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