(!prealloc);
 985
 986                 /*
 987                  * Avoid to free 'prealloc' if it can be merged with
 988                  * the later extent.
 989                  */
 990                 err = insert_state(tree, prealloc, start, this_end,
 991                                    NULL, NULL, &bits, changeset);
 992                 if (err)
 993                         extent_io_tree_panic(tree, err);
 994
 995                 cache_state(prealloc, cached_state);
 996                 prealloc = NULL;
 997                 start = this_end + 1;
 998                 goto search_again;
 999         }
1000         /*
1001          * | ---- desired range ---- |
1002          *                        | state |
1003          * We need to split the extent, and set the bit
1004          * on the first half
1005          */
1006         if (state->start <= end && state->end > end) {
1007                 if (state->state & exclusive_bits) {
1008                         *failed_start = start;
1009                         err = -EEXIST;
1010                         goto out;
1011                 }
1012
1013                 prealloc = alloc_extent_state_atomic(prealloc);
1014                 BUG_ON(!prealloc);
1015                 err = split_state(tree, state, prealloc, end + 1);
1016                 if (err)
1017                         extent_io_tree_panic(tree, err);
1018
1019                 set_state_bits(tree, prealloc, &bits, changeset);
1020                 cache_state(prealloc, cached_state);
1021                 merge_state(tree, prealloc);
1022                 prealloc = NULL;
1023                 goto out;
1024         }
1025
1026 search_again:
1027         if (start > end)
1028                 goto out;
1029         spin_unlock(&tree->lock);
1030         if (gfpflags_allow_blocking(mask))
1031                 cond_resched();
1032         goto again;
1033
1034 out:
1035         spin_unlock(&tree->lock);
1036         if (prealloc)
1037                 free_extent_state(prealloc);
1038
1039         return err;
1040
1041 }
1042
1043 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1044                    unsigned bits, u64 * failed_start,
1045                    struct extent_state **cached_state, gfp_t mask)
1046 {
1047         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1048                                 cached_state, mask, NULL);
1049 }
1050
1051
1052 /**
1053  * convert_extent_bit - convert all bits in a given range from one bit to
1054  *                      another
1055  * @tree:       the io tree to search
1056  * @start:      the start offset in bytes
1057  * @end:        the end offset in bytes (inclusive)
1058  * @bits:       the bits to set in this range
1059  * @clear_bits: the bits to clear in this range
1060  * @cached_state:       state that we're going to cache
1061  *
1062  * This will go through and set bits for the given range.  If any states exist
1063  * already in this range they are set with the given bit and cleared of the
1064  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1065  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1066  * boundary bits like LOCK.
1067  *
1068  * All allocations are done with GFP_NOFS.
1069  */
1070 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1071                        unsigned bits, unsigned clear_bits,
1072                        struct extent_state **cached_state)
1073 {
1074         struct extent_state *state;
1075         struct extent_state *prealloc = NULL;
1076         struct rb_node *node;
1077         struct rb_node **p;
1078         struct rb_node *parent;
1079         int err = 0;
1080         u64 last_start;
1081         u64 last_end;
1082         bool first_iteration = true;
1083
1084         btrfs_debug_check_extent_io_range(tree, start, end);
1085
1086 again:
1087         if (!prealloc) {
1088                 /*
1089                  * Best effort, don't worry if extent state allocation fails
1090                  * here for the first iteration. We might have a cached state
1091                  * that matches exactly the target range, in which case no
1092                  * extent state allocations are needed. We'll only know this
1093                  * after locking the tree.
1094                  */
1095                 prealloc = alloc_extent_state(GFP_NOFS);
1096                 if (!prealloc && !first_iteration)
1097                         return -ENOMEM;
1098         }
1099
1100         spin_lock(&tree->lock);
1101         if (cached_state && *cached_state) {
1102                 state = *cached_state;
1103                 if (state->start <= start && state->end > start &&
1104                     extent_state_in_tree(state)) {
1105                         node = &state->rb_node;
1106                         goto hit_next;
1107                 }
1108         }
1109
1110         /*
1111          * this search will find all the extents that end after
1112          * our range starts.
1113          */
1114         node = tree_search_for_insert(tree, start, &p, &parent);
1115         if (!node) {
1116                 prealloc = alloc_extent_state_atomic(prealloc);
1117                 if (!prealloc) {
1118                         err = -ENOMEM;
1119                         goto out;
1120                 }
1121                 err = insert_state(tree, prealloc, start, end,
1122                                    &p, &parent, &bits, NULL);
1123                 if (err)
1124                         extent_io_tree_panic(tree, err);
1125                 cache_state(prealloc, cached_state);
1126                 prealloc = NULL;
1127                 goto out;
1128         }
1129         state = rb_entry(node, struct extent_state, rb_node);
1130 hit_next:
1131         last_start = state->start;
1132         last_end = state->end;
1133
1134         /*
1135          * | ---- desired range ---- |
1136          * | state |
1137          *
1138          * Just lock what we found and keep going
1139          */
1140         if (state->start == start && state->end <= end) {
1141                 set_state_bits(tree, state, &bits, NULL);
1142                 cache_state(state, cached_state);
1143                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1144                 if (last_end == (u64)-1)
1145                         goto out;
1146                 start = last_end + 1;
1147                 if (start < end && state && state->start == start &&
1148                     !need_resched())
1149                         goto hit_next;
1150                 goto search_again;
1151         }
1152
1153         /*
1154          *     | ---- desired range ---- |
1155          * | state |
1156          *   or
1157          * | ------------- state -------------- |
1158          *
1159          * We need to split the extent we found, and may flip bits on
1160          * second half.
1161          *
1162          * If the extent we found extends past our
1163          * range, we just split and search again.  It'll get split
1164          * again the next time though.
1165          *
1166          * If the extent we found is inside our range, we set the
1167          * desired bit on it.
1168          */
1169         if (state->start < start) {
1170                 prealloc = alloc_extent_state_atomic(prealloc);
1171                 if (!prealloc) {
1172                         err = -ENOMEM;
1173                         goto out;
1174                 }
1175                 err = split_state(tree, state, prealloc, start);
1176                 if (err)
1177                         extent_io_tree_panic(tree, err);
1178                 prealloc = NULL;
1179                 if (err)
1180                         goto out;
1181                 if (state->end <= end) {
1182                         set_state_bits(tree, state, &bits, NULL);
1183                         cache_state(state, cached_state);
1184                         state = clear_state_bit(tree, state, &clear_bits, 0,
1185                                                 NULL);
1186                         if (last_end == (u64)-1)
1187                                 goto out;
1188                         start = last_end + 1;
1189                         if (start < end && state && state->start == start &&
1190                             !need_resched())
1191                                 goto hit_next;
1192                 }
1193                 goto search_again;
1194         }
1195         /*
1196          * | ---- desired range ---- |
1197          *     | state | or               | state |
1198          *
1199          * There's a hole, we need to insert something in it and
1200          * ignore the extent we found.
1201          */
1202         if (state->start > start) {
1203                 u64 this_end;
1204                 if (end < last_start)
1205                         this_end = end;
1206                 else
1207                         this_end = last_start - 1;
1208
1209                 prealloc = alloc_extent_state_atomic(prealloc);
1210                 if (!prealloc) {
1211                         err = -ENOMEM;
1212                         goto out;
1213                 }
1214
1215                 /*
1216                  * Avoid to free 'prealloc' if it can be merged with
1217                  * the later extent.
1218                  */
1219                 err = insert_state(tree, prealloc, start, this_end,
1220                                    NULL, NULL, &bits, NULL);
1221                 if (err)
1222                         extent_io_tree_panic(tree, err);
1223                 cache_state(prealloc, cached_state);
1224                 prealloc = NULL;
1225                 start = this_end + 1;
1226                 goto search_again;
1227         }
1228         /*
1229          * | ---- desired range ---- |
1230          *                        | state |
1231          * We need to split the extent, and set the bit
1232          * on the first half
1233          */
1234         if (state->start <= end && state->end > end) {
1235                 prealloc = alloc_extent_state_atomic(prealloc);
1236                 if (!prealloc) {
1237                         err = -ENOMEM;
1238                         goto out;
1239                 }
1240
1241                 err = split_state(tree, state, prealloc, end + 1);
1242                 if (err)
1243                         extent_io_tree_panic(tree, err);
1244
1245                 set_state_bits(tree, prealloc, &bits, NULL);
1246                 cache_state(prealloc, cached_state);
1247                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1248                 prealloc = NULL;
1249                 goto out;
1250         }
1251
1252 search_again:
1253         if (start > end)
1254                 goto out;
1255         spin_unlock(&tree->lock);
1256         cond_resched();
1257         first_iteration = false;
1258         goto again;
1259
1260 out:
1261         spin_unlock(&tree->lock);
1262         if (prealloc)
1263                 free_extent_state(prealloc);
1264
1265         return err;
1266 }
1267
1268 /* wrappers around set/clear extent bit */
1269 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1270                            unsigned bits, struct extent_changeset *changeset)
1271 {
1272         /*
1273          * We don't support EXTENT_LOCKED yet, as current changeset will
1274          * record any bits changed, so for EXTENT_LOCKED case, it will
1275          * either fail with -EEXIST or changeset will record the whole
1276          * range.
1277          */
1278         BUG_ON(bits & EXTENT_LOCKED);
1279
1280         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1281                                 changeset);
1282 }
1283
1284 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1285                      unsigned bits, int wake, int delete,
1286                      struct extent_state **cached)
1287 {
1288         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1289                                   cached, GFP_NOFS, NULL);
1290 }
1291
1292 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1293                 unsigned bits, struct extent_changeset *changeset)
1294 {
1295         /*
1296          * Don't support EXTENT_LOCKED case, same reason as
1297          * set_record_extent_bits().
1298          */
1299         BUG_ON(bits & EXTENT_LOCKED);
1300
1301         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1302                                   changeset);
1303 }
1304
1305 /*
1306  * either insert or lock state struct between start and end use mask to tell
1307  * us if waiting is desired.
1308  */
1309 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1310                      struct extent_state **cached_state)
1311 {
1312         int err;
1313         u64 failed_start;
1314
1315         while (1) {
1316                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1317                                        EXTENT_LOCKED, &failed_start,
1318                                        cached_state, GFP_NOFS, NULL);
1319                 if (err == -EEXIST) {
1320                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1321                         start = failed_start;
1322                 } else
1323                         break;
1324                 WARN_ON(start > end);
1325         }
1326         return err;
1327 }
1328
1329 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1330 {
1331         int err;
1332         u64 failed_start;
1333
1334         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1335                                &failed_start, NULL, GFP_NOFS, NULL);
1336         if (err == -EEXIST) {
1337                 if (failed_start > start)
1338                         clear_extent_bit(tree, start, failed_start - 1,
1339                                          EXTENT_LOCKED, 1, 0, NULL);
1340                 return 0;
1341         }
1342         return 1;
1343 }
1344
1345 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1346 {
1347         unsigned long index = start >> PAGE_SHIFT;
1348         unsigned long end_index = end >> PAGE_SHIFT;
1349         struct page *page;
1350
1351         while (index <= end_index) {
1352                 page = find_get_page(inode->i_mapping, index);
1353                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1354                 clear_page_dirty_for_io(page);
1355                 put_page(page);
1356                 index++;
1357         }
1358 }
1359
1360 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1361 {
1362         unsigned long index = start >> PAGE_SHIFT;
1363         unsigned long end_index = end >> PAGE_SHIFT;
1364         struct page *page;
1365
1366         while (index <= end_index) {
1367                 page = find_get_page(inode->i_mapping, index);
1368                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1369                 __set_page_dirty_nobuffers(page);
1370                 account_page_redirty(page);
1371                 put_page(page);
1372                 index++;
1373         }
1374 }
1375
1376 /* find the first state struct with 'bits' set after 'start', and
1377  * return it.  tree->lock must be held.  NULL will returned if
1378  * nothing was found after 'start'
1379  */
1380 static struct extent_state *
1381 find_first_extent_bit_state(struct extent_io_tree *tree,
1382                             u64 start, unsigned bits)
1383 {
1384         struct rb_node *node;
1385         struct extent_state *state;
1386
1387         /*
1388          * this search will find all the extents that end after
1389          * our range starts.
1390          */
1391         node = tree_search(tree, start);
1392         if (!node)
1393                 goto out;
1394
1395         while (1) {
1396                 state = rb_entry(node, struct extent_state, rb_node);
1397                 if (state->end >= start && (state->state & bits))
1398                         return state;
1399
1400                 node = rb_next(node);
1401                 if (!node)
1402                         break;
1403         }
1404 out:
1405         return NULL;
1406 }
1407
1408 /*
1409  * find the first offset in the io tree with 'bits' set. zero is
1410  * returned if we find something, and *start_ret and *end_ret are
1411  * set to reflect the state struct that was found.
1412  *
1413  * If nothing was found, 1 is returned. If found something, return 0.
1414  */
1415 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1416                           u64 *start_ret, u64 *end_ret, unsigned bits,
1417                           struct extent_state **cached_state)
1418 {
1419         struct extent_state *state;
1420         int ret = 1;
1421
1422         spin_lock(&tree->lock);
1423         if (cached_state && *cached_state) {
1424                 state = *cached_state;
1425                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1426                         while ((state = next_state(state)) != NULL) {
1427                                 if (state->state & bits)
1428                                         goto got_it;
1429                         }
1430                         free_extent_state(*cached_state);
1431                         *cached_state = NULL;
1432                         goto out;
1433                 }
1434                 free_extent_state(*cached_state);
1435                 *cached_state = NULL;
1436         }
1437
1438         state = find_first_extent_bit_state(tree, start, bits);
1439 got_it:
1440         if (state) {
1441                 cache_state_if_flags(state, cached_state, 0);
1442                 *start_ret = state->start;
1443                 *end_ret = state->end;
1444                 ret = 0;
1445         }
1446 out:
1447         spin_unlock(&tree->lock);
1448         return ret;
1449 }
1450
1451 /*
1452  * find a contiguous range of bytes in the file marked as delalloc, not
1453  * more than 'max_bytes'.  start and end are used to return the range,
1454  *
1455  * 1 is returned if we find something, 0 if nothing was in the tree
1456  */
1457 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1458                                         u64 *start, u64 *end, u64 max_bytes,
1459                                         struct extent_state **cached_state)
1460 {
1461         struct rb_node *node;
1462         struct extent_state *state;
1463         u64 cur_start = *start;
1464         u64 found = 0;
1465         u64 total_bytes = 0;
1466
1467         spin_lock(&tree->lock);
1468
1469         /*
1470          * this search will find all the extents that end after
1471          * our range starts.
1472          */
1473         node = tree_search(tree, cur_start);
1474         if (!node) {
1475                 if (!found)
1476                         *end = (u64)-1;
1477                 goto out;
1478         }
1479
1480         while (1) {
1481                 state = rb_entry(node, struct extent_state, rb_node);
1482                 if (found && (state->start != cur_start ||
1483                               (state->state & EXTENT_BOUNDARY))) {
1484                         goto out;
1485                 }
1486                 if (!(state->state & EXTENT_DELALLOC)) {
1487                         if (!found)
1488                                 *end = state->end;
1489                         goto out;
1490                 }
1491                 if (!found) {
1492                         *start = state->start;
1493                         *cached_state = state;
1494                         refcount_inc(&state->refs);
1495                 }
1496                 found++;
1497                 *end = state->end;
1498                 cur_start = state->end + 1;
1499                 node = rb_next(node);
1500                 total_bytes += state->end - state->start + 1;
1501                 if (total_bytes >= max_bytes)
1502                         break;
1503                 if (!node)
1504                         break;
1505         }
1506 out:
1507         spin_unlock(&tree->lock);
1508         return found;
1509 }
1510
1511 static int __process_pages_contig(struct address_space *mapping,
1512                                   struct page *locked_page,
1513                                   pgoff_t start_index, pgoff_t end_index,
1514                                   unsigned long page_ops, pgoff_t *index_ret);
1515
1516 static noinline void __unlock_for_delalloc(struct inode *inode,
1517                                            struct page *locked_page,
1518                                            u64 start, u64 end)
1519 {
1520         unsigned long index = start >> PAGE_SHIFT;
1521         unsigned long end_index = end >> PAGE_SHIFT;
1522
1523         ASSERT(locked_page);
1524         if (index == locked_page->index && end_index == index)
1525                 return;
1526
1527         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1528                                PAGE_UNLOCK, NULL);
1529 }
1530
1531 static noinline int lock_delalloc_pages(struct inode *inode,
1532                                         struct page *locked_page,
1533                                         u64 delalloc_start,
1534                                         u64 delalloc_end)
1535 {
1536         unsigned long index = delalloc_start >> PAGE_SHIFT;
1537         unsigned long index_ret = index;
1538         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1539         int ret;
1540
1541         ASSERT(locked_page);
1542         if (index == locked_page->index && index == end_index)
1543                 return 0;
1544
1545         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1546                                      end_index, PAGE_LOCK, &index_ret);
1547         if (ret == -EAGAIN)
1548                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1549                                       (u64)index_ret << PAGE_SHIFT);
1550         return ret;
1551 }
1552
1553 /*
1554  * find a contiguous range of bytes in the file marked as delalloc, not
1555  * more than 'max_bytes'.  start and end are used to return the range,
1556  *
1557  * 1 is returned if we find something, 0 if nothing was in the tree
1558  */
1559 static noinline_for_stack u64 find_lock_delalloc_range(struct inode *inode,
1560                                     struct extent_io_tree *tree,
1561                                     struct page *locked_page, u64 *start,
1562                                     u64 *end, u64 max_bytes)
1563 {
1564         u64 delalloc_start;
1565         u64 delalloc_end;
1566         u64 found;
1567         struct extent_state *cached_state = NULL;
1568         int ret;
1569         int loops = 0;
1570
1571 again:
1572         /* step one, find a bunch of delalloc bytes starting at start */
1573         delalloc_start = *start;
1574         delalloc_end = 0;
1575         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1576                                     max_bytes, &cached_state);
1577         if (!found || delalloc_end <= *start) {
1578                 *start = delalloc_start;
1579                 *end = delalloc_end;
1580                 free_extent_state(cached_state);
1581                 return 0;
1582         }
1583
1584         /*
1585          * start comes from the offset of locked_page.  We have to lock
1586          * pages in order, so we can't process delalloc bytes before
1587          * locked_page
1588          */
1589         if (delalloc_start < *start)
1590                 delalloc_start = *start;
1591
1592         /*
1593          * make sure to limit the number of pages we try to lock down
1594          */
1595         if (delalloc_end + 1 - delalloc_start > max_bytes)
1596                 delalloc_end = delalloc_start + max_bytes - 1;
1597
1598         /* step two, lock all the pages after the page that has start */
1599         ret = lock_delalloc_pages(inode, locked_page,
1600                                   delalloc_start, delalloc_end);
1601         if (ret == -EAGAIN) {
1602                 /* some of the pages are gone, lets avoid looping by
1603                  * shortening the size of the delalloc range we're searching
1604                  */
1605                 free_extent_state(cached_state);
1606                 cached_state = NULL;
1607                 if (!loops) {
1608                         max_bytes = PAGE_SIZE;
1609                         loops = 1;
1610                         goto again;
1611                 } else {
1612                         found = 0;
1613                         goto out_failed;
1614                 }
1615         }
1616         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1617
1618         /* step three, lock the state bits for the whole range */
1619         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1620
1621         /* then test to make sure it is all still delalloc */
1622         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1623                              EXTENT_DELALLOC, 1, cached_state);
1624         if (!ret) {
1625                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1626                                      &cached_state);
1627                 __unlock_for_delalloc(inode, locked_page,
1628                               delalloc_start, delalloc_end);
1629                 cond_resched();
1630                 goto again;
1631         }
1632         free_extent_state(cached_state);
1633         *start = delalloc_start;
1634         *end = delalloc_end;
1635 out_failed:
1636         return found;
1637 }
1638
1639 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1640 u64 btrfs_find_lock_delalloc_range(struct inode *inode,
1641                                     struct extent_io_tree *tree,
1642                                     struct page *locked_page, u64 *start,
1643                                     u64 *end, u64 max_bytes)
1644 {
1645         return find_lock_delalloc_range(inode, tree, locked_page, start, end,
1646                         max_bytes);
1647 }
1648 #endif
1649
1650 static int __process_pages_contig(struct address_space *mapping,
1651                                   struct page *locked_page,
1652                                   pgoff_t start_index, pgoff_t end_index,
1653                                   unsigned long page_ops, pgoff_t *index_ret)
1654 {
1655         unsigned long nr_pages = end_index - start_index + 1;
1656         unsigned long pages_locked = 0;
1657         pgoff_t index = start_index;
1658         struct page *pages[16];
1659         unsigned ret;
1660         int err = 0;
1661         int i;
1662
1663         if (page_ops & PAGE_LOCK) {
1664                 ASSERT(page_ops == PAGE_LOCK);
1665                 ASSERT(index_ret && *index_ret == start_index);
1666         }
1667
1668         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1669                 mapping_set_error(mapping, -EIO);
1670
1671         while (nr_pages > 0) {
1672                 ret = find_get_pages_contig(mapping, index,
1673                                      min_t(unsigned long,
1674                                      nr_pages, ARRAY_SIZE(pages)), pages);
1675                 if (ret == 0) {
1676                         /*
1677                          * Only if we're going to lock these pages,
1678                          * can we find nothing at @index.
1679                          */
1680                         ASSERT(page_ops & PAGE_LOCK);
1681                         err = -EAGAIN;
1682                         goto out;
1683                 }
1684
1685                 for (i = 0; i < ret; i++) {
1686                         if (page_ops & PAGE_SET_PRIVATE2)
1687                                 SetPagePrivate2(pages[i]);
1688
1689                         if (pages[i] == locked_page) {
1690                                 put_page(pages[i]);
1691                                 pages_locked++;
1692                                 continue;
1693                         }
1694                         if (page_ops & PAGE_CLEAR_DIRTY)
1695                                 clear_page_dirty_for_io(pages[i]);
1696                         if (page_ops & PAGE_SET_WRITEBACK)
1697                                 set_page_writeback(pages[i]);
1698                         if (page_ops & PAGE_SET_ERROR)
1699                                 SetPageError(pages[i]);
1700                         if (page_ops & PAGE_END_WRITEBACK)
1701                                 end_page_writeback(pages[i]);
1702                         if (page_ops & PAGE_UNLOCK)
1703                                 unlock_page(pages[i]);
1704                         if (page_ops & PAGE_LOCK) {
1705                                 lock_page(pages[i]);
1706                                 if (!PageDirty(pages[i]) ||
1707                                     pages[i]->mapping != mapping) {
1708                                         unlock_page(pages[i]);
1709                                         put_page(pages[i]);
1710                                         err = -EAGAIN;
1711                                         goto out;
1712                                 }
1713                         }
1714                         put_page(pages[i]);
1715                         pages_locked++;
1716                 }
1717                 nr_pages -= ret;
1718                 index += ret;
1719                 cond_resched();
1720         }
1721 out:
1722         if (err && index_ret)
1723                 *index_ret = start_index + pages_locked - 1;
1724         return err;
1725 }
1726
1727 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1728                                  u64 delalloc_end, struct page *locked_page,
1729                                  unsigned clear_bits,
1730                                  unsigned long page_ops)
1731 {
1732         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1733                          NULL);
1734
1735         __process_pages_contig(inode->i_mapping, locked_page,
1736                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1737                                page_ops, NULL);
1738 }
1739
1740 /*
1741  * count the number of bytes in the tree that have a given bit(s)
1742  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1743  * cached.  The total number found is returned.
1744  */
1745 u64 count_range_bits(struct extent_io_tree *tree,
1746                      u64 *start, u64 search_end, u64 max_bytes,
1747                      unsigned bits, int contig)
1748 {
1749         struct rb_node *node;
1750         struct extent_state *state;
1751         u64 cur_start = *start;
1752         u64 total_bytes = 0;
1753         u64 last = 0;
1754         int found = 0;
1755
1756         if (WARN_ON(search_end <= cur_start))
1757                 return 0;
1758
1759         spin_lock(&tree->lock);
1760         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1761                 total_bytes = tree->dirty_bytes;
1762                 goto out;
1763         }
1764         /*
1765          * this search will find all the extents that end after
1766          * our range starts.
1767          */
1768         node = tree_search(tree, cur_start);
1769         if (!node)
1770                 goto out;
1771
1772         while (1) {
1773                 state = rb_entry(node, struct extent_state, rb_node);
1774                 if (state->start > search_end)
1775                         break;
1776                 if (contig && found && state->start > last + 1)
1777                         break;
1778                 if (state->end >= cur_start && (state->state & bits) == bits) {
1779                         total_bytes += min(search_end, state->end) + 1 -
1780                                        max(cur_start, state->start);
1781                         if (total_bytes >= max_bytes)
1782                                 break;
1783                         if (!found) {
1784                                 *start = max(cur_start, state->start);
1785                                 found = 1;
1786                         }
1787                         last = state->end;
1788                 } else if (contig && found) {
1789                         break;
1790                 }
1791                 node = rb_next(node);
1792                 if (!node)
1793                         break;
1794         }
1795 out:
1796         spin_unlock(&tree->lock);
1797         return total_bytes;
1798 }
1799
1800 /*
1801  * set the private field for a given byte offset in the tree.  If there isn't
1802  * an extent_state there already, this does nothing.
1803  */
1804 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1805                 struct io_failure_record *failrec)
1806 {
1807         struct rb_node *node;
1808         struct extent_state *state;
1809         int ret = 0;
1810
1811         spin_lock(&tree->lock);
1812         /*
1813          * this search will find all the extents that end after
1814          * our range starts.
1815          */
1816         node = tree_search(tree, start);
1817         if (!node) {
1818                 ret = -ENOENT;
1819                 goto out;
1820         }
1821         state = rb_entry(node, struct extent_state, rb_node);
1822         if (state->start != start) {
1823                 ret = -ENOENT;
1824                 goto out;
1825         }
1826         state->failrec = failrec;
1827 out:
1828         spin_unlock(&tree->lock);
1829         return ret;
1830 }
1831
1832 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1833                 struct io_failure_record **failrec)
1834 {
1835         struct rb_node *node;
1836         struct extent_state *state;
1837         int ret = 0;
1838
1839         spin_lock(&tree->lock);
1840         /*
1841          * this search will find all the extents that end after
1842          * our range starts.
1843          */
1844         node = tree_search(tree, start);
1845         if (!node) {
1846                 ret = -ENOENT;
1847                 goto out;
1848         }
1849         state = rb_entry(node, struct extent_state, rb_node);
1850         if (state->start != start) {
1851                 ret = -ENOENT;
1852                 goto out;
1853         }
1854         *failrec = state->failrec;
1855 out:
1856         spin_unlock(&tree->lock);
1857         return ret;
1858 }
1859
1860 /*
1861  * searches a range in the state tree for a given mask.
1862  * If 'filled' == 1, this returns 1 only if every extent in the tree
1863  * has the bits set.  Otherwise, 1 is returned if any bit in the
1864  * range is found set.
1865  */
1866 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1867                    unsigned bits, int filled, struct extent_state *cached)
1868 {
1869         struct extent_state *state = NULL;
1870         struct rb_node *node;
1871         int bitset = 0;
1872
1873         spin_lock(&tree->lock);
1874         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1875             cached->end > start)
1876                 node = &cached->rb_node;
1877         else
1878                 node = tree_search(tree, start);
1879         while (node && start <= end) {
1880                 state = rb_entry(node, struct extent_state, rb_node);
1881
1882                 if (filled && state->start > start) {
1883                         bitset = 0;
1884                         break;
1885                 }
1886
1887                 if (state->start > end)
1888                         break;
1889
1890                 if (state->state & bits) {
1891                         bitset = 1;
1892                         if (!filled)
1893                                 break;
1894                 } else if (filled) {
1895                         bitset = 0;
1896                         break;
1897                 }
1898
1899                 if (state->end == (u64)-1)
1900                         break;
1901
1902                 start = state->end + 1;
1903                 if (start > end)
1904                         break;
1905                 node = rb_next(node);
1906                 if (!node) {
1907                         if (filled)
1908                                 bitset = 0;
1909                         break;
1910                 }
1911         }
1912         spin_unlock(&tree->lock);
1913         return bitset;
1914 }
1915
1916 /*
1917  * helper function to set a given page up to date if all the
1918  * extents in the tree for that page are up to date
1919  */
1920 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1921 {
1922         u64 start = page_offset(page);
1923         u64 end = start + PAGE_SIZE - 1;
1924         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1925                 SetPageUptodate(page);
1926 }
1927
1928 int free_io_failure(struct extent_io_tree *failure_tree,
1929                     struct extent_io_tree *io_tree,
1930                     struct io_failure_record *rec)
1931 {
1932         int ret;
1933         int err = 0;
1934
1935         set_state_failrec(failure_tree, rec->start, NULL);
1936         ret = clear_extent_bits(failure_tree, rec->start,
1937                                 rec->start + rec->len - 1,
1938                                 EXTENT_LOCKED | EXTENT_DIRTY);
1939         if (ret)
1940                 err = ret;
1941
1942         ret = clear_extent_bits(io_tree, rec->start,
1943                                 rec->start + rec->len - 1,
1944                                 EXTENT_DAMAGED);
1945         if (ret && !err)
1946                 err = ret;
1947
1948         kfree(rec);
1949         return err;
1950 }
1951
1952 /*
1953  * this bypasses the standard btrfs submit functions deliberately, as
1954  * the standard behavior is to write all copies in a raid setup. here we only
1955  * want to write the one bad copy. so we do the mapping for ourselves and issue
1956  * submit_bio directly.
1957  * to avoid any synchronization issues, wait for the data after writing, which
1958  * actually prevents the read that triggered the error from finishing.
1959  * currently, there can be no more than two copies of every data bit. thus,
1960  * exactly one rewrite is required.
1961  */
1962 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
1963                       u64 length, u64 logical, struct page *page,
1964                       unsigned int pg_offset, int mirror_num)
1965 {
1966         struct bio *bio;
1967         struct btrfs_device *dev;
1968         u64 map_length = 0;
1969         u64 sector;
1970         struct btrfs_bio *bbio = NULL;
1971         int ret;
1972
1973         ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
1974         BUG_ON(!mirror_num);
1975
1976         bio = btrfs_io_bio_alloc(1);
1977         bio->bi_iter.bi_size = 0;
1978         map_length = length;
1979
1980         /*
1981          * Avoid races with device replace and make sure our bbio has devices
1982          * associated to its stripes that don't go away while we are doing the
1983          * read repair operation.
1984          */
1985         btrfs_bio_counter_inc_blocked(fs_info);
1986         if (btrfs_is_parity_mirror(fs_info, logical, length)) {
1987                 /*
1988                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
1989                  * to update all raid stripes, but here we just want to correct
1990                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
1991                  * stripe's dev and sector.
1992                  */
1993                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
1994                                       &map_length, &bbio, 0);
1995                 if (ret) {
1996                         btrfs_bio_counter_dec(fs_info);
1997                         bio_put(bio);
1998                         return -EIO;
1999                 }
2000                 ASSERT(bbio->mirror_num == 1);
2001         } else {
2002                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2003                                       &map_length, &bbio, mirror_num);
2004                 if (ret) {
2005                         btrfs_bio_counter_dec(fs_info);
2006                         bio_put(bio);
2007                         return -EIO;
2008                 }
2009                 BUG_ON(mirror_num != bbio->mirror_num);
2010         }
2011
2012         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2013         bio->bi_iter.bi_sector = sector;
2014         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2015         btrfs_put_bbio(bbio);
2016         if (!dev || !dev->bdev ||
2017             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2018                 btrfs_bio_counter_dec(fs_info);
2019                 bio_put(bio);
2020                 return -EIO;
2021         }
2022         bio_set_dev(bio, dev->bdev);
2023         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2024         bio_add_page(bio, page, length, pg_offset);
2025
2026         if (btrfsic_submit_bio_wait(bio)) {
2027                 /* try to remap that extent elsewhere? */
2028                 btrfs_bio_counter_dec(fs_info);
2029                 bio_put(bio);
2030                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2031                 return -EIO;
2032         }
2033
2034         btrfs_info_rl_in_rcu(fs_info,
2035                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2036                                   ino, start,
2037                                   rcu_str_deref(dev->name), sector);
2038         btrfs_bio_counter_dec(fs_info);
2039         bio_put(bio);
2040         return 0;
2041 }
2042
2043 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2044                          struct extent_buffer *eb, int mirror_num)
2045 {
2046         u64 start = eb->start;
2047         int i, num_pages = num_extent_pages(eb);
2048         int ret = 0;
2049
2050         if (sb_rdonly(fs_info->sb))
2051                 return -EROFS;
2052
2053         for (i = 0; i < num_pages; i++) {
2054                 struct page *p = eb->pages[i];
2055
2056                 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2057                                         start - page_offset(p), mirror_num);
2058                 if (ret)
2059                         break;
2060                 start += PAGE_SIZE;
2061         }
2062
2063         return ret;
2064 }
2065
2066 /*
2067  * each time an IO finishes, we do a fast check in the IO failure tree
2068  * to see if we need to process or clean up an io_failure_record
2069  */
2070 int clean_io_failure(struct btrfs_fs_info *fs_info,
2071                      struct extent_io_tree *failure_tree,
2072                      struct extent_io_tree *io_tree, u64 start,
2073                      struct page *page, u64 ino, unsigned int pg_offset)
2074 {
2075         u64 private;
2076         struct io_failure_record *failrec;
2077         struct extent_state *state;
2078         int num_copies;
2079         int ret;
2080
2081         private = 0;
2082         ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2083                                EXTENT_DIRTY, 0);
2084         if (!ret)
2085                 return 0;
2086
2087         ret = get_state_failrec(failure_tree, start, &failrec);
2088         if (ret)
2089                 return 0;
2090
2091         BUG_ON(!failrec->this_mirror);
2092
2093         if (failrec->in_validation) {
2094                 /* there was no real error, just free the record */
2095                 btrfs_debug(fs_info,
2096                         "clean_io_failure: freeing dummy error at %llu",
2097                         failrec->start);
2098                 goto out;
2099         }
2100         if (sb_rdonly(fs_info->sb))
2101                 goto out;
2102
2103         spin_lock(&io_tree->lock);
2104         state = find_first_extent_bit_state(io_tree,
2105                                             failrec->start,
2106                                             EXTENT_LOCKED);
2107         spin_unlock(&io_tree->lock);
2108
2109         if (state && state->start <= failrec->start &&
2110             state->end >= failrec->start + failrec->len - 1) {
2111                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2112                                               failrec->len);
2113                 if (num_copies > 1)  {
2114                         repair_io_failure(fs_info, ino, start, failrec->len,
2115                                           failrec->logical, page, pg_offset,
2116                                           failrec->failed_mirror);
2117                 }
2118         }
2119
2120 out:
2121         free_io_failure(failure_tree, io_tree, failrec);
2122
2123         return 0;
2124 }
2125
2126 /*
2127  * Can be called when
2128  * - hold extent lock
2129  * - under ordered extent
2130  * - the inode is freeing
2131  */
2132 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2133 {
2134         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2135         struct io_failure_record *failrec;
2136         struct extent_state *state, *next;
2137
2138         if (RB_EMPTY_ROOT(&failure_tree->state))
2139                 return;
2140
2141         spin_lock(&failure_tree->lock);
2142         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2143         while (state) {
2144                 if (state->start > end)
2145                         break;
2146
2147                 ASSERT(state->end <= end);
2148
2149                 next = next_state(state);
2150
2151                 failrec = state->failrec;
2152                 free_extent_state(state);
2153                 kfree(failrec);
2154
2155                 state = next;
2156         }
2157         spin_unlock(&failure_tree->lock);
2158 }
2159
2160 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2161                 struct io_failure_record **failrec_ret)
2162 {
2163         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2164         struct io_failure_record *failrec;
2165         struct extent_map *em;
2166         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2167         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2168         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2169         int ret;
2170         u64 logical;
2171
2172         ret = get_state_failrec(failure_tree, start, &failrec);
2173         if (ret) {
2174                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2175                 if (!failrec)
2176                         return -ENOMEM;
2177
2178                 failrec->start = start;
2179                 failrec->len = end - start + 1;
2180                 failrec->this_mirror = 0;
2181                 failrec->bio_flags = 0;
2182                 failrec->in_validation = 0;
2183
2184                 read_lock(&em_tree->lock);
2185                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2186                 if (!em) {
2187                         read_unlock(&em_tree->lock);
2188                         kfree(failrec);
2189                         return -EIO;
2190                 }
2191
2192                 if (em->start > start || em->start + em->len <= start) {
2193                         free_extent_map(em);
2194                         em = NULL;
2195                 }
2196                 read_unlock(&em_tree->lock);
2197                 if (!em) {
2198                         kfree(failrec);
2199                         return -EIO;
2200                 }
2201
2202                 logical = start - em->start;
2203                 logical = em->block_start + logical;
2204                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2205                         logical = em->block_start;
2206                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2207                         extent_set_compress_type(&failrec->bio_flags,
2208                                                  em->compress_type);
2209                 }
2210
2211                 btrfs_debug(fs_info,
2212                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2213                         logical, start, failrec->len);
2214
2215                 failrec->logical = logical;
2216                 free_extent_map(em);
2217
2218                 /* set the bits in the private failure tree */
2219                 ret = set_extent_bits(failure_tree, start, end,
2220                                         EXTENT_LOCKED | EXTENT_DIRTY);
2221                 if (ret >= 0)
2222                         ret = set_state_failrec(failure_tree, start, failrec);
2223                 /* set the bits in the inode's tree */
2224                 if (ret >= 0)
2225                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2226                 if (ret < 0) {
2227                         kfree(failrec);
2228                         return ret;
2229                 }
2230         } else {
2231                 btrfs_debug(fs_info,
2232                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2233                         failrec->logical, failrec->start, failrec->len,
2234                         failrec->in_validation);
2235                 /*
2236                  * when data can be on disk more than twice, add to failrec here
2237                  * (e.g. with a list for failed_mirror) to make
2238                  * clean_io_failure() clean all those errors at once.
2239                  */
2240         }
2241
2242         *failrec_ret = failrec;
2243
2244         return 0;
2245 }
2246
2247 bool btrfs_check_repairable(struct inode *inode, unsigned failed_bio_pages,
2248                            struct io_failure_record *failrec, int failed_mirror)
2249 {
2250         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2251         int num_copies;
2252
2253         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2254         if (num_copies == 1) {
2255                 /*
2256                  * we only have a single copy of the data, so don't bother with
2257                  * all the retry and error correction code that follows. no
2258                  * matter what the error is, it is very likely to persist.
2259                  */
2260                 btrfs_debug(fs_info,
2261                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2262                         num_copies, failrec->this_mirror, failed_mirror);
2263                 return false;
2264         }
2265
2266         /*
2267          * there are two premises:
2268          *      a) deliver good data to the caller
2269          *      b) correct the bad sectors on disk
2270          */
2271         if (failed_bio_pages > 1) {
2272                 /*
2273                  * to fulfill b), we need to know the exact failing sectors, as
2274                  * we don't want to rewrite any more than the failed ones. thus,
2275                  * we need separate read requests for the failed bio
2276                  *
2277                  * if the following BUG_ON triggers, our validation request got
2278                  * merged. we need separate requests for our algorithm to work.
2279                  */
2280                 BUG_ON(failrec->in_validation);
2281                 failrec->in_validation = 1;
2282                 failrec->this_mirror = failed_mirror;
2283         } else {
2284                 /*
2285                  * we're ready to fulfill a) and b) alongside. get a good copy
2286                  * of the failed sector and if we succeed, we have setup
2287                  * everything for repair_io_failure to do the rest for us.
2288                  */
2289                 if (failrec->in_validation) {
2290                         BUG_ON(failrec->this_mirror != failed_mirror);
2291                         failrec->in_validation = 0;
2292                         failrec->this_mirror = 0;
2293                 }
2294                 failrec->failed_mirror = failed_mirror;
2295                 failrec->this_mirror++;
2296                 if (failrec->this_mirror == failed_mirror)
2297                         failrec->this_mirror++;
2298         }
2299
2300         if (failrec->this_mirror > num_copies) {
2301                 btrfs_debug(fs_info,
2302                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2303                         num_copies, failrec->this_mirror, failed_mirror);
2304                 return false;
2305         }
2306
2307         return true;
2308 }
2309
2310
2311 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2312                                     struct io_failure_record *failrec,
2313                                     struct page *page, int pg_offset, int icsum,
2314                                     bio_end_io_t *endio_func, void *data)
2315 {
2316         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2317         struct bio *bio;
2318         struct btrfs_io_bio *btrfs_failed_bio;
2319         struct btrfs_io_bio *btrfs_bio;
2320
2321         bio = btrfs_io_bio_alloc(1);
2322         bio->bi_end_io = endio_func;
2323         bio->bi_iter.bi_sector = failrec->logical >> 9;
2324         bio_set_dev(bio, fs_info->fs_devices->latest_bdev);
2325         bio->bi_iter.bi_size = 0;
2326         bio->bi_private = data;
2327
2328         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2329         if (btrfs_failed_bio->csum) {
2330                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2331
2332                 btrfs_bio = btrfs_io_bio(bio);
2333                 btrfs_bio->csum = btrfs_bio->csum_inline;
2334                 icsum *= csum_size;
2335                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2336                        csum_size);
2337         }
2338
2339         bio_add_page(bio, page, failrec->len, pg_offset);
2340
2341         return bio;
2342 }
2343
2344 /*
2345  * this is a generic handler for readpage errors (default
2346  * readpage_io_failed_hook). if other copies exist, read those and write back
2347  * good data to the failed position. does not investigate in remapping the
2348  * failed extent elsewhere, hoping the device will be smart enough to do this as
2349  * needed
2350  */
2351
2352 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2353                               struct page *page, u64 start, u64 end,
2354                               int failed_mirror)
2355 {
2356         struct io_failure_record *failrec;
2357         struct inode *inode = page->mapping->host;
2358         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2359         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2360         struct bio *bio;
2361         int read_mode = 0;
2362         blk_status_t status;
2363         int ret;
2364         unsigned failed_bio_pages = bio_pages_all(failed_bio);
2365
2366         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2367
2368         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2369         if (ret)
2370                 return ret;
2371
2372         if (!btrfs_check_repairable(inode, failed_bio_pages, failrec,
2373                                     failed_mirror)) {
2374                 free_io_failure(failure_tree, tree, failrec);
2375                 return -EIO;
2376         }
2377
2378         if (failed_bio_pages > 1)
2379                 read_mode |= REQ_FAILFAST_DEV;
2380
2381         phy_offset >>= inode->i_sb->s_blocksize_bits;
2382         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2383                                       start - page_offset(page),
2384                                       (int)phy_offset, failed_bio->bi_end_io,
2385                                       NULL);
2386         bio->bi_opf = REQ_OP_READ | read_mode;
2387
2388         btrfs_debug(btrfs_sb(inode->i_sb),
2389                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2390                 read_mode, failrec->this_mirror, failrec->in_validation);
2391
2392         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2393                                          failrec->bio_flags, 0);
2394         if (status) {
2395                 free_io_failure(failure_tree, tree, failrec);
2396                 bio_put(bio);
2397                 ret = blk_status_to_errno(status);
2398         }
2399
2400         return ret;
2401 }
2402
2403 /* lots and lots of room for performance fixes in the end_bio funcs */
2404
2405 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2406 {
2407         int uptodate = (err == 0);
2408         int ret = 0;
2409
2410         btrfs_writepage_endio_finish_ordered(page, start, end, uptodate);
2411
2412         if (!uptodate) {
2413                 ClearPageUptodate(page);
2414                 SetPageError(page);
2415                 ret = err < 0 ? err : -EIO;
2416                 mapping_set_error(page->mapping, ret);
2417         }
2418 }
2419
2420 /*
2421  * after a writepage IO is done, we need to:
2422  * clear the uptodate bits on error
2423  * clear the writeback bits in the extent tree for this IO
2424  * end_page_writeback if the page has no more pending IO
2425  *
2426  * Scheduling is not allowed, so the extent state tree is expected
2427  * to have one and only one object corresponding to this IO.
2428  */
2429 static void end_bio_extent_writepage(struct bio *bio)
2430 {
2431         int error = blk_status_to_errno(bio->bi_status);
2432         struct bio_vec *bvec;
2433         u64 start;
2434         u64 end;
2435         int i;
2436
2437         ASSERT(!bio_flagged(bio, BIO_CLONED));
2438         bio_for_each_segment_all(bvec, bio, i) {
2439                 struct page *page = bvec->bv_page;
2440                 struct inode *inode = page->mapping->host;
2441                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2442
2443                 /* We always issue full-page reads, but if some block
2444                  * in a page fails to read, blk_update_request() will
2445                  * advance bv_offset and adjust bv_len to compensate.
2446                  * Print a warning for nonzero offsets, and an error
2447                  * if they don't add up to a full page.  */
2448                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2449                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2450                                 btrfs_err(fs_info,
2451                                    "partial page write in btrfs with offset %u and length %u",
2452                                         bvec->bv_offset, bvec->bv_len);
2453                         else
2454                                 btrfs_info(fs_info,
2455                                    "incomplete page write in btrfs with offset %u and length %u",
2456                                         bvec->bv_offset, bvec->bv_len);
2457                 }
2458
2459                 start = page_offset(page);
2460                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2461
2462                 end_extent_writepage(page, error, start, end);
2463                 end_page_writeback(page);
2464         }
2465
2466         bio_put(bio);
2467 }
2468
2469 static void
2470 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2471                               int uptodate)
2472 {
2473         struct extent_state *cached = NULL;
2474         u64 end = start + len - 1;
2475
2476         if (uptodate && tree->track_uptodate)
2477                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2478         unlock_extent_cached_atomic(tree, start, end, &cached);
2479 }
2480
2481 /*
2482  * after a readpage IO is done, we need to:
2483  * clear the uptodate bits on error
2484  * set the uptodate bits if things worked
2485  * set the page up to date if all extents in the tree are uptodate
2486  * clear the lock bit in the extent tree
2487  * unlock the page if there are no other extents locked for it
2488  *
2489  * Scheduling is not allowed, so the extent state tree is expected
2490  * to have one and only one object corresponding to this IO.
2491  */
2492 static void end_bio_extent_readpage(struct bio *bio)
2493 {
2494         struct bio_vec *bvec;
2495         int uptodate = !bio->bi_status;
2496         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2497         struct extent_io_tree *tree, *failure_tree;
2498         u64 offset = 0;
2499         u64 start;
2500         u64 end;
2501         u64 len;
2502         u64 extent_start = 0;
2503         u64 extent_len = 0;
2504         int mirror;
2505         int ret;
2506         int i;
2507
2508         ASSERT(!bio_flagged(bio, BIO_CLONED));
2509         bio_for_each_segment_all(bvec, bio, i) {
2510                 struct page *page = bvec->bv_page;
2511                 struct inode *inode = page->mapping->host;
2512                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2513
2514                 btrfs_debug(fs_info,
2515                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2516                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2517                         io_bio->mirror_num);
2518                 tree = &BTRFS_I(inode)->io_tree;
2519                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2520
2521                 /* We always issue full-page reads, but if some block
2522                  * in a page fails to read, blk_update_request() will
2523                  * advance bv_offset and adjust bv_len to compensate.
2524                  * Print a warning for nonzero offsets, and an error
2525                  * if they don't add up to a full page.  */
2526                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2527                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2528                                 btrfs_err(fs_info,
2529                                         "partial page read in btrfs with offset %u and length %u",
2530                                         bvec->bv_offset, bvec->bv_len);
2531                         else
2532                                 btrfs_info(fs_info,
2533                                         "incomplete page read in btrfs with offset %u and length %u",
2534                                         bvec->bv_offset, bvec->bv_len);
2535                 }
2536
2537                 start = page_offset(page);
2538                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2539                 len = bvec->bv_len;
2540
2541                 mirror = io_bio->mirror_num;
2542                 if (likely(uptodate && tree->ops)) {
2543                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2544                                                               page, start, end,
2545                                                               mirror);
2546                         if (ret)
2547                                 uptodate = 0;
2548                         else
2549                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2550                                                  failure_tree, tree, start,
2551                                                  page,
2552                                                  btrfs_ino(BTRFS_I(inode)), 0);
2553                 }
2554
2555                 if (likely(uptodate))
2556                         goto readpage_ok;
2557
2558                 if (tree->ops) {
2559                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2560                         if (ret == -EAGAIN) {
2561                                 /*
2562                                  * Data inode's readpage_io_failed_hook() always
2563                                  * returns -EAGAIN.
2564                                  *
2565                                  * The generic bio_readpage_error handles errors
2566                                  * the following way: If possible, new read
2567                                  * requests are created and submitted and will
2568                                  * end up in end_bio_extent_readpage as well (if
2569                                  * we're lucky, not in the !uptodate case). In
2570                                  * that case it returns 0 and we just go on with
2571                                  * the next page in our bio. If it can't handle
2572                                  * the error it will return -EIO and we remain
2573                                  * responsible for that page.
2574                                  */
2575                                 ret = bio_readpage_error(bio, offset, page,
2576                                                          start, end, mirror);
2577                                 if (ret == 0) {
2578                                         uptodate = !bio->bi_status;
2579                                         offset += len;
2580                                         continue;
2581                                 }
2582                         }
2583
2584                         /*
2585                          * metadata's readpage_io_failed_hook() always returns
2586                          * -EIO and fixes nothing.  -EIO is also returned if
2587                          * data inode error could not be fixed.
2588                          */
2589                         ASSERT(ret == -EIO);
2590                 }
2591 readpage_ok:
2592                 if (likely(uptodate)) {
2593                         loff_t i_size = i_size_read(inode);
2594                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2595                         unsigned off;
2596
2597                         /* Zero out the end if this page straddles i_size */
2598                         off = i_size & (PAGE_SIZE-1);
2599                         if (page->index == end_index && off)
2600                                 zero_user_segment(page, off, PAGE_SIZE);
2601                         SetPageUptodate(page);
2602                 } else {
2603                         ClearPageUptodate(page);
2604                         SetPageError(page);
2605                 }
2606                 unlock_page(page);
2607                 offset += len;
2608
2609                 if (unlikely(!uptodate)) {
2610                         if (extent_len) {
2611                                 endio_readpage_release_extent(tree,
2612                                                               extent_start,
2613                                                               extent_len, 1);
2614                                 extent_start = 0;
2615                                 extent_len = 0;
2616                         }
2617                         endio_readpage_release_extent(tree, start,
2618                                                       end - start + 1, 0);
2619                 } else if (!extent_len) {
2620                         extent_start = start;
2621                         extent_len = end + 1 - start;
2622                 } else if (extent_start + extent_len == start) {
2623                         extent_len += end + 1 - start;
2624                 } else {
2625                         endio_readpage_release_extent(tree, extent_start,
2626                                                       extent_len, uptodate);
2627                         extent_start = start;
2628                         extent_len = end + 1 - start;
2629                 }
2630         }
2631
2632         if (extent_len)
2633                 endio_readpage_release_extent(tree, extent_start, extent_len,
2634                                               uptodate);
2635         if (io_bio->end_io)
2636                 io_bio->end_io(io_bio, blk_status_to_errno(bio->bi_status));
2637         bio_put(bio);
2638 }
2639
2640 /*
2641  * Initialize the members up to but not including 'bio'. Use after allocating a
2642  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2643  * 'bio' because use of __GFP_ZERO is not supported.
2644  */
2645 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2646 {
2647         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2648 }
2649
2650 /*
2651  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2652  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2653  * for the appropriate container_of magic
2654  */
2655 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte)
2656 {
2657         struct bio *bio;
2658
2659         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, &btrfs_bioset);
2660         bio_set_dev(bio, bdev);
2661         bio->bi_iter.bi_sector = first_byte >> 9;
2662         btrfs_io_bio_init(btrfs_io_bio(bio));
2663         return bio;
2664 }
2665
2666 struct bio *btrfs_bio_clone(struct bio *bio)
2667 {
2668         struct btrfs_io_bio *btrfs_bio;
2669         struct bio *new;
2670
2671         /* Bio allocation backed by a bioset does not fail */
2672         new = bio_clone_fast(bio, GFP_NOFS, &btrfs_bioset);
2673         btrfs_bio = btrfs_io_bio(new);
2674         btrfs_io_bio_init(btrfs_bio);
2675         btrfs_bio->iter = bio->bi_iter;
2676         return new;
2677 }
2678
2679 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2680 {
2681         struct bio *bio;
2682
2683         /* Bio allocation backed by a bioset does not fail */
2684         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, &btrfs_bioset);
2685         btrfs_io_bio_init(btrfs_io_bio(bio));
2686         return bio;
2687 }
2688
2689 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2690 {
2691         struct bio *bio;
2692         struct btrfs_io_bio *btrfs_bio;
2693
2694         /* this will never fail when it's backed by a bioset */
2695         bio = bio_clone_fast(orig, GFP_NOFS, &btrfs_bioset);
2696         ASSERT(bio);
2697
2698         btrfs_bio = btrfs_io_bio(bio);
2699         btrfs_io_bio_init(btrfs_bio);
2700
2701         bio_trim(bio, offset >> 9, size >> 9);
2702         btrfs_bio->iter = bio->bi_iter;
2703         return bio;
2704 }
2705
2706 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2707                                        unsigned long bio_flags)
2708 {
2709         blk_status_t ret = 0;
2710         struct bio_vec *bvec = bio_last_bvec_all(bio);
2711         struct page *page = bvec->bv_page;
2712         struct extent_io_tree *tree = bio->bi_private;
2713         u64 start;
2714
2715         start = page_offset(page) + bvec->bv_offset;
2716
2717         bio->bi_private = NULL;
2718
2719         if (tree->ops)
2720                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
2721                                            mirror_num, bio_flags, start);
2722         else
2723                 btrfsic_submit_bio(bio);
2724
2725         return blk_status_to_errno(ret);
2726 }
2727
2728 /*
2729  * @opf:        bio REQ_OP_* and REQ_* flags as one value
2730  * @tree:       tree so we can call our merge_bio hook
2731  * @wbc:        optional writeback control for io accounting
2732  * @page:       page to add to the bio
2733  * @pg_offset:  offset of the new bio or to check whether we are adding
2734  *              a contiguous page to the previous one
2735  * @size:       portion of page that we want to write
2736  * @offset:     starting offset in the page
2737  * @bdev:       attach newly created bios to this bdev
2738  * @bio_ret:    must be valid pointer, newly allocated bio will be stored there
2739  * @end_io_func:     end_io callback for new bio
2740  * @mirror_num:      desired mirror to read/write
2741  * @prev_bio_flags:  flags of previous bio to see if we can merge the current one
2742  * @bio_flags:  flags of the current bio to see if we can merge them
2743  */
2744 static int submit_extent_page(unsigned int opf, struct extent_io_tree *tree,
2745                               struct writeback_control *wbc,
2746                               struct page *page, u64 offset,
2747                               size_t size, unsigned long pg_offset,
2748                               struct block_device *bdev,
2749                               struct bio **bio_ret,
2750                               bio_end_io_t end_io_func,
2751                               int mirror_num,
2752                               unsigned long prev_bio_flags,
2753                               unsigned long bio_flags,
2754                               bool force_bio_submit)
2755 {
2756         int ret = 0;
2757         struct bio *bio;
2758         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2759         sector_t sector = offset >> 9;
2760
2761         ASSERT(bio_ret);
2762
2763         if (*bio_ret) {
2764                 bool contig;
2765                 bool can_merge = true;
2766
2767                 bio = *bio_ret;
2768                 if (prev_bio_flags & EXTENT_BIO_COMPRESSED)
2769                         contig = bio->bi_iter.bi_sector == sector;
2770                 else
2771                         contig = bio_end_sector(bio) == sector;
2772
2773                 if (tree->ops && btrfs_merge_bio_hook(page, offset, page_size,
2774                                                       bio, bio_flags))
2775                         can_merge = false;
2776
2777                 if (prev_bio_flags != bio_flags || !contig || !can_merge ||
2778                     force_bio_submit ||
2779                     bio_add_page(bio, page, page_size, pg_offset) < page_size) {
2780                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2781                         if (ret < 0) {
2782                                 *bio_ret = NULL;
2783                                 return ret;
2784                         }
2785                         bio = NULL;
2786                 } else {
2787                         if (wbc)
2788                                 wbc_account_io(wbc, page, page_size);
2789                         return 0;
2790                 }
2791         }
2792
2793         bio = btrfs_bio_alloc(bdev, offset);
2794         bio_add_page(bio, page, page_size, pg_offset);
2795         bio->bi_end_io = end_io_func;
2796         bio->bi_private = tree;
2797         bio->bi_write_hint = page->mapping->host->i_write_hint;
2798         bio->bi_opf = opf;
2799         if (wbc) {
2800                 wbc_init_bio(wbc, bio);
2801                 wbc_account_io(wbc, page, page_size);
2802         }
2803
2804         *bio_ret = bio;
2805
2806         return ret;
2807 }
2808
2809 static void attach_extent_buffer_page(struct extent_buffer *eb,
2810                                       struct page *page)
2811 {
2812         if (!PagePrivate(page)) {
2813                 SetPagePrivate(page);
2814                 get_page(page);
2815                 set_page_private(page, (unsigned long)eb);
2816         } else {
2817                 WARN_ON(page->private != (unsigned long)eb);
2818         }
2819 }
2820
2821 void set_page_extent_mapped(struct page *page)
2822 {
2823         if (!PagePrivate(page)) {
2824                 SetPagePrivate(page);
2825                 get_page(page);
2826                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2827         }
2828 }
2829
2830 static struct extent_map *
2831 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2832                  u64 start, u64 len, get_extent_t *get_extent,
2833                  struct extent_map **em_cached)
2834 {
2835         struct extent_map *em;
2836
2837         if (em_cached && *em_cached) {
2838                 em = *em_cached;
2839                 if (extent_map_in_tree(em) && start >= em->start &&
2840                     start < extent_map_end(em)) {
2841                         refcount_inc(&em->refs);
2842                         return em;
2843                 }
2844
2845                 free_extent_map(em);
2846                 *em_cached = NULL;
2847         }
2848
2849         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2850         if (em_cached && !IS_ERR_OR_NULL(em)) {
2851                 BUG_ON(*em_cached);
2852                 refcount_inc(&em->refs);
2853                 *em_cached = em;
2854         }
2855         return em;
2856 }
2857 /*
2858  * basic readpage implementation.  Locked extent state structs are inserted
2859  * into the tree that are removed when the IO is done (by the end_io
2860  * handlers)
2861  * XXX JDM: This needs looking at to ensure proper page locking
2862  * return 0 on success, otherwise return error
2863  */
2864 static int __do_readpage(struct extent_io_tree *tree,
2865                          struct page *page,
2866                          get_extent_t *get_extent,
2867                          struct extent_map **em_cached,
2868                          struct bio **bio, int mirror_num,
2869                          unsigned long *bio_flags, unsigned int read_flags,
2870                          u64 *prev_em_start)
2871 {
2872         struct inode *inode = page->mapping->host;
2873         u64 start = page_offset(page);
2874         const u64 end = start + PAGE_SIZE - 1;
2875         u64 cur = start;
2876         u64 extent_offset;
2877         u64 last_byte = i_size_read(inode);
2878         u64 block_start;
2879         u64 cur_end;
2880         struct extent_map *em;
2881         struct block_device *bdev;
2882         int ret = 0;
2883         int nr = 0;
2884         size_t pg_offset = 0;
2885         size_t iosize;
2886         size_t disk_io_size;
2887         size_t blocksize = inode->i_sb->s_blocksize;
2888         unsigned long this_bio_flag = 0;
2889
2890         set_page_extent_mapped(page);
2891
2892         if (!PageUptodate(page)) {
2893                 if (cleancache_get_page(page) == 0) {
2894                         BUG_ON(blocksize != PAGE_SIZE);
2895                         unlock_extent(tree, start, end);
2896                         goto out;
2897                 }
2898         }
2899
2900         if (page->index == last_byte >> PAGE_SHIFT) {
2901                 char *userpage;
2902                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2903
2904                 if (zero_offset) {
2905                         iosize = PAGE_SIZE - zero_offset;
2906                         userpage = kmap_atomic(page);
2907                         memset(userpage + zero_offset, 0, iosize);
2908                         flush_dcache_page(page);
2909                         kunmap_atomic(userpage);
2910                 }
2911         }
2912         while (cur <= end) {
2913                 bool force_bio_submit = false;
2914                 u64 offset;
2915
2916                 if (cur >= last_byte) {
2917                         char *userpage;
2918                         struct extent_state *cached = NULL;
2919
2920                         iosize = PAGE_SIZE - pg_offset;
2921                         userpage = kmap_atomic(page);
2922                         memset(userpage + pg_offset, 0, iosize);
2923                         flush_dcache_page(page);
2924                         kunmap_atomic(userpage);
2925                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2926                                             &cached, GFP_NOFS);
2927                         unlock_extent_cached(tree, cur,
2928                                              cur + iosize - 1, &cached);
2929                         break;
2930                 }
2931                 em = __get_extent_map(inode, page, pg_offset, cur,
2932                                       end - cur + 1, get_extent, em_cached);
2933                 if (IS_ERR_OR_NULL(em)) {
2934                         SetPageError(page);
2935                         unlock_extent(tree, cur, end);
2936                         break;
2937                 }
2938                 extent_offset = cur - em->start;
2939                 BUG_ON(extent_map_end(em) <= cur);
2940                 BUG_ON(end < cur);
2941
2942                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2943                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2944                         extent_set_compress_type(&this_bio_flag,
2945                                                  em->compress_type);
2946                 }
2947
2948                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2949                 cur_end = min(extent_map_end(em) - 1, end);
2950                 iosize = ALIGN(iosize, blocksize);
2951                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2952                         disk_io_size = em->block_len;
2953                         offset = em->block_start;
2954                 } else {
2955                         offset = em->block_start + extent_offset;
2956                         disk_io_size = iosize;
2957                 }
2958                 bdev = em->bdev;
2959                 block_start = em->block_start;
2960                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2961                         block_start = EXTENT_MAP_HOLE;
2962
2963                 /*
2964                  * If we have a file range that points to a compressed extent
2965                  * and it's followed by a consecutive file range that points to
2966                  * to the same compressed extent (possibly with a different
2967                  * offset and/or length, so it either points to the whole extent
2968                  * or only part of it), we must make sure we do not submit a
2969                  * single bio to populate the pages for the 2 ranges because
2970                  * this makes the compressed extent read zero out the pages
2971                  * belonging to the 2nd range. Imagine the following scenario:
2972                  *
2973                  *  File layout
2974                  *  [0 - 8K]                     [8K - 24K]
2975                  *    |                               |
2976                  *    |                               |
2977                  * points to extent X,         points to extent X,
2978                  * offset 4K, length of 8K     offset 0, length 16K
2979                  *
2980                  * [extent X, compressed length = 4K uncompressed length = 16K]
2981                  *
2982                  * If the bio to read the compressed extent covers both ranges,
2983                  * it will decompress extent X into the pages belonging to the
2984                  * first range and then it will stop, zeroing out the remaining
2985                  * pages that belong to the other range that points to extent X.
2986                  * So here we make sure we submit 2 bios, one for the first
2987                  * range and another one for the third range. Both will target
2988                  * the same physical extent from disk, but we can't currently
2989                  * make the compressed bio endio callback populate the pages
2990                  * for both ranges because each compressed bio is tightly
2991                  * coupled with a single extent map, and each range can have
2992                  * an extent map with a different offset value relative to the
2993                  * uncompressed data of our extent and different lengths. This
2994                  * is a corner case so we prioritize correctness over
2995                  * non-optimal behavior (submitting 2 bios for the same extent).
2996                  */
2997                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
2998                     prev_em_start && *prev_em_start != (u64)-1 &&
2999                     *prev_em_start != em->orig_start)
3000                         force_bio_submit = true;
3001
3002                 if (prev_em_start)
3003                         *prev_em_start = em->orig_start;
3004
3005                 free_extent_map(em);
3006                 em = NULL;
3007
3008                 /* we've found a hole, just zero and go on */
3009                 if (block_start == EXTENT_MAP_HOLE) {
3010                         char *userpage;
3011                         struct extent_state *cached = NULL;
3012
3013                         userpage = kmap_atomic(page);
3014                         memset(userpage + pg_offset, 0, iosize);
3015                         flush_dcache_page(page);
3016                         kunmap_atomic(userpage);
3017
3018                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3019                                             &cached, GFP_NOFS);
3020                         unlock_extent_cached(tree, cur,
3021                                              cur + iosize - 1, &cached);
3022                         cur = cur + iosize;
3023                         pg_offset += iosize;
3024                         continue;
3025                 }
3026                 /* the get_extent function already copied into the page */
3027                 if (test_range_bit(tree, cur, cur_end,
3028                                    EXTENT_UPTODATE, 1, NULL)) {
3029                         check_page_uptodate(tree, page);
3030                         unlock_extent(tree, cur, cur + iosize - 1);
3031                         cur = cur + iosize;
3032                         pg_offset += iosize;
3033                         continue;
3034                 }
3035                 /* we have an inline extent but it didn't get marked up
3036                  * to date.  Error out
3037                  */
3038                 if (block_start == EXTENT_MAP_INLINE) {
3039                         SetPageError(page);
3040                         unlock_extent(tree, cur, cur + iosize - 1);
3041                         cur = cur + iosize;
3042                         pg_offset += iosize;
3043                         continue;
3044                 }
3045
3046                 ret = submit_extent_page(REQ_OP_READ | read_flags, tree, NULL,
3047                                          page, offset, disk_io_size,
3048                                          pg_offset, bdev, bio,
3049                                          end_bio_extent_readpage, mirror_num,
3050                                          *bio_flags,
3051                                          this_bio_flag,
3052                                          force_bio_submit);
3053                 if (!ret) {
3054                         nr++;
3055                         *bio_flags = this_bio_flag;
3056                 } else {
3057                         SetPageError(page);
3058                         unlock_extent(tree, cur, cur + iosize - 1);
3059                         goto out;
3060                 }
3061                 cur = cur + iosize;
3062                 pg_offset += iosize;
3063         }
3064 out:
3065         if (!nr) {
3066                 if (!PageError(page))
3067                         SetPageUptodate(page);
3068                 unlock_page(page);
3069         }
3070         return ret;
3071 }
3072
3073 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3074                                              struct page *pages[], int nr_pages,
3075                                              u64 start, u64 end,
3076                                              struct extent_map **em_cached,
3077                                              struct bio **bio,
3078                                              unsigned long *bio_flags,
3079                                              u64 *prev_em_start)
3080 {
3081         struct inode *inode;
3082         struct btrfs_ordered_extent *ordered;
3083         int index;
3084
3085         inode = pages[0]->mapping->host;
3086         while (1) {
3087                 lock_extent(tree, start, end);
3088                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3089                                                      end - start + 1);
3090                 if (!ordered)
3091                         break;
3092                 unlock_extent(tree, start, end);
3093                 btrfs_start_ordered_extent(inode, ordered, 1);
3094                 btrfs_put_ordered_extent(ordered);
3095         }
3096
3097         for (index = 0; index < nr_pages; index++) {
3098                 __do_readpage(tree, pages[index], btrfs_get_extent, em_cached,
3099                                 bio, 0, bio_flags, REQ_RAHEAD, prev_em_start);
3100                 put_page(pages[index]);
3101         }
3102 }
3103
3104 static void __extent_readpages(struct extent_io_tree *tree,
3105                                struct page *pages[],
3106                                int nr_pages,
3107                                struct extent_map **em_cached,
3108                                struct bio **bio, unsigned long *bio_flags,
3109                                u64 *prev_em_start)
3110 {
3111         u64 start = 0;
3112         u64 end = 0;
3113         u64 page_start;
3114         int index;
3115         int first_index = 0;
3116
3117         for (index = 0; index < nr_pages; index++) {
3118        &nb