fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18 #include "volumes.h"
  19 #include "check-integrity.h"
  20 #include "locking.h"
  21 #include "rcu-string.h"
  22 #include "backref.h"
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26 static struct bio_set *btrfs_bioset;
  27
  28 #ifdef CONFIG_BTRFS_DEBUG
  29 static LIST_HEAD(buffers);
  30 static LIST_HEAD(states);
  31
  32 static DEFINE_SPINLOCK(leak_lock);
  33
  34 static inline
  35 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  36 {
  37         unsigned long flags;
  38
  39         spin_lock_irqsave(&leak_lock, flags);
  40         list_add(new, head);
  41         spin_unlock_irqrestore(&leak_lock, flags);
  42 }
  43
  44 static inline
  45 void btrfs_leak_debug_del(struct list_head *entry)
  46 {
  47         unsigned long flags;
  48
  49         spin_lock_irqsave(&leak_lock, flags);
  50         list_del(entry);
  51         spin_unlock_irqrestore(&leak_lock, flags);
  52 }
  53
  54 static inline
  55 void btrfs_leak_debug_check(void)
  56 {
  57         struct extent_state *state;
  58         struct extent_buffer *eb;
  59
  60         while (!list_empty(&states)) {
  61                 state = list_entry(states.next, struct extent_state, leak_list);
  62                 printk(KERN_ERR "BTRFS: state leak: start %llu end %llu "
  63                        "state %lu in tree %p refs %d\n",
  64                        state->start, state->end, state->state, state->tree,
  65                        atomic_read(&state->refs));
  66                 list_del(&state->leak_list);
  67                 kmem_cache_free(extent_state_cache, state);
  68         }
  69
  70         while (!list_empty(&buffers)) {
  71                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  72                 printk(KERN_ERR "BTRFS: buffer leak start %llu len %lu "
  73                        "refs %d\n",
  74                        eb->start, eb->len, atomic_read(&eb->refs));
  75                 list_del(&eb->leak_list);
  76                 kmem_cache_free(extent_buffer_cache, eb);
  77         }
  78 }
  79
  80 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  81         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  82 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  83                 struct extent_io_tree *tree, u64 start, u64 end)
  84 {
  85         struct inode *inode;
  86         u64 isize;
  87
  88         if (!tree->mapping)
  89                 return;
  90
  91         inode = tree->mapping->host;
  92         isize = i_size_read(inode);
  93         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  94                 printk_ratelimited(KERN_DEBUG
  95                     "BTRFS: %s: ino %llu isize %llu odd range [%llu,%llu]\n",
  96                                 caller, btrfs_ino(inode), isize, start, end);
  97         }
  98 }
  99 #else
 100 #define btrfs_leak_debug_add(new, head) do {} while (0)
 101 #define btrfs_leak_debug_del(entry)     do {} while (0)
 102 #define btrfs_leak_debug_check()        do {} while (0)
 103 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 104 #endif
 105
 106 #define BUFFER_LRU_MAX 64
 107
 108 struct tree_entry {
 109         u64 start;
 110         u64 end;
 111         struct rb_node rb_node;
 112 };
 113
 114 struct extent_page_data {
 115         struct bio *bio;
 116         struct extent_io_tree *tree;
 117         get_extent_t *get_extent;
 118         unsigned long bio_flags;
 119
 120         /* tells writepage not to lock the state bits for this range
 121          * it still does the unlocking
 122          */
 123         unsigned int extent_locked:1;
 124
 125         /* tells the submit_bio code to use a WRITE_SYNC */
 126         unsigned int sync_io:1;
 127 };
 128
 129 static noinline void flush_write_bio(void *data);
 130 static inline struct btrfs_fs_info *
 131 tree_fs_info(struct extent_io_tree *tree)
 132 {
 133         if (!tree->mapping)
 134                 return NULL;
 135         return btrfs_sb(tree->mapping->host->i_sb);
 136 }
 137
 138 int __init extent_io_init(void)
 139 {
 140         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 141                         sizeof(struct extent_state), 0,
 142                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 143         if (!extent_state_cache)
 144                 return -ENOMEM;
 145
 146         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 147                         sizeof(struct extent_buffer), 0,
 148                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 149         if (!extent_buffer_cache)
 150                 goto free_state_cache;
 151
 152         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 153                                      offsetof(struct btrfs_io_bio, bio));
 154         if (!btrfs_bioset)
 155                 goto free_buffer_cache;
 156
 157         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 158                 goto free_bioset;
 159
 160         return 0;
 161
 162 free_bioset:
 163         bioset_free(btrfs_bioset);
 164         btrfs_bioset = NULL;
 165
 166 free_buffer_cache:
 167         kmem_cache_destroy(extent_buffer_cache);
 168         extent_buffer_cache = NULL;
 169
 170 free_state_cache:
 171         kmem_cache_destroy(extent_state_cache);
 172         extent_state_cache = NULL;
 173         return -ENOMEM;
 174 }
 175
 176 void extent_io_exit(void)
 177 {
 178         btrfs_leak_debug_check();
 179
 180         /*
 181          * Make sure all delayed rcu free are flushed before we
 182          * destroy caches.
 183          */
 184         rcu_barrier();
 185         if (extent_state_cache)
 186                 kmem_cache_destroy(extent_state_cache);
 187         if (extent_buffer_cache)
 188                 kmem_cache_destroy(extent_buffer_cache);
 189         if (btrfs_bioset)
 190                 bioset_free(btrfs_bioset);
 191 }
 192
 193 void extent_io_tree_init(struct extent_io_tree *tree,
 194                          struct address_space *mapping)
 195 {
 196         tree->state = RB_ROOT;
 197         tree->ops = NULL;
 198         tree->dirty_bytes = 0;
 199         spin_lock_init(&tree->lock);
 200         tree->mapping = mapping;
 201 }
 202
 203 static struct extent_state *alloc_extent_state(gfp_t mask)
 204 {
 205         struct extent_state *state;
 206
 207         state = kmem_cache_alloc(extent_state_cache, mask);
 208         if (!state)
 209                 return state;
 210         state->state = 0;
 211         state->private = 0;
 212         state->tree = NULL;
 213         btrfs_leak_debug_add(&state->leak_list, &states);
 214         atomic_set(&state->refs, 1);
 215         init_waitqueue_head(&state->wq);
 216         trace_alloc_extent_state(state, mask, _RET_IP_);
 217         return state;
 218 }
 219
 220 void free_extent_state(struct extent_state *state)
 221 {
 222         if (!state)
 223                 return;
 224         if (atomic_dec_and_test(&state->refs)) {
 225                 WARN_ON(state->tree);
 226                 btrfs_leak_debug_del(&state->leak_list);
 227                 trace_free_extent_state(state, _RET_IP_);
 228                 kmem_cache_free(extent_state_cache, state);
 229         }
 230 }
 231
 232 static struct rb_node *tree_insert(struct rb_root *root,
 233                                    struct rb_node *search_start,
 234                                    u64 offset,
 235                                    struct rb_node *node,
 236                                    struct rb_node ***p_in,
 237                                    struct rb_node **parent_in)
 238 {
 239         struct rb_node **p;
 240         struct rb_node *parent = NULL;
 241         struct tree_entry *entry;
 242
 243         if (p_in && parent_in) {
 244                 p = *p_in;
 245                 parent = *parent_in;
 246                 goto do_insert;
 247         }
 248
 249         p = search_start ? &search_start : &root->rb_node;
 250         while (*p) {
 251                 parent = *p;
 252                 entry = rb_entry(parent, struct tree_entry, rb_node);
 253
 254                 if (offset < entry->start)
 255                         p = &(*p)->rb_left;
 256                 else if (offset > entry->end)
 257                         p = &(*p)->rb_right;
 258                 else
 259                         return parent;
 260         }
 261
 262 do_insert:
 263         rb_link_node(node, parent, p);
 264         rb_insert_color(node, root);
 265         return NULL;
 266 }
 267
 268 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 269                                       struct rb_node **prev_ret,
 270                                       struct rb_node **next_ret,
 271                                       struct rb_node ***p_ret,
 272                                       struct rb_node **parent_ret)
 273 {
 274         struct rb_root *root = &tree->state;
 275         struct rb_node **n = &root->rb_node;
 276         struct rb_node *prev = NULL;
 277         struct rb_node *orig_prev = NULL;
 278         struct tree_entry *entry;
 279         struct tree_entry *prev_entry = NULL;
 280
 281         while (*n) {
 282                 prev = *n;
 283                 entry = rb_entry(prev, struct tree_entry, rb_node);
 284                 prev_entry = entry;
 285
 286                 if (offset < entry->start)
 287                         n = &(*n)->rb_left;
 288                 else if (offset > entry->end)
 289                         n = &(*n)->rb_right;
 290                 else
 291                         return *n;
 292         }
 293
 294         if (p_ret)
 295                 *p_ret = n;
 296         if (parent_ret)
 297                 *parent_ret = prev;
 298
 299         if (prev_ret) {
 300                 orig_prev = prev;
 301                 while (prev && offset > prev_entry->end) {
 302                         prev = rb_next(prev);
 303                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 304                 }
 305                 *prev_ret = prev;
 306                 prev = orig_prev;
 307         }
 308
 309         if (next_ret) {
 310                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 311                 while (prev && offset < prev_entry->start) {
 312                         prev = rb_prev(prev);
 313                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 314                 }
 315                 *next_ret = prev;
 316         }
 317         return NULL;
 318 }
 319
 320 static inline struct rb_node *
 321 tree_search_for_insert(struct extent_io_tree *tree,
 322                        u64 offset,
 323                        struct rb_node ***p_ret,
 324                        struct rb_node **parent_ret)
 325 {
 326         struct rb_node *prev = NULL;
 327         struct rb_node *ret;
 328
 329         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 330         if (!ret)
 331                 return prev;
 332         return ret;
 333 }
 334
 335 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 336                                           u64 offset)
 337 {
 338         return tree_search_for_insert(tree, offset, NULL, NULL);
 339 }
 340
 341 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 342                      struct extent_state *other)
 343 {
 344         if (tree->ops && tree->ops->merge_extent_hook)
 345                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 346                                              other);
 347 }
 348
 349 /*
 350  * utility function to look for merge candidates inside a given range.
 351  * Any extents with matching state are merged together into a single
 352  * extent in the tree.  Extents with EXTENT_IO in their state field
 353  * are not merged because the end_io handlers need to be able to do
 354  * operations on them without sleeping (or doing allocations/splits).
 355  *
 356  * This should be called with the tree lock held.
 357  */
 358 static void merge_state(struct extent_io_tree *tree,
 359                         struct extent_state *state)
 360 {
 361         struct extent_state *other;
 362         struct rb_node *other_node;
 363
 364         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 365                 return;
 366
 367         other_node = rb_prev(&state->rb_node);
 368         if (other_node) {
 369                 other = rb_entry(other_node, struct extent_state, rb_node);
 370                 if (other->end == state->start - 1 &&
 371                     other->state == state->state) {
 372                         merge_cb(tree, state, other);
 373                         state->start = other->start;
 374                         other->tree = NULL;
 375                         rb_erase(&other->rb_node, &tree->state);
 376                         free_extent_state(other);
 377                 }
 378         }
 379         other_node = rb_next(&state->rb_node);
 380         if (other_node) {
 381                 other = rb_entry(other_node, struct extent_state, rb_node);
 382                 if (other->start == state->end + 1 &&
 383                     other->state == state->state) {
 384                         merge_cb(tree, state, other);
 385                         state->end = other->end;
 386                         other->tree = NULL;
 387                         rb_erase(&other->rb_node, &tree->state);
 388                         free_extent_state(other);
 389                 }
 390         }
 391 }
 392
 393 static void set_state_cb(struct extent_io_tree *tree,
 394                          struct extent_state *state, unsigned long *bits)
 395 {
 396         if (tree->ops && tree->ops->set_bit_hook)
 397                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 398 }
 399
 400 static void clear_state_cb(struct extent_io_tree *tree,
 401                            struct extent_state *state, unsigned long *bits)
 402 {
 403         if (tree->ops && tree->ops->clear_bit_hook)
 404                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 405 }
 406
 407 static void set_state_bits(struct extent_io_tree *tree,
 408                            struct extent_state *state, unsigned long *bits);
 409
 410 /*
 411  * insert an extent_state struct into the tree.  'bits' are set on the
 412  * struct before it is inserted.
 413  *
 414  * This may return -EEXIST if the extent is already there, in which case the
 415  * state struct is freed.
 416  *
 417  * The tree lock is not taken internally.  This is a utility function and
 418  * probably isn't what you want to call (see set/clear_extent_bit).
 419  */
 420 static int insert_state(struct extent_io_tree *tree,
 421                         struct extent_state *state, u64 start, u64 end,
 422                         struct rb_node ***p,
 423                         struct rb_node **parent,
 424                         unsigned long *bits)
 425 {
 426         struct rb_node *node;
 427
 428         if (end < start)
 429                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 430                        end, start);
 431         state->start = start;
 432         state->end = end;
 433
 434         set_state_bits(tree, state, bits);
 435
 436         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 437         if (node) {
 438                 struct extent_state *found;
 439                 found = rb_entry(node, struct extent_state, rb_node);
 440                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
 441                        "%llu %llu\n",
 442                        found->start, found->end, start, end);
 443                 return -EEXIST;
 444         }
 445         state->tree = tree;
 446         merge_state(tree, state);
 447         return 0;
 448 }
 449
 450 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 451                      u64 split)
 452 {
 453         if (tree->ops && tree->ops->split_extent_hook)
 454                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 455 }
 456
 457 /*
 458  * split a given extent state struct in two, inserting the preallocated
 459  * struct 'prealloc' as the newly created second half.  'split' indicates an
 460  * offset inside 'orig' where it should be split.
 461  *
 462  * Before calling,
 463  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 464  * are two extent state structs in the tree:
 465  * prealloc: [orig->start, split - 1]
 466  * orig: [ split, orig->end ]
 467  *
 468  * The tree locks are not taken by this function. They need to be held
 469  * by the caller.
 470  */
 471 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 472                        struct extent_state *prealloc, u64 split)
 473 {
 474         struct rb_node *node;
 475
 476         split_cb(tree, orig, split);
 477
 478         prealloc->start = orig->start;
 479         prealloc->end = split - 1;
 480         prealloc->state = orig->state;
 481         orig->start = split;
 482
 483         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 484                            &prealloc->rb_node, NULL, NULL);
 485         if (node) {
 486                 free_extent_state(prealloc);
 487                 return -EEXIST;
 488         }
 489         prealloc->tree = tree;
 490         return 0;
 491 }
 492
 493 static struct extent_state *next_state(struct extent_state *state)
 494 {
 495         struct rb_node *next = rb_next(&state->rb_node);
 496         if (next)
 497                 return rb_entry(next, struct extent_state, rb_node);
 498         else
 499                 return NULL;
 500 }
 501
 502 /*
 503  * utility function to clear some bits in an extent state struct.
 504  * it will optionally wake up any one waiting on this state (wake == 1).
 505  *
 506  * If no bits are set on the state struct after clearing things, the
 507  * struct is freed and removed from the tree
 508  */
 509 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 510                                             struct extent_state *state,
 511                                             unsigned long *bits, int wake)
 512 {
 513         struct extent_state *next;
 514         unsigned long bits_to_clear = *bits & ~EXTENT_CTLBITS;
 515
 516         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 517                 u64 range = state->end - state->start + 1;
 518                 WARN_ON(range > tree->dirty_bytes);
 519                 tree->dirty_bytes -= range;
 520         }
 521         clear_state_cb(tree, state, bits);
 522         state->state &= ~bits_to_clear;
 523         if (wake)
 524                 wake_up(&state->wq);
 525         if (state->state == 0) {
 526                 next = next_state(state);
 527                 if (state->tree) {
 528                         rb_erase(&state->rb_node, &tree->state);
 529                         state->tree = NULL;
 530                         free_extent_state(state);
 531                 } else {
 532                         WARN_ON(1);
 533                 }
 534         } else {
 535                 merge_state(tree, state);
 536                 next = next_state(state);
 537         }
 538         return next;
 539 }
 540
 541 static struct extent_state *
 542 alloc_extent_state_atomic(struct extent_state *prealloc)
 543 {
 544         if (!prealloc)
 545                 prealloc = alloc_extent_state(GFP_ATOMIC);
 546
 547         return prealloc;
 548 }
 549
 550 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 551 {
 552         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 553                     "Extent tree was modified by another "
 554                     "thread while locked.");
 555 }
 556
 557 /*
 558  * clear some bits on a range in the tree.  This may require splitting
 559  * or inserting elements in the tree, so the gfp mask is used to
 560  * indicate which allocations or sleeping are allowed.
 561  *
 562  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 563  * the given range from the tree regardless of state (ie for truncate).
 564  *
 565  * the range [start, end] is inclusive.
 566  *
 567  * This takes the tree lock, and returns 0 on success and < 0 on error.
 568  */
 569 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 570                      unsigned long bits, int wake, int delete,
 571                      struct extent_state **cached_state,
 572                      gfp_t mask)
 573 {
 574         struct extent_state *state;
 575         struct extent_state *cached;
 576         struct extent_state *prealloc = NULL;
 577         struct rb_node *node;
 578         u64 last_end;
 579         int err;
 580         int clear = 0;
 581
 582         btrfs_debug_check_extent_io_range(tree, start, end);
 583
 584         if (bits & EXTENT_DELALLOC)
 585                 bits |= EXTENT_NORESERVE;
 586
 587         if (delete)
 588                 bits |= ~EXTENT_CTLBITS;
 589         bits |= EXTENT_FIRST_DELALLOC;
 590
 591         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 592                 clear = 1;
 593 again:
 594         if (!prealloc && (mask & __GFP_WAIT)) {
 595                 prealloc = alloc_extent_state(mask);
 596                 if (!prealloc)
 597                         return -ENOMEM;
 598         }
 599
 600         spin_lock(&tree->lock);
 601         if (cached_state) {
 602                 cached = *cached_state;
 603
 604                 if (clear) {
 605                         *cached_state = NULL;
 606                         cached_state = NULL;
 607                 }
 608
 609                 if (cached && cached->tree && cached->start <= start &&
 610                     cached->end > start) {
 611                         if (clear)
 612                                 atomic_dec(&cached->refs);
 613                         state = cached;
 614                         goto hit_next;
 615                 }
 616                 if (clear)
 617                         free_extent_state(cached);
 618         }
 619         /*
 620          * this search will find the extents that end after
 621          * our range starts
 622          */
 623         node = tree_search(tree, start);
 624         if (!node)
 625                 goto out;
 626         state = rb_entry(node, struct extent_state, rb_node);
 627 hit_next:
 628         if (state->start > end)
 629                 goto out;
 630         WARN_ON(state->end < start);
 631         last_end = state->end;
 632
 633         /* the state doesn't have the wanted bits, go ahead */
 634         if (!(state->state & bits)) {
 635                 state = next_state(state);
 636                 goto next;
 637         }
 638
 639         /*
 640          *     | ---- desired range ---- |
 641          *  | state | or
 642          *  | ------------- state -------------- |
 643          *
 644          * We need to split the extent we found, and may flip
 645          * bits on second half.
 646          *
 647          * If the extent we found extends past our range, we
 648          * just split and search again.  It'll get split again
 649          * the next time though.
 650          *
 651          * If the extent we found is inside our range, we clear
 652          * the desired bit on it.
 653          */
 654
 655         if (state->start < start) {
 656                 prealloc = alloc_extent_state_atomic(prealloc);
 657                 BUG_ON(!prealloc);
 658                 err = split_state(tree, state, prealloc, start);
 659                 if (err)
 660                         extent_io_tree_panic(tree, err);
 661
 662                 prealloc = NULL;
 663                 if (err)
 664                         goto out;
 665                 if (state->end <= end) {
 666                         state = clear_state_bit(tree, state, &bits, wake);
 667                         goto next;
 668                 }
 669                 goto search_again;
 670         }
 671         /*
 672          * | ---- desired range ---- |
 673          *                        | state |
 674          * We need to split the extent, and clear the bit
 675          * on the first half
 676          */
 677         if (state->start <= end && state->end > end) {
 678                 prealloc = alloc_extent_state_atomic(prealloc);
 679                 BUG_ON(!prealloc);
 680                 err = split_state(tree, state, prealloc, end + 1);
 681                 if (err)
 682                         extent_io_tree_panic(tree, err);
 683
 684                 if (wake)
 685                         wake_up(&state->wq);
 686
 687                 clear_state_bit(tree, prealloc, &bits, wake);
 688
 689                 prealloc = NULL;
 690                 goto out;
 691         }
 692
 693         state = clear_state_bit(tree, state, &bits, wake);
 694 next:
 695         if (last_end == (u64)-1)
 696                 goto out;
 697         start = last_end + 1;
 698         if (start <= end && state && !need_resched())
 699                 goto hit_next;
 700         goto search_again;
 701
 702 out:
 703         spin_unlock(&tree->lock);
 704         if (prealloc)
 705                 free_extent_state(prealloc);
 706
 707         return 0;
 708
 709 search_again:
 710         if (start > end)
 711                 goto out;
 712         spin_unlock(&tree->lock);
 713         if (mask & __GFP_WAIT)
 714                 cond_resched();
 715         goto again;
 716 }
 717
 718 static void wait_on_state(struct extent_io_tree *tree,
 719                           struct extent_state *state)
 720                 __releases(tree->lock)
 721                 __acquires(tree->lock)
 722 {
 723         DEFINE_WAIT(wait);
 724         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 725         spin_unlock(&tree->lock);
 726         schedule();
 727         spin_lock(&tree->lock);
 728         finish_wait(&state->wq, &wait);
 729 }
 730
 731 /*
 732  * waits for one or more bits to clear on a range in the state tree.
 733  * The range [start, end] is inclusive.
 734  * The tree lock is taken by this function
 735  */
 736 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 737                             unsigned long bits)
 738 {
 739         struct extent_state *state;
 740         struct rb_node *node;
 741
 742         btrfs_debug_check_extent_io_range(tree, start, end);
 743
 744         spin_lock(&tree->lock);
 745 again:
 746         while (1) {
 747                 /*
 748                  * this search will find all the extents that end after
 749                  * our range starts
 750                  */
 751                 node = tree_search(tree, start);
 752 process_node:
 753                 if (!node)
 754                         break;
 755
 756                 state = rb_entry(node, struct extent_state, rb_node);
 757
 758                 if (state->start > end)
 759                         goto out;
 760
 761                 if (state->state & bits) {
 762                         start = state->start;
 763                         atomic_inc(&state->refs);
 764                         wait_on_state(tree, state);
 765                         free_extent_state(state);
 766                         goto again;
 767                 }
 768                 start = state->end + 1;
 769
 770                 if (start > end)
 771                         break;
 772
 773                 if (!cond_resched_lock(&tree->lock)) {
 774                         node = rb_next(node);
 775                         goto process_node;
 776                 }
 777         }
 778 out:
 779         spin_unlock(&tree->lock);
 780 }
 781
 782 static void set_state_bits(struct extent_io_tree *tree,
 783                            struct extent_state *state,
 784                            unsigned long *bits)
 785 {
 786         unsigned long bits_to_set = *bits & ~EXTENT_CTLBITS;
 787
 788         set_state_cb(tree, state, bits);
 789         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 790                 u64 range = state->end - state->start + 1;
 791                 tree->dirty_bytes += range;
 792         }
 793         state->state |= bits_to_set;
 794 }
 795
 796 static void cache_state(struct extent_state *state,
 797                         struct extent_state **cached_ptr)
 798 {
 799         if (cached_ptr && !(*cached_ptr)) {
 800                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 801                         *cached_ptr = state;
 802                         atomic_inc(&state->refs);
 803                 }
 804         }
 805 }
 806
 807 /*
 808  * set some bits on a range in the tree.  This may require allocations or
 809  * sleeping, so the gfp mask is used to indicate what is allowed.
 810  *
 811  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 812  * part of the range already has the desired bits set.  The start of the
 813  * existing range is returned in failed_start in this case.
 814  *
 815  * [start, end] is inclusive This takes the tree lock.
 816  */
 817
 818 static int __must_check
 819 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 820                  unsigned long bits, unsigned long exclusive_bits,
 821                  u64 *failed_start, struct extent_state **cached_state,
 822                  gfp_t mask)
 823 {
 824         struct extent_state *state;
 825         struct extent_state *prealloc = NULL;
 826         struct rb_node *node;
 827         struct rb_node **p;
 828         struct rb_node *parent;
 829         int err = 0;
 830         u64 last_start;
 831         u64 last_end;
 832
 833         btrfs_debug_check_extent_io_range(tree, start, end);
 834
 835         bits |= EXTENT_FIRST_DELALLOC;
 836 again:
 837         if (!prealloc && (mask & __GFP_WAIT)) {
 838                 prealloc = alloc_extent_state(mask);
 839                 BUG_ON(!prealloc);
 840         }
 841
 842         spin_lock(&tree->lock);
 843         if (cached_state && *cached_state) {
 844                 state = *cached_state;
 845                 if (state->start <= start && state->end > start &&
 846                     state->tree) {
 847                         node = &state->rb_node;
 848                         goto hit_next;
 849                 }
 850         }
 851         /*
 852          * this search will find all the extents that end after
 853          * our range starts.
 854          */
 855         node = tree_search_for_insert(tree, start, &p, &parent);
 856         if (!node) {
 857                 prealloc = alloc_extent_state_atomic(prealloc);
 858                 BUG_ON(!prealloc);
 859                 err = insert_state(tree, prealloc, start, end,
 860                                    &p, &parent, &bits);
 861                 if (err)
 862                         extent_io_tree_panic(tree, err);
 863
 864                 cache_state(prealloc, cached_state);
 865                 prealloc = NULL;
 866                 goto out;
 867         }
 868         state = rb_entry(node, struct extent_state, rb_node);
 869 hit_next:
 870         last_start = state->start;
 871         last_end = state->end;
 872
 873         /*
 874          * | ---- desired range ---- |
 875          * | state |
 876          *
 877          * Just lock what we found and keep going
 878          */
 879         if (state->start == start && state->end <= end) {
 880                 if (state->state & exclusive_bits) {
 881                         *failed_start = state->start;
 882                         err = -EEXIST;
 883                         goto out;
 884                 }
 885
 886                 set_state_bits(tree, state, &bits);
 887                 cache_state(state, cached_state);
 888                 merge_state(tree, state);
 889                 if (last_end == (u64)-1)
 890                         goto out;
 891                 start = last_end + 1;
 892                 state = next_state(state);
 893                 if (start < end && state && state->start == start &&
 894                     !need_resched())
 895                         goto hit_next;
 896                 goto search_again;
 897         }
 898
 899         /*
 900          *     | ---- desired range ---- |
 901          * | state |
 902          *   or
 903          * | ------------- state -------------- |
 904          *
 905          * We need to split the extent we found, and may flip bits on
 906          * second half.
 907          *
 908          * If the extent we found extends past our
 909          * range, we just split and search again.  It'll get split
 910          * again the next time though.
 911          *
 912          * If the extent we found is inside our range, we set the
 913          * desired bit on it.
 914          */
 915         if (state->start < start) {
 916                 if (state->state & exclusive_bits) {
 917                         *failed_start = start;
 918                         err = -EEXIST;
 919                         goto out;
 920                 }
 921
 922                 prealloc = alloc_extent_state_atomic(prealloc);
 923                 BUG_ON(!prealloc);
 924                 err = split_state(tree, state, prealloc, start);
 925                 if (err)
 926                         extent_io_tree_panic(tree, err);
 927
 928                 prealloc = NULL;
 929                 if (err)
 930                         goto out;
 931                 if (state->end <= end) {
 932                         set_state_bits(tree, state, &bits);
 933                         cache_state(state, cached_state);
 934                         merge_state(tree, state);
 935                         if (last_end == (u64)-1)
 936                                 goto out;
 937                         start = last_end + 1;
 938                         state = next_state(state);
 939                         if (start < end && state && state->start == start &&
 940                             !need_resched())
 941                                 goto hit_next;
 942                 }
 943                 goto search_again;
 944         }
 945         /*
 946          * | ---- desired range ---- |
 947          *     | state | or               | state |
 948          *
 949          * There's a hole, we need to insert something in it and
 950          * ignore the extent we found.
 951          */
 952         if (state->start > start) {
 953                 u64 this_end;
 954                 if (end < last_start)
 955                         this_end = end;
 956                 else
 957                         this_end = last_start - 1;
 958
 959                 prealloc = alloc_extent_state_atomic(prealloc);
 960                 BUG_ON(!prealloc);
 961
 962                 /*
 963                  * Avoid to free 'prealloc' if it can be merged with
 964                  * the later extent.
 965                  */
 966                 err = insert_state(tree, prealloc, start, this_end,
 967                                    NULL, NULL, &bits);
 968                 if (err)
 969                         extent_io_tree_panic(tree, err);
 970
 971                 cache_state(prealloc, cached_state);
 972                 prealloc = NULL;
 973                 start = this_end + 1;
 974                 goto search_again;
 975         }
 976         /*
 977          * | ---- desired range ---- |
 978          *                        | state |
 979          * We need to split the extent, and set the bit
 980          * on the first half
 981          */
 982         if (state->start <= end && state->end > end) {
 983                 if (state->state & exclusive_bits) {
 984                         *failed_start = start;
 985                         err = -EEXIST;
 986                         goto out;
 987                 }
 988
 989                 prealloc = alloc_extent_state_atomic(prealloc);
 990                 BUG_ON(!prealloc);
 991                 err = split_state(tree, state, prealloc, end + 1);
 992                 if (err)
 993                         extent_io_tree_panic(tree, err);
 994
 995                 set_state_bits(tree, prealloc, &bits);
 996                 cache_state(prealloc, cached_state);
 997                 merge_state(tree, prealloc);
 998                 prealloc = NULL;
 999                 goto out;
1000         }
1001
1002         goto search_again;
1003
1004 out:
1005         spin_unlock(&tree->lock);
1006         if (prealloc)
1007                 free_extent_state(prealloc);
1008
1009         return err;
1010
1011 search_again:
1012         if (start > end)
1013                 goto out;
1014         spin_unlock(&tree->lock);
1015         if (mask & __GFP_WAIT)
1016                 cond_resched();
1017         goto again;
1018 }
1019
1020 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1021                    unsigned long bits, u64 * failed_start,
1022                    struct extent_state **cached_state, gfp_t mask)
1023 {
1024         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1025                                 cached_state, mask);
1026 }
1027
1028
1029 /**
1030  * convert_extent_bit - convert all bits in a given range from one bit to
1031  *                      another
1032  * @tree:       the io tree to search
1033  * @start:      the start offset in bytes
1034  * @end:        the end offset in bytes (inclusive)
1035  * @bits:       the bits to set in this range
1036  * @clear_bits: the bits to clear in this range
1037  * @cached_state:       state that we're going to cache
1038  * @mask:       the allocation mask
1039  *
1040  * This will go through and set bits for the given range.  If any states exist
1041  * already in this range they are set with the given bit and cleared of the
1042  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1043  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1044  * boundary bits like LOCK.
1045  */
1046 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1047                        unsigned long bits, unsigned long clear_bits,
1048                        struct extent_state **cached_state, gfp_t mask)
1049 {
1050         struct extent_state *state;
1051         struct extent_state *prealloc = NULL;
1052         struct rb_node *node;
1053         struct rb_node **p;
1054         struct rb_node *parent;
1055         int err = 0;
1056         u64 last_start;
1057         u64 last_end;
1058
1059         btrfs_debug_check_extent_io_range(tree, start, end);
1060
1061 again:
1062         if (!prealloc && (mask & __GFP_WAIT)) {
1063                 prealloc = alloc_extent_state(mask);
1064                 if (!prealloc)
1065                         return -ENOMEM;
1066         }
1067
1068         spin_lock(&tree->lock);
1069         if (cached_state && *cached_state) {
1070                 state = *cached_state;
1071                 if (state->start <= start && state->end > start &&
1072                     state->tree) {
1073                         node = &state->rb_node;
1074                         goto hit_next;
1075                 }
1076         }
1077
1078         /*
1079          * this search will find all the extents that end after
1080          * our range starts.
1081          */
1082         node = tree_search_for_insert(tree, start, &p, &parent);
1083         if (!node) {
1084                 prealloc = alloc_extent_state_atomic(prealloc);
1085                 if (!prealloc) {
1086                         err = -ENOMEM;
1087                         goto out;
1088                 }
1089                 err = insert_state(tree, prealloc, start, end,
1090                                    &p, &parent, &bits);
1091                 if (err)
1092                         extent_io_tree_panic(tree, err);
1093                 cache_state(prealloc, cached_state);
1094                 prealloc = NULL;
1095                 goto out;
1096         }
1097         state = rb_entry(node, struct extent_state, rb_node);
1098 hit_next:
1099         last_start = state->start;
1100         last_end = state->end;
1101
1102         /*
1103          * | ---- desired range ---- |
1104          * | state |
1105          *
1106          * Just lock what we found and keep going
1107          */
1108         if (state->start == start && state->end <= end) {
1109                 set_state_bits(tree, state, &bits);
1110                 cache_state(state, cached_state);
1111                 state = clear_state_bit(tree, state, &clear_bits, 0);
1112                 if (last_end == (u64)-1)
1113                         goto out;
1114                 start = last_end + 1;
1115                 if (start < end && state && state->start == start &&
1116                     !need_resched())
1117                         goto hit_next;
1118                 goto search_again;
1119         }
1120
1121         /*
1122          *     | ---- desired range ---- |
1123          * | state |
1124          *   or
1125          * | ------------- state -------------- |
1126          *
1127          * We need to split the extent we found, and may flip bits on
1128          * second half.
1129          *
1130          * If the extent we found extends past our
1131          * range, we just split and search again.  It'll get split
1132          * again the next time though.
1133          *
1134          * If the extent we found is inside our range, we set the
1135          * desired bit on it.
1136          */
1137         if (state->start < start) {
1138                 prealloc = alloc_extent_state_atomic(prealloc);
1139                 if (!prealloc) {
1140                         err = -ENOMEM;
1141                         goto out;
1142                 }
1143                 err = split_state(tree, state, prealloc, start);
1144                 if (err)
1145                         extent_io_tree_panic(tree, err);
1146                 prealloc = NULL;
1147                 if (err)
1148                         goto out;
1149                 if (state->end <= end) {
1150                         set_state_bits(tree, state, &bits);
1151                         cache_state(state, cached_state);
1152                         state = clear_state_bit(tree, state, &clear_bits, 0);
1153                         if (last_end == (u64)-1)
1154                                 goto out;
1155                         start = last_end + 1;
1156                         if (start < end && state && state->start == start &&
1157                             !need_resched())
1158                                 goto hit_next;
1159                 }
1160                 goto search_again;
1161         }
1162         /*
1163          * | ---- desired range ---- |
1164          *     | state | or               | state |
1165          *
1166          * There's a hole, we need to insert something in it and
1167          * ignore the extent we found.
1168          */
1169         if (state->start > start) {
1170                 u64 this_end;
1171                 if (end < last_start)
1172                         this_end = end;
1173                 else
1174                         this_end = last_start - 1;
1175
1176                 prealloc = alloc_extent_state_atomic(prealloc);
1177                 if (!prealloc) {
1178                         err = -ENOMEM;
1179                         goto out;
1180                 }
1181
1182                 /*
1183                  * Avoid to free 'prealloc' if it can be merged with
1184                  * the later extent.
1185                  */
1186                 err = insert_state(tree, prealloc, start, this_end,
1187                                    NULL, NULL, &bits);
1188                 if (err)
1189                         extent_io_tree_panic(tree, err);
1190                 cache_state(prealloc, cached_state);
1191                 prealloc = NULL;
1192                 start = this_end + 1;
1193                 goto search_again;
1194         }
1195         /*
1196          * | ---- desired range ---- |
1197          *                        | state |
1198          * We need to split the extent, and set the bit
1199          * on the first half
1200          */
1201         if (state->start <= end && state->end > end) {
1202                 prealloc = alloc_extent_state_atomic(prealloc);
1203                 if (!prealloc) {
1204                         err = -ENOMEM;
1205                         goto out;
1206                 }
1207
1208                 err = split_state(tree, state, prealloc, end + 1);
1209                 if (err)
1210                         extent_io_tree_panic(tree, err);
1211
1212                 set_state_bits(tree, prealloc, &bits);
1213                 cache_state(prealloc, cached_state);
1214                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1215                 prealloc = NULL;
1216                 goto out;
1217         }
1218
1219         goto search_again;
1220
1221 out:
1222         spin_unlock(&tree->lock);
1223         if (prealloc)
1224                 free_extent_state(prealloc);
1225
1226         return err;
1227
1228 search_again:
1229         if (start > end)
1230                 goto out;
1231         spin_unlock(&tree->lock);
1232         if (mask & __GFP_WAIT)
1233                 cond_resched();
1234         goto again;
1235 }
1236
1237 /* wrappers around set/clear extent bit */
1238 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1239                      gfp_t mask)
1240 {
1241         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1242                               NULL, mask);
1243 }
1244
1245 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1246                     unsigned long bits, gfp_t mask)
1247 {
1248         return set_extent_bit(tree, start, end, bits, NULL,
1249                               NULL, mask);
1250 }
1251
1252 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1253                       unsigned long bits, gfp_t mask)
1254 {
1255         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1256 }
1257
1258 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1259                         struct extent_state **cached_state, gfp_t mask)
1260 {
1261         return set_extent_bit(tree, start, end,
1262                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1263                               NULL, cached_state, mask);
1264 }
1265
1266 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1267                       struct extent_state **cached_state, gfp_t mask)
1268 {
1269         return set_extent_bit(tree, start, end,
1270                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1271                               NULL, cached_state, mask);
1272 }
1273
1274 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1275                        gfp_t mask)
1276 {
1277         return clear_extent_bit(tree, start, end,
1278                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1279                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1280 }
1281
1282 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1283                      gfp_t mask)
1284 {
1285         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1286                               NULL, mask);
1287 }
1288
1289 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1290                         struct extent_state **cached_state, gfp_t mask)
1291 {
1292         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1293                               cached_state, mask);
1294 }
1295
1296 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1297                           struct extent_state **cached_state, gfp_t mask)
1298 {
1299         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1300                                 cached_state, mask);
1301 }
1302
1303 /*
1304  * either insert or lock state struct between start and end use mask to tell
1305  * us if waiting is desired.
1306  */
1307 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1308                      unsigned long bits, struct extent_state **cached_state)
1309 {
1310         int err;
1311         u64 failed_start;
1312         while (1) {
1313                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1314                                        EXTENT_LOCKED, &failed_start,
1315                                        cached_state, GFP_NOFS);
1316                 if (err == -EEXIST) {
1317                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1318                         start = failed_start;
1319                 } else
1320                         break;
1321                 WARN_ON(start > end);
1322         }
1323         return err;
1324 }
1325
1326 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1327 {
1328         return lock_extent_bits(tree, start, end, 0, NULL);
1329 }
1330
1331 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1332 {
1333         int err;
1334         u64 failed_start;
1335
1336         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1337                                &failed_start, NULL, GFP_NOFS);
1338         if (err == -EEXIST) {
1339                 if (failed_start > start)
1340                         clear_extent_bit(tree, start, failed_start - 1,
1341                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1342                 return 0;
1343         }
1344         return 1;
1345 }
1346
1347 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1348                          struct extent_state **cached, gfp_t mask)
1349 {
1350         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1351                                 mask);
1352 }
1353
1354 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1355 {
1356         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1357                                 GFP_NOFS);
1358 }
1359
1360 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1361 {
1362         unsigned long index = start >> PAGE_CACHE_SHIFT;
1363         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1364         struct page *page;
1365
1366         while (index <= end_index) {
1367                 page = find_get_page(inode->i_mapping, index);
1368                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1369                 clear_page_dirty_for_io(page);
1370                 page_cache_release(page);
1371                 index++;
1372         }
1373         return 0;
1374 }
1375
1376 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1377 {
1378         unsigned long index = start >> PAGE_CACHE_SHIFT;
1379         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1380         struct page *page;
1381
1382         while (index <= end_index) {
1383                 page = find_get_page(inode->i_mapping, index);
1384                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1385                 account_page_redirty(page);
1386                 __set_page_dirty_nobuffers(page);
1387                 page_cache_release(page);
1388                 index++;
1389         }
1390         return 0;
1391 }
1392
1393 /*
1394  * helper function to set both pages and extents in the tree writeback
1395  */
1396 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1397 {
1398         unsigned long index = start >> PAGE_CACHE_SHIFT;
1399         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1400         struct page *page;
1401
1402         while (index <= end_index) {
1403                 page = find_get_page(tree->mapping, index);
1404                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1405                 set_page_writeback(page);
1406                 page_cache_release(page);
1407                 index++;
1408         }
1409         return 0;
1410 }
1411
1412 /* find the first state struct with 'bits' set after 'start', and
1413  * return it.  tree->lock must be held.  NULL will returned if
1414  * nothing was found after 'start'
1415  */
1416 static struct extent_state *
1417 find_first_extent_bit_state(struct extent_io_tree *tree,
1418                             u64 start, unsigned long bits)
1419 {
1420         struct rb_node *node;
1421         struct extent_state *state;
1422
1423         /*
1424          * this search will find all the extents that end after
1425          * our range starts.
1426          */
1427         node = tree_search(tree, start);
1428         if (!node)
1429                 goto out;
1430
1431         while (1) {
1432                 state = rb_entry(node, struct extent_state, rb_node);
1433                 if (state->end >= start && (state->state & bits))
1434                         return state;
1435
1436                 node = rb_next(node);
1437                 if (!node)
1438                         break;
1439         }
1440 out:
1441         return NULL;
1442 }
1443
1444 /*
1445  * find the first offset in the io tree with 'bits' set. zero is
1446  * returned if we find something, and *start_ret and *end_ret are
1447  * set to reflect the state struct that was found.
1448  *
1449  * If nothing was found, 1 is returned. If found something, return 0.
1450  */
1451 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1452                           u64 *start_ret, u64 *end_ret, unsigned long bits,
1453                           struct extent_state **cached_state)
1454 {
1455         struct extent_state *state;
1456         struct rb_node *n;
1457         int ret = 1;
1458
1459         spin_lock(&tree->lock);
1460         if (cached_state && *cached_state) {
1461                 state = *cached_state;
1462                 if (state->end == start - 1 && state->tree) {
1463                         n = rb_next(&state->rb_node);
1464                         while (n) {
1465                                 state = rb_entry(n, struct extent_state,
1466                                                  rb_node);
1467                                 if (state->state & bits)
1468                                         goto got_it;
1469                                 n = rb_next(n);
1470                         }
1471                         free_extent_state(*cached_state);
1472                         *cached_state = NULL;
1473                         goto out;
1474                 }
1475                 free_extent_state(*cached_state);
1476                 *cached_state = NULL;
1477         }
1478
1479         state = find_first_extent_bit_state(tree, start, bits);
1480 got_it:
1481         if (state) {
1482                 cache_state(state, cached_state);
1483                 *start_ret = state->start;
1484                 *end_ret = state->end;
1485                 ret = 0;
1486         }
1487 out:
1488         spin_unlock(&tree->lock);
1489         return ret;
1490 }
1491
1492 /*
1493  * find a contiguous range of bytes in the file marked as delalloc, not
1494  * more than 'max_bytes'.  start and end are used to return the range,
1495  *
1496  * 1 is returned if we find something, 0 if nothing was in the tree
1497  */
1498 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1499                                         u64 *start, u64 *end, u64 max_bytes,
1500                                         struct extent_state **cached_state)
1501 {
1502         struct rb_node *node;
1503         struct extent_state *state;
1504         u64 cur_start = *start;
1505         u64 found = 0;
1506         u64 total_bytes = 0;
1507
1508         spin_lock(&tree->lock);
1509
1510         /*
1511          * this search will find all the extents that end after
1512          * our range starts.
1513          */
1514         node = tree_search(tree, cur_start);
1515         if (!node) {
1516                 if (!found)
1517                         *end = (u64)-1;
1518                 goto out;
1519         }
1520
1521         while (1) {
1522                 state = rb_entry(node, struct extent_state, rb_node);
1523                 if (found && (state->start != cur_start ||
1524                               (state->state & EXTENT_BOUNDARY))) {
1525                         goto out;
1526                 }
1527                 if (!(state->state & EXTENT_DELALLOC)) {
1528                         if (!found)
1529                                 *end = state->end;
1530                         goto out;
1531                 }
1532                 if (!found) {
1533                         *start = state->start;
1534                         *cached_state = state;
1535                         atomic_inc(&state->refs);
1536                 }
1537                 found++;
1538                 *end = state->end;
1539                 cur_start = state->end + 1;
1540                 node = rb_next(node);
1541                 total_bytes += state->end - state->start + 1;
1542                 if (total_bytes >= max_bytes)
1543                         break;
1544                 if (!node)
1545                         break;
1546         }
1547 out:
1548         spin_unlock(&tree->lock);
1549         return found;
1550 }
1551
1552 static noinline void __unlock_for_delalloc(struct inode *inode,
1553                                            struct page *locked_page,
1554                                            u64 start, u64 end)
1555 {
1556         int ret;
1557         struct page *pages[16];
1558         unsigned long index = start >> PAGE_CACHE_SHIFT;
1559         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1560         unsigned long nr_pages = end_index - index + 1;
1561         int i;
1562
1563         if (index == locked_page->index && end_index == index)
1564                 return;
1565
1566         while (nr_pages > 0) {
1567                 ret = find_get_pages_contig(inode->i_mapping, index,
1568                                      min_t(unsigned long, nr_pages,
1569                                      ARRAY_SIZE(pages)), pages);
1570                 for (i = 0; i < ret; i++) {
1571                         if (pages[i] != locked_page)
1572                                 unlock_page(pages[i]);
1573                         page_cache_release(pages[i]);
1574                 }
1575                 nr_pages -= ret;
1576                 index += ret;
1577                 cond_resched();
1578         }
1579 }
1580
1581 static noinline int lock_delalloc_pages(struct inode *inode,
1582                                         struct page *locked_page,
1583                                         u64 delalloc_start,
1584                                         u64 delalloc_end)
1585 {
1586         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1587         unsigned long start_index = index;
1588         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1589         unsigned long pages_locked = 0;
1590         struct page *pages[16];
1591         unsigned long nrpages;
1592         int ret;
1593         int i;
1594
1595         /* the caller is responsible for locking the start index */
1596         if (index == locked_page->index && index == end_index)
1597                 return 0;
1598
1599         /* skip the page at the start index */
1600         nrpages = end_index - index + 1;
1601         while (nrpages > 0) {
1602                 ret = find_get_pages_contig(inode->i_mapping, index,
1603                                      min_t(unsigned long,
1604                                      nrpages, ARRAY_SIZE(pages)), pages);
1605                 if (ret == 0) {
1606                         ret = -EAGAIN;
1607                         goto done;
1608                 }
1609                 /* now we have an array of pages, lock them all */
1610                 for (i = 0; i < ret; i++) {
1611                         /*
1612                          * the caller is taking responsibility for
1613                          * locked_page
1614                          */
1615                         if (pages[i] != locked_page) {
1616                                 lock_page(pages[i]);
1617                                 if (!PageDirty(pages[i]) ||
1618                                     pages[i]->mapping != inode->i_mapping) {
1619                                         ret = -EAGAIN;
1620                                         unlock_page(pages[i]);
1621                                         page_cache_release(pages[i]);
1622                                         goto done;
1623                                 }
1624                         }
1625                         page_cache_release(pages[i]);
1626                         pages_locked++;
1627                 }
1628                 nrpages -= ret;
1629                 index += ret;
1630                 cond_resched();
1631         }
1632         ret = 0;
1633 done:
1634         if (ret && pages_locked) {
1635                 __unlock_for_delalloc(inode, locked_page,
1636                               delalloc_start,
1637                               ((u64)(start_index + pages_locked - 1)) <<
1638                               PAGE_CACHE_SHIFT);
1639         }
1640         return ret;
1641 }
1642
1643 /*
1644  * find a contiguous range of bytes in the file marked as delalloc, not
1645  * more than 'max_bytes'.  start and end are used to return the range,
1646  *
1647  * 1 is returned if we find something, 0 if nothing was in the tree
1648  */
1649 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1650                                     struct extent_io_tree *tree,
1651                                     struct page *locked_page, u64 *start,
1652                                     u64 *end, u64 max_bytes)
1653 {
1654         u64 delalloc_start;
1655         u64 delalloc_end;
1656         u64 found;
1657         struct extent_state *cached_state = NULL;
1658         int ret;
1659         int loops = 0;
1660
1661 again:
1662         /* step one, find a bunch of delalloc bytes starting at start */
1663         delalloc_start = *start;
1664         delalloc_end = 0;
1665         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1666                                     max_bytes, &cached_state);
1667         if (!found || delalloc_end <= *start) {
1668                 *start = delalloc_start;
1669                 *end = delalloc_end;
1670                 free_extent_state(cached_state);
1671                 return 0;
1672         }
1673
1674         /*
1675          * start comes from the offset of locked_page.  We have to lock
1676          * pages in order, so we can't process delalloc bytes before
1677          * locked_page
1678          */
1679         if (delalloc_start < *start)
1680                 delalloc_start = *start;
1681
1682         /*
1683          * make sure to limit the number of pages we try to lock down
1684          */
1685         if (delalloc_end + 1 - delalloc_start > max_bytes)
1686                 delalloc_end = delalloc_start + max_bytes - 1;
1687
1688         /* step two, lock all the pages after the page that has start */
1689         ret = lock_delalloc_pages(inode, locked_page,
1690                                   delalloc_start, delalloc_end);
1691         if (ret == -EAGAIN) {
1692                 /* some of the pages are gone, lets avoid looping by
1693                  * shortening the size of the delalloc range we're searching
1694                  */
1695                 free_extent_state(cached_state);
1696                 cached_state = NULL;
1697                 if (!loops) {
1698                         max_bytes = PAGE_CACHE_SIZE;
1699                         loops = 1;
1700                         goto again;
1701                 } else {
1702                         found = 0;
1703                         goto out_failed;
1704                 }
1705         }
1706         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1707
1708         /* step three, lock the state bits for the whole range */
1709         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1710
1711         /* then test to make sure it is all still delalloc */
1712         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1713                              EXTENT_DELALLOC, 1, cached_state);
1714         if (!ret) {
1715                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1716                                      &cached_state, GFP_NOFS);
1717                 __unlock_for_delalloc(inode, locked_page,
1718                               delalloc_start, delalloc_end);
1719                 cond_resched();
1720                 goto again;
1721         }
1722         free_extent_state(cached_state);
1723         *start = delalloc_start;
1724         *end = delalloc_end;
1725 out_failed:
1726         return found;
1727 }
1728
1729 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1730                                  struct page *locked_page,
1731                                  unsigned long clear_bits,
1732                                  unsigned long page_ops)
1733 {
1734         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1735         int ret;
1736         struct page *pages[16];
1737         unsigned long index = start >> PAGE_CACHE_SHIFT;
1738         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1739         unsigned long nr_pages = end_index - index + 1;
1740         int i;
1741
1742         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1743         if (page_ops == 0)
1744                 return 0;
1745
1746         while (nr_pages > 0) {
1747                 ret = find_get_pages_contig(inode->i_mapping, index,
1748                                      min_t(unsigned long,
1749                                      nr_pages, ARRAY_SIZE(pages)), pages);
1750                 for (i = 0; i < ret; i++) {
1751
1752                         if (page_ops & PAGE_SET_PRIVATE2)
1753                                 SetPagePrivate2(pages[i]);
1754
1755                         if (pages[i] == locked_page) {
1756                                 page_cache_release(pages[i]);
1757                                 continue;
1758                         }
1759                         if (page_ops & PAGE_CLEAR_DIRTY)
1760                                 clear_page_dirty_for_io(pages[i]);
1761                         if (page_ops & PAGE_SET_WRITEBACK)
1762                                 set_page_writeback(pages[i]);
1763                         if (page_ops & PAGE_END_WRITEBACK)
1764                                 end_page_writeback(pages[i]);
1765                         if (page_ops & PAGE_UNLOCK)
1766                                 unlock_page(pages[i]);
1767                         page_cache_release(pages[i]);
1768                 }
1769                 nr_pages -= ret;
1770                 index += ret;
1771                 cond_resched();
1772         }
1773         return 0;
1774 }
1775
1776 /*
1777  * count the number of bytes in the tree that have a given bit(s)
1778  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1779  * cached.  The total number found is returned.
1780  */
1781 u64 count_range_bits(struct extent_io_tree *tree,
1782                      u64 *start, u64 search_end, u64 max_bytes,
1783                      unsigned long bits, int contig)
1784 {
1785         struct rb_node *node;
1786         struct extent_state *state;
1787         u64 cur_start = *start;
1788         u64 total_bytes = 0;
1789         u64 last = 0;
1790         int found = 0;
1791
1792         if (WARN_ON(search_end <= cur_start))
1793                 return 0;
1794
1795         spin_lock(&tree->lock);
1796         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1797                 total_bytes = tree->dirty_bytes;
1798                 goto out;
1799         }
1800         /*
1801          * this search will find all the extents that end after
1802          * our range starts.
1803          */
1804         node = tree_search(tree, cur_start);
1805         if (!node)
1806                 goto out;
1807
1808         while (1) {
1809                 state = rb_entry(node, struct extent_state, rb_node);
1810                 if (state->start > search_end)
1811                         break;
1812                 if (contig && found && state->start > last + 1)
1813                         break;
1814                 if (state->end >= cur_start && (state->state & bits) == bits) {
1815                         total_bytes += min(search_end, state->end) + 1 -
1816                                        max(cur_start, state->start);
1817                         if (total_bytes >= max_bytes)
1818                                 break;
1819                         if (!found) {
1820                                 *start = max(cur_start, state->start);
1821                                 found = 1;
1822                         }
1823                         last = state->end;
1824                 } else if (contig && found) {
1825                         break;
1826                 }
1827                 node = rb_next(node);
1828                 if (!node)
1829                         break;
1830         }
1831 out:
1832         spin_unlock(&tree->lock);
1833         return total_bytes;
1834 }
1835
1836 /*
1837  * set the private field for a given byte offset in the tree.  If there isn't
1838  * an extent_state there already, this does nothing.
1839  */
1840 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
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         state->private = private;
1862 out:
1863         spin_unlock(&tree->lock);
1864         return ret;
1865 }
1866
1867 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1868 {
1869         struct rb_node *node;
1870         struct extent_state *state;
1871         int ret = 0;
1872
1873         spin_lock(&tree->lock);
1874         /*
1875          * this search will find all the extents that end after
1876          * our range starts.
1877          */
1878         node = tree_search(tree, start);
1879         if (!node) {
1880                 ret = -ENOENT;
1881                 goto out;
1882         }
1883         state = rb_entry(node, struct extent_state, rb_node);
1884         if (state->start != start) {
1885                 ret = -ENOENT;
1886                 goto out;
1887         }
1888         *private = state->private;
1889 out:
1890         spin_unlock(&tree->lock);
1891         return ret;
1892 }
1893
1894 /*
1895  * searches a range in the state tree for a given mask.
1896  * If 'filled' == 1, this returns 1 only if every extent in the tree
1897  * has the bits set.  Otherwise, 1 is returned if any bit in the
1898  * range is found set.
1899  */
1900 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1901                    unsigned long bits, int filled, struct extent_state *cached)
1902 {
1903         struct extent_state *state = NULL;
1904         struct rb_node *node;
1905         int bitset = 0;
1906
1907         spin_lock(&tree->lock);
1908         if (cached && cached->tree && cached->start <= start &&
1909             cached->end > start)
1910                 node = &cached->rb_node;
1911         else
1912                 node = tree_search(tree, start);
1913         while (node && start <= end) {
1914                 state = rb_entry(node, struct extent_state, rb_node);
1915
1916                 if (filled && state->start > start) {
1917                         bitset = 0;
1918                         break;
1919                 }
1920
1921                 if (state->start > end)
1922                         break;
1923
1924                 if (state->state & bits) {
1925                         bitset = 1;
1926                         if (!filled)
1927                                 break;
1928                 } else if (filled) {
1929                         bitset = 0;
1930                         break;
1931                 }
1932
1933                 if (state->end == (u64)-1)
1934                         break;
1935
1936                 start = state->end + 1;
1937                 if (start > end)
1938                         break;
1939                 node = rb_next(node);
1940                 if (!node) {
1941                         if (filled)
1942                                 bitset = 0;
1943                         break;
1944                 }
1945         }
1946         spin_unlock(&tree->lock);
1947         return bitset;
1948 }
1949
1950 /*
1951  * helper function to set a given page up to date if all the
1952  * extents in the tree for that page are up to date
1953  */
1954 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1955 {
1956         u64 start = page_offset(page);
1957         u64 end = start + PAGE_CACHE_SIZE - 1;
1958         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1959                 SetPageUptodate(page);
1960 }
1961
1962 /*
1963  * When IO fails, either with EIO or csum verification fails, we
1964  * try other mirrors that might have a good copy of the data.  This
1965  * io_failure_record is used to record state as we go through all the
1966  * mirrors.  If another mirror has good data, the page is set up to date
1967  * and things continue.  If a good mirror can't be found, the original
1968  * bio end_io callback is called to indicate things have failed.
1969  */
1970 struct io_failure_record {
1971         struct page *page;
1972         u64 start;
1973         u64 len;
1974         u64 logical;
1975         unsigned long bio_flags;
1976         int this_mirror;
1977         int failed_mirror;
1978         int in_validation;
1979 };
1980
1981 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1982                                 int did_repair)
1983 {
1984         int ret;
1985         int err = 0;
1986         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1987
1988         set_state_private(failure_tree, rec->start, 0);
1989         ret = clear_extent_bits(failure_tree, rec->start,
1990                                 rec->start + rec->len - 1,
1991                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1992         if (ret)
1993                 err = ret;
1994
1995         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1996                                 rec->start + rec->len - 1,
1997                                 EXTENT_DAMAGED, GFP_NOFS);
1998         if (ret && !err)
1999                 err = ret;
2000
2001         kfree(rec);
2002         return err;
2003 }
2004
2005 /*
2006  * this bypasses the standard btrfs submit functions deliberately, as
2007  * the standard behavior is to write all copies in a raid setup. here we only
2008  * want to write the one bad copy. so we do the mapping for ourselves and issue
2009  * submit_bio directly.
2010  * to avoid any synchronization issues, wait for the data after writing, which
2011  * actually prevents the read that triggered the error from finishing.
2012  * currently, there can be no more than two copies of every data bit. thus,
2013  * exactly one rewrite is required.
2014  */
2015 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 start,
2016                         u64 length, u64 logical, struct page *page,
2017                         int mirror_num)
2018 {
2019         struct bio *bio;
2020         struct btrfs_device *dev;
2021         u64 map_length = 0;
2022         u64 sector;
2023         struct btrfs_bio *bbio = NULL;
2024         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2025         int ret;
2026
2027         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2028         BUG_ON(!mirror_num);
2029
2030         /* we can't repair anything in raid56 yet */
2031         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2032                 return 0;
2033
2034         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2035         if (!bio)
2036                 return -EIO;
2037         bio->bi_iter.bi_size = 0;
2038         map_length = length;
2039
2040         ret = btrfs_map_block(fs_info, WRITE, logical,
2041                               &map_length, &bbio, mirror_num);
2042         if (ret) {
2043                 bio_put(bio);
2044                 return -EIO;
2045         }
2046         BUG_ON(mirror_num != bbio->mirror_num);
2047         sector = bbio->stripes[mirror_num-1].physical >> 9;
2048         bio->bi_iter.bi_sector = sector;
2049         dev = bbio->stripes[mirror_num-1].dev;
2050         kfree(bbio);
2051         if (!dev || !dev->bdev || !dev->writeable) {
2052                 bio_put(bio);
2053                 return -EIO;
2054         }
2055         bio->bi_bdev = dev->bdev;
2056         bio_add_page(bio, page, length, start - page_offset(page));
2057
2058         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2059                 /* try to remap that extent elsewhere? */
2060                 bio_put(bio);
2061                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2062                 return -EIO;
2063         }
2064
2065         printk_ratelimited_in_rcu(KERN_INFO
2066                         "BTRFS: read error corrected: ino %lu off %llu "
2067                     "(dev %s sector %llu)\n", page->mapping->host->i_ino,
2068                     start, rcu_str_deref(dev->name), sector);
2069
2070         bio_put(bio);
2071         return 0;
2072 }
2073
2074 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2075                          int mirror_num)
2076 {
2077         u64 start = eb->start;
2078         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2079         int ret = 0;
2080
2081         if (root->fs_info->sb->s_flags & MS_RDONLY)
2082                 return -EROFS;
2083
2084         for (i = 0; i < num_pages; i++) {
2085                 struct page *p = extent_buffer_page(eb, i);
2086                 ret = repair_io_failure(root->fs_info, start, PAGE_CACHE_SIZE,
2087                                         start, p, mirror_num);
2088                 if (ret)
2089                         break;
2090                 start += PAGE_CACHE_SIZE;
2091         }
2092
2093         return ret;
2094 }
2095
2096 /*
2097  * each time an IO finishes, we do a fast check in the IO failure tree
2098  * to see if we need to process or clean up an io_failure_record
2099  */
2100 static int clean_io_failure(u64 start, struct page *page)
2101 {
2102         u64 private;
2103         u64 private_failure;
2104         struct io_failure_record *failrec;
2105         struct inode *inode = page->mapping->host;
2106         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2107         struct extent_state *state;
2108         int num_copies;
2109         int did_repair = 0;
2110         int ret;
2111
2112         private = 0;
2113         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2114                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2115         if (!ret)
2116                 return 0;
2117
2118         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2119                                 &private_failure);
2120         if (ret)
2121                 return 0;
2122
2123         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2124         BUG_ON(!failrec->this_mirror);
2125
2126         if (failrec->in_validation) {
2127                 /* there was no real error, just free the record */
2128                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2129                          failrec->start);
2130                 did_repair = 1;
2131                 goto out;
2132         }
2133         if (fs_info->sb->s_flags & MS_RDONLY)
2134                 goto out;
2135
2136         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2137         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2138                                             failrec->start,
2139                                             EXTENT_LOCKED);
2140         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2141
2142         if (state && state->start <= failrec->start &&
2143             state->end >= failrec->start + failrec->len - 1) {
2144                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2145                                               failrec->len);
2146                 if (num_copies > 1)  {
2147                         ret = repair_io_failure(fs_info, start, failrec->len,
2148                                                 failrec->logical, page,
2149                                                 failrec->failed_mirror);
2150                         did_repair = !ret;
2151                 }
2152                 ret = 0;
2153         }
2154
2155 out:
2156         if (!ret)
2157                 ret = free_io_failure(inode, failrec, did_repair);
2158
2159         return ret;
2160 }
2161
2162 /*
2163  * this is a generic handler for readpage errors (default
2164  * readpage_io_failed_hook). if other copies exist, read those and write back
2165  * good data to the failed position. does not investigate in remapping the
2166  * failed extent elsewhere, hoping the device will be smart enough to do this as
2167  * needed
2168  */
2169
2170 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2171                               struct page *page, u64 start, u64 end,
2172                               int failed_mirror)
2173 {
2174         struct io_failure_record *failrec = NULL;
2175         u64 private;
2176         struct extent_map *em;
2177         struct inode *inode = page->mapping->host;
2178         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2179         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2180         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2181         struct bio *bio;
2182         struct btrfs_io_bio *btrfs_failed_bio;
2183         struct btrfs_io_bio *btrfs_bio;
2184         int num_copies;
2185         int ret;
2186         int read_mode;
2187         u64 logical;
2188
2189         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2190
2191         ret = get_state_private(failure_tree, start, &private);
2192         if (ret) {
2193                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2194                 if (!failrec)
2195                         return -ENOMEM;
2196                 failrec->start = start;
2197                 failrec->len = end - start + 1;
2198                 failrec->this_mirror = 0;
2199                 failrec->bio_flags = 0;
2200                 failrec->in_validation = 0;
2201
2202                 read_lock(&em_tree->lock);
2203                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2204                 if (!em) {
2205                         read_unlock(&em_tree->lock);
2206                         kfree(failrec);
2207                         return -EIO;
2208                 }
2209
2210                 if (em->start > start || em->start + em->len <= start) {
2211                         free_extent_map(em);
2212                         em = NULL;
2213                 }
2214                 read_unlock(&em_tree->lock);
2215
2216                 if (!em) {
2217                         kfree(failrec);
2218                         return -EIO;
2219                 }
2220                 logical = start - em->start;
2221                 logical = em->block_start + logical;
2222                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2223                         logical = em->block_start;
2224                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2225                         extent_set_compress_type(&failrec->bio_flags,
2226                                                  em->compress_type);
2227                 }
2228                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2229                          "len=%llu\n", logical, start, failrec->len);
2230                 failrec->logical = logical;
2231                 free_extent_map(em);
2232
2233                 /* set the bits in the private failure tree */
2234                 ret = set_extent_bits(failure_tree, start, end,
2235                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2236                 if (ret >= 0)
2237                         ret = set_state_private(failure_tree, start,
2238                                                 (u64)(unsigned long)failrec);
2239                 /* set the bits in the inode's tree */
2240                 if (ret >= 0)
2241                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2242                                                 GFP_NOFS);
2243                 if (ret < 0) {
2244                         kfree(failrec);
2245                         return ret;
2246                 }
2247         } else {
2248                 failrec = (struct io_failure_record *)(unsigned long)private;
2249                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2250                          "start=%llu, len=%llu, validation=%d\n",
2251                          failrec->logical, failrec->start, failrec->len,
2252                          failrec->in_validation);
2253                 /*
2254                  * when data can be on disk more than twice, add to failrec here
2255                  * (e.g. with a list for failed_mirror) to make
2256                  * clean_io_failure() clean all those errors at once.
2257                  */
2258         }
2259         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2260                                       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                 pr_debug("bio_readpage_error: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2268                          num_copies, failrec->this_mirror, failed_mirror);
2269                 free_io_failure(inode, failrec, 0);
2270                 return -EIO;
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->bi_vcnt > 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                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2291         } else {
2292                 /*
2293                  * we're ready to fulfill a) and b) alongside. get a good copy
2294                  * of the failed sector and if we succeed, we have setup
2295                  * everything for repair_io_failure to do the rest for us.
2296                  */
2297                 if (failrec->in_validation) {
2298                         BUG_ON(failrec->this_mirror != failed_mirror);
2299                         failrec->in_validation = 0;
2300                         failrec->this_mirror = 0;
2301                 }
2302                 failrec->failed_mirror = failed_mirror;
2303                 failrec->this_mirror++;
2304                 if (failrec->this_mirror == failed_mirror)
2305                         failrec->this_mirror++;
2306                 read_mode = READ_SYNC;
2307         }
2308
2309         if (failrec->this_mirror > num_copies) {
2310                 pr_debug("bio_readpage_error: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2311                          num_copies, failrec->this_mirror, failed_mirror);
2312                 free_io_failure(inode, failrec, 0);
2313                 return -EIO;
2314         }
2315
2316         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2317         if (!bio) {
2318                 free_io_failure(inode, failrec, 0);
2319                 return -EIO;
2320         }
2321         bio->bi_end_io = failed_bio->bi_end_io;
2322         bio->bi_iter.bi_sector = failrec->logical >> 9;
2323         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2324         bio->bi_iter.bi_size = 0;
2325
2326         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2327         if (btrfs_failed_bio->csum) {
2328                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2329                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2330
2331                 btrfs_bio = btrfs_io_bio(bio);
2332                 btrfs_bio->csum = btrfs_bio->csum_inline;
2333                 phy_offset >>= inode->i_sb->s_blocksize_bits;
2334                 phy_offset *= csum_size;
2335                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + phy_offset,
2336                        csum_size);
2337         }
2338
2339         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2340
2341         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2342                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2343                  failrec->this_mirror, num_copies, failrec->in_validation);
2344
2345         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2346                                          failrec->this_mirror,
2347                                          failrec->bio_flags, 0);
2348         return ret;
2349 }
2350
2351 /* lots and lots of room for performance fixes in the end_bio funcs */
2352
2353 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2354 {
2355         int uptodate = (err == 0);
2356         struct extent_io_tree *tree;
2357         int ret;
2358
2359         tree = &BTRFS_I(page->mapping->host)->io_tree;
2360
2361         if (tree->ops && tree->ops->writepage_end_io_hook) {
2362                 ret = tree->ops->writepage_end_io_hook(page, start,
2363                                                end, NULL, uptodate);
2364                 if (ret)
2365                         uptodate = 0;
2366         }
2367
2368         if (!uptodate) {
2369                 ClearPageUptodate(page);
2370                 SetPageError(page);
2371                 ret = ret < 0 ? ret : -EIO;
2372                 mapping_set_error(page->mapping, ret);
2373         }
2374         return 0;
2375 }
2376
2377 /*
2378  * after a writepage IO is done, we need to:
2379  * clear the uptodate bits on error
2380  * clear the writeback bits in the extent tree for this IO
2381  * end_page_writeback if the page has no more pending IO
2382  *
2383  * Scheduling is not allowed, so the extent state tree is expected
2384  * to have one and only one object corresponding to this IO.
2385  */
2386 static void end_bio_extent_writepage(struct bio *bio, int err)
2387 {
2388         struct bio_vec *bvec;
2389         u64 start;
2390         u64 end;
2391         int i;
2392
2393         bio_for_each_segment_all(bvec, bio, i) {
2394                 struct page *page = bvec->bv_page;
2395
2396                 /* We always issue full-page reads, but if some block
2397                  * in a page fails to read, blk_update_request() will
2398                  * advance bv_offset and adjust bv_len to compensate.
2399                  * Print a warning for nonzero offsets, and an error
2400                  * if they don't add up to a full page.  */
2401                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2402                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2403                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2404                                    "partial page write in btrfs with offset %u and length %u",
2405                                         bvec->bv_offset, bvec->bv_len);
2406                         else
2407                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2408                                    "incomplete page write in btrfs with offset %u and "
2409                                    "length %u",
2410                                         bvec->bv_offset, bvec->bv_len);
2411                 }
2412
2413                 start = page_offset(page);
2414                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2415
2416                 if (end_extent_writepage(page, err, start, end))
2417                         continue;
2418
2419                 end_page_writeback(page);
2420         }
2421
2422         bio_put(bio);
2423 }
2424
2425 static void
2426 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2427                               int uptodate)
2428 {
2429         struct extent_state *cached = NULL;
2430         u64 end = start + len - 1;
2431
2432         if (uptodate && tree->track_uptodate)
2433                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2434         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2435 }
2436
2437 /*
2438  * after a readpage IO is done, we need to:
2439  * clear the uptodate bits on error
2440  * set the uptodate bits if things worked
2441  * set the page up to date if all extents in the tree are uptodate
2442  * clear the lock bit in the extent tree
2443  * unlock the page if there are no other extents locked for it
2444  *
2445  * Scheduling is not allowed, so the extent state tree is expected
2446  * to have one and only one object corresponding to this IO.
2447  */
2448 static void end_bio_extent_readpage(struct bio *bio, int err)
2449 {
2450         struct bio_vec *bvec;
2451         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2452         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2453         struct extent_io_tree *tree;
2454         u64 offset = 0;
2455         u64 start;
2456         u64 end;
2457         u64 len;
2458         u64 extent_start = 0;
2459         u64 extent_len = 0;
2460         int mirror;
2461         int ret;
2462         int i;
2463
2464         if (err)
2465                 uptodate = 0;
2466
2467         bio_for_each_segment_all(bvec, bio, i) {
2468                 struct page *page = bvec->bv_page;
2469                 struct inode *inode = page->mapping->host;
2470
2471                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2472                          "mirror=%lu\n", (u64)bio->bi_iter.bi_sector, err,
2473                          io_bio->mirror_num);
2474                 tree = &BTRFS_I(inode)->io_tree;
2475
2476                 /* We always issue full-page reads, but if some block
2477                  * in a page fails to read, blk_update_request() will
2478                  * advance bv_offset and adjust bv_len to compensate.
2479                  * Print a warning for nonzero offsets, and an error
2480                  * if they don't add up to a full page.  */
2481                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2482                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2483                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2484                                    "partial page read in btrfs with offset %u and length %u",
2485                                         bvec->bv_offset, bvec->bv_len);
2486                         else
2487                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2488                                    "incomplete page read in btrfs with offset %u and "
2489                                    "length %u",
2490                                         bvec->bv_offset, bvec->bv_len);
2491                 }
2492
2493                 start = page_offset(page);
2494                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2495                 len = bvec->bv_len;
2496
2497                 mirror = io_bio->mirror_num;
2498                 if (likely(uptodate && tree->ops &&
2499                            tree->ops->readpage_end_io_hook)) {
2500                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2501                                                               page, start, end,
2502                                                               mirror);
2503                         if (ret)
2504                                 uptodate = 0;
2505                         else
2506                                 clean_io_failure(start, page);
2507                 }
2508
2509                 if (likely(uptodate))
2510                         goto readpage_ok;
2511
2512                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2513                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2514                         if (!ret && !err &&
2515                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2516                                 uptodate = 1;
2517                 } else {
2518                         /*
2519                          * The generic bio_readpage_error handles errors the
2520                          * following way: If possible, new read requests are
2521                          * created and submitted and will end up in
2522                          * end_bio_extent_readpage as well (if we're lucky, not
2523                          * in the !uptodate case). In that case it returns 0 and
2524                          * we just go on with the next page in our bio. If it
2525                          * can't handle the error it will return -EIO and we
2526                          * remain responsible for that page.
2527                          */
2528                         ret = bio_readpage_error(bio, offset, page, start, end,
2529                                                  mirror);
2530                         if (ret == 0) {
2531                                 uptodate =
2532                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2533                                 if (err)
2534                                         uptodate = 0;
2535                                 continue;
2536                         }
2537                 }
2538 readpage_ok:
2539                 if (likely(uptodate)) {
2540                         loff_t i_size = i_size_read(inode);
2541                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2542                         unsigned offset;
2543
2544                         /* Zero out the end if this page straddles i_size */
2545                         offset = i_size & (PAGE_CACHE_SIZE-1);
2546                         if (page->index == end_index && offset)
2547                                 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2548                         SetPageUptodate(page);
2549                 } else {
2550                         ClearPageUptodate(page);
2551                         SetPageError(page);
2552                 }
2553                 unlock_page(page);
2554                 offset += len;
2555
2556                 if (unlikely(!uptodate)) {
2557                         if (extent_len) {
2558                                 endio_readpage_release_extent(tree,
2559                                                               extent_start,
2560                                                               extent_len, 1);
2561                                 extent_start = 0;
2562                                 extent_len = 0;
2563                         }
2564                         endio_readpage_release_extent(tree, start,
2565                                                       end - start + 1, 0);
2566                 } else if (!extent_len) {
2567                         extent_start = start;
2568                         extent_len = end + 1 - start;
2569                 } else if (extent_start + extent_len == start) {
2570                         extent_len += end + 1 - start;
2571                 } else {
2572                         endio_readpage_release_extent(tree, extent_start,
2573                                                       extent_len, uptodate);
2574                         extent_start = start;
2575                         extent_len = end + 1 - start;
2576                 }
2577         }
2578
2579         if (extent_len)
2580                 endio_readpage_release_extent(tree, extent_start, extent_len,
2581                                               uptodate);
2582         if (io_bio->end_io)
2583                 io_bio->end_io(io_bio, err);
2584         bio_put(bio);
2585 }
2586
2587 /*
2588  * this allocates from the btrfs_bioset.  We're returning a bio right now
2589  * but you can call btrfs_io_bio for the appropriate container_of magic
2590  */
2591 struct bio *
2592 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2593                 gfp_t gfp_flags)
2594 {
2595         struct btrfs_io_bio *btrfs_bio;
2596         struct bio *bio;
2597
2598         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2599
2600         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2601                 while (!bio && (nr_vecs /= 2)) {
2602                         bio = bio_alloc_bioset(gfp_flags,
2603                                                nr_vecs, btrfs_bioset);
2604                 }
2605         }
2606
2607         if (bio) {
2608                 bio->bi_bdev = bdev;
2609                 bio->bi_iter.bi_sector = first_sector;
2610                 btrfs_bio = btrfs_io_bio(bio);
2611                 btrfs_bio->csum = NULL;
2612                 btrfs_bio->csum_allocated = NULL;
2613                 btrfs_bio->end_io = NULL;
2614         }
2615         return bio;
2616 }
2617
2618 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2619 {
2620         return bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2621 }
2622
2623
2624 /* this also allocates from the btrfs_bioset */
2625 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2626 {
2627         struct btrfs_io_bio *btrfs_bio;
2628         struct bio *bio;
2629
2630         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2631         if (bio) {
2632                 btrfs_bio = btrfs_io_bio(bio);
2633                 btrfs_bio->csum = NULL;
2634                 btrfs_bio->csum_allocated = NULL;
2635                 btrfs_bio->end_io = NULL;
2636         }
2637         return bio;
2638 }
2639
2640
2641 static int __must_check submit_one_bio(int rw, struct bio *bio,
2642                                        int mirror_num, unsigned long bio_flags)
2643 {
2644         int ret = 0;
2645         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2646         struct page *page = bvec->bv_page;
2647         struct extent_io_tree *tree = bio->bi_private;
2648         u64 start;
2649
2650         start = page_offset(page) + bvec->bv_offset;
2651
2652         bio->bi_private = NULL;
2653
2654         bio_get(bio);
2655
2656         if (tree->ops && tree->ops->submit_bio_hook)
2657                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2658                                            mirror_num, bio_flags, start);
2659         else
2660                 btrfsic_submit_bio(rw, bio);
2661
2662         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2663                 ret = -EOPNOTSUPP;
2664         bio_put(bio);
2665         return ret;
2666 }
2667
2668 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2669                      unsigned long offset, size_t size, struct bio *bio,
2670                      unsigned long bio_flags)
2671 {
2672         int ret = 0;
2673         if (tree->ops && tree->ops->merge_bio_hook)
2674                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2675                                                 bio_flags);
2676         BUG_ON(ret < 0);
2677         return ret;
2678
2679 }
2680
2681 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2682                               struct page *page, sector_t sector,
2683                               size_t size, unsigned long offset,
2684                               struct block_device *bdev,
2685                               struct bio **bio_ret,
2686                               unsigned long max_pages,
2687                               bio_end_io_t end_io_func,
2688                               int mirror_num,
2689                               unsigned long prev_bio_flags,
2690                               unsigned long bio_flags)
2691 {
2692         int ret = 0;
2693         struct bio *bio;
2694         int nr;
2695         int contig = 0;
2696         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2697         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2698         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2699
2700         if (bio_ret && *bio_ret) {
2701                 bio = *bio_ret;
2702                 if (old_compressed)
2703                         contig = bio->bi_iter.bi_sector == sector;
2704                 else
2705                         contig = bio_end_sector(bio) == sector;
2706
2707                 if (prev_bio_flags != bio_flags || !contig ||
2708                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2709                     bio_add_page(bio, page, page_size, offset) < page_size) {
2710                         ret = submit_one_bio(rw, bio, mirror_num,
2711                                              prev_bio_flags);
2712                         if (ret < 0)
2713                                 return ret;
2714                         bio = NULL;
2715                 } else {
2716                         return 0;
2717                 }
2718         }
2719         if (this_compressed)
2720                 nr = BIO_MAX_PAGES;
2721         else
2722                 nr = bio_get_nr_vecs(bdev);
2723
2724         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2725         if (!bio)
2726                 return -ENOMEM;
2727
2728         bio_add_page(bio, page, page_size, offset);
2729         bio->bi_end_io = end_io_func;
2730         bio->bi_private = tree;
2731
2732         if (bio_ret)
2733                 *bio_ret = bio;
2734         else
2735                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2736
2737         return ret;
2738 }
2739
2740 static void attach_extent_buffer_page(struct extent_buffer *eb,
2741                                       struct page *page)
2742 {
2743         if (!PagePrivate(page)) {
2744                 SetPagePrivate(page);
2745                 page_cache_get(page);
2746                 set_page_private(page, (unsigned long)eb);
2747         } else {
2748                 WARN_ON(page->private != (unsigned long)eb);
2749         }
2750 }
2751
2752 void set_page_extent_mapped(struct page *page)
2753 {
2754         if (!PagePrivate(page)) {
2755                 SetPagePrivate(page);
2756                 page_cache_get(page);
2757                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2758         }
2759 }
2760
2761 static struct extent_map *
2762 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2763                  u64 start, u64 len, get_extent_t *get_extent,
2764                  struct extent_map **em_cached)
2765 {
2766         struct extent_map *em;
2767
2768         if (em_cached && *em_cached) {
2769                 em = *em_cached;
2770                 if (extent_map_in_tree(em) && start >= em->start &&
2771                     start < extent_map_end(em)) {
2772                         atomic_inc(&em->refs);
2773                         return em;
2774                 }
2775
2776                 free_extent_map(em);
2777                 *em_cached = NULL;
2778         }
2779
2780         em = get_extent(inode, page, pg_offset, start, len, 0);
2781         if (em_cached && !IS_ERR_OR_NULL(em)) {
2782                 BUG_ON(*em_cached);
2783                 atomic_inc(&em->refs);
2784                 *em_cached = em;
2785         }
2786         return em;
2787 }
2788 /*
2789  * basic readpage implementation.  Locked extent state structs are inserted
2790  * into the tree that are removed when the IO is done (by the end_io
2791  * handlers)
2792  * XXX JDM: This needs looking at to ensure proper page locking
2793  */
2794 static int __do_readpage(struct extent_io_tree *tree,
2795                          struct page *page,
2796                          get_extent_t *get_extent,
2797                          struct extent_map **em_cached,
2798                          struct bio **bio, int mirror_num,
2799                          unsigned long *bio_flags, int rw)
2800 {
2801         struct inode *inode = page->mapping->host;
2802         u64 start = page_offset(page);
2803         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2804         u64 end;
2805         u64 cur = start;
2806         u64 extent_offset;
2807         u64 last_byte = i_size_read(inode);
2808         u64 block_start;
2809         u64 cur_end;
2810         sector_t sector;
2811         struct extent_map *em;
2812         struct block_device *bdev;
2813         int ret;
2814         int nr = 0;
2815         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2816         size_t pg_offset = 0;
2817         size_t iosize;
2818         size_t disk_io_size;
2819         size_t blocksize = inode->i_sb->s_blocksize;
2820         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2821
2822         set_page_extent_mapped(page);
2823
2824         end = page_end;
2825         if (!PageUptodate(page)) {
2826                 if (cleancache_get_page(page) == 0) {
2827                         BUG_ON(blocksize != PAGE_SIZE);
2828                         unlock_extent(tree, start, end);
2829                         goto out;
2830                 }
2831         }
2832
2833         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2834                 char *userpage;
2835                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2836
2837                 if (zero_offset) {
2838                         iosize = PAGE_CACHE_SIZE - zero_offset;
2839                         userpage = kmap_atomic(page);
2840                         memset(userpage + zero_offset, 0, iosize);
2841                         flush_dcache_page(page);
2842                         kunmap_atomic(userpage);
2843                 }
2844         }
2845         while (cur <= end) {
2846                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2847
2848                 if (cur >= last_byte) {
2849                         char *userpage;
2850                         struct extent_state *cached = NULL;
2851
2852                         iosize = PAGE_CACHE_SIZE - pg_offset;
2853                         userpage = kmap_atomic(page);
2854                         memset(userpage + pg_offset, 0, iosize);
2855                         flush_dcache_page(page);
2856                         kunmap_atomic(userpage);
2857                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2858                                             &cached, GFP_NOFS);
2859                         if (!parent_locked)
2860                                 unlock_extent_cached(tree, cur,
2861                                                      cur + iosize - 1,
2862                                                      &cached, GFP_NOFS);
2863                         break;
2864                 }
2865                 em = __get_extent_map(inode, page, pg_offset, cur,
2866                                       end - cur + 1, get_extent, em_cached);
2867                 if (IS_ERR_OR_NULL(em)) {
2868                         SetPageError(page);
2869                         if (!parent_locked)
2870                                 unlock_extent(tree, cur, end);
2871                         break;
2872                 }
2873                 extent_offset = cur - em->start;
2874                 BUG_ON(extent_map_end(em) <= cur);
2875                 BUG_ON(end < cur);
2876
2877                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2878                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2879                         extent_set_compress_type(&this_bio_flag,
2880                                                  em->compress_type);
2881                 }
2882
2883                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2884                 cur_end = min(extent_map_end(em) - 1, end);
2885                 iosize = ALIGN(iosize, blocksize);
2886                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2887                         disk_io_size = em->block_len;
2888                         sector = em->block_start >> 9;
2889                 } else {
2890                         sector = (em->block_start + extent_offset) >> 9;
2891                         disk_io_size = iosize;
2892                 }
2893                 bdev = em->bdev;
2894                 block_start = em->block_start;
2895                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2896                         block_start = EXTENT_MAP_HOLE;
2897                 free_extent_map(em);
2898                 em = NULL;
2899
2900                 /* we've found a hole, just zero and go on */
2901                 if (block_start == EXTENT_MAP_HOLE) {
2902                         char *userpage;
2903                         struct extent_state *cached = NULL;
2904
2905                         userpage = kmap_atomic(page);
2906                         memset(userpage + pg_offset, 0, iosize);
2907                         flush_dcache_page(page);
2908                         kunmap_atomic(userpage);
2909
2910                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2911                                             &cached, GFP_NOFS);
2912                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2913                                              &cached, GFP_NOFS);
2914                         cur = cur + iosize;
2915                         pg_offset += iosize;
2916                         continue;
2917                 }
2918                 /* the get_extent function already copied into the page */
2919                 if (test_range_bit(tree, cur, cur_end,
2920                                    EXTENT_UPTODATE, 1, NULL)) {
2921                         check_page_uptodate(tree, page);
2922                         if (!parent_locked)
2923                                 unlock_extent(tree, cur, cur + iosize - 1);
2924                         cur = cur + iosize;
2925                         pg_offset += iosize;
2926                         continue;
2927                 }
2928                 /* we have an inline extent but it didn't get marked up
2929                  * to date.  Error out
2930                  */
2931                 if (block_start == EXTENT_MAP_INLINE) {
2932                         SetPageError(page);
2933                         if (!parent_locked)
2934                                 unlock_extent(tree, cur, cur + iosize - 1);
2935                         cur = cur + iosize;
2936                         pg_offset += iosize;
2937                         continue;
2938                 }
2939
2940                 pnr -= page->index;
2941                 ret = submit_extent_page(rw, tree, page,
2942                                          sector, disk_io_size, pg_offset,
2943                                          bdev, bio, pnr,
2944                                          end_bio_extent_readpage, mirror_num,
2945                                          *bio_flags,
2946                                          this_bio_flag);
2947                 if (!ret) {
2948                         nr++;
2949                         *bio_flags = this_bio_flag;
2950                 } else {
2951                         SetPageError(page);
2952                         if (!parent_locked)
2953                                 unlock_extent(tree, cur, cur + iosize - 1);
2954                 }
2955                 cur = cur + iosize;
2956                 pg_offset += iosize;
2957         }
2958 out:
2959         if (!nr) {
2960                 if (!PageError(page))
2961                         SetPageUptodate(page);
2962                 unlock_page(page);
2963         }
2964         return 0;
2965 }
2966
2967 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
2968                                              struct page *pages[], int nr_pages,
2969                                              u64 start, u64 end,
2970                                              get_extent_t *get_extent,
2971                                              struct extent_map **em_cached,
2972                                              struct bio **bio, int mirror_num,
2973                                              unsigned long *bio_flags, int rw)
2974 {
2975         struct inode *inode;
2976         struct btrfs_ordered_extent *ordered;
2977         int index;
2978
2979         inode = pages[0]->mapping->host;
2980         while (1) {
2981                 lock_extent(tree, start, end);
2982                 ordered = btrfs_lookup_ordered_range(inode, start,
2983                                                      end - start + 1);
2984                 if (!ordered)
2985                         break;
2986                 unlock_extent(tree, start, end);
2987                 btrfs_start_ordered_extent(inode, ordered, 1);
2988                 btrfs_put_ordered_extent(ordered);
2989         }
2990
2991         for (index = 0; index < nr_pages; index++) {
2992                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
2993                               mirror_num, bio_flags, rw);
2994                 page_cache_release(pages[index]);
2995         }
2996 }
2997
2998 static void __extent_readpages(struct extent_io_tree *tree,
2999                                struct page *pages[],
3000                                int nr_pages, get_extent_t *get_extent,
3001                                struct extent_map **em_cached,
3002                                struct bio **bio, int mirror_num,
3003                                unsigned long *bio_flags, int rw)
3004 {
3005         u64 start = 0;
3006         u64 end = 0;
3007         u64 page_start;
3008         int index;
3009         int first_index = 0;
3010
3011         for (index = 0; index < nr_pages; index++) {
3012                 page_start = page_offset(pages[index]);
3013                 if (!end) {
3014                         start = page_start;
3015                         end = start + PAGE_CACHE_SIZE - 1;
3016                         first_index = index;
3017                 } else if (end + 1 == page_start) {
3018                         end += PAGE_CACHE_SIZE;
3019                 } else {
3020                         __do_contiguous_readpages(tree, &pages[first_index],
3021                                                   index - first_index, start,
3022                                                   end, get_extent, em_cached,
3023                                                   bio, mirror_num, bio_flags,
3024                                                   rw);
3025                         start = page_start;
3026                         end = start + PAGE_CACHE_SIZE - 1;
3027                         first_index = index;
3028                 }
3029         }
3030
3031         if (end)
3032                 __do_contiguous_readpages(tree, &pages[first_index],
3033                                           index - first_index, start,
3034                                           end, get_extent, em_cached, bio,
3035                                           mirror_num, bio_flags, rw);
3036 }
3037
3038 static int __extent_read_full_page(struct extent_io_tree *tree,
3039                                    struct page *page,
3040                                    get_extent_t *get_extent,
3041                                    struct bio **bio, int mirror_num,
3042                                    unsigned long *bio_flags, int rw)
3043 {
3044         struct inode *inode = page->mapping->host;
3045         struct btrfs_ordered_extent *ordered;
3046         u64 start = page_offset(page);
3047         u64 end = start + PAGE_CACHE_SIZE - 1;
3048         int ret;
3049
3050         while (1) {
3051                 lock_extent(tree, start, end);
3052                 ordered = btrfs_lookup_ordered_extent(inode, start);
3053                 if (!ordered)
3054                         break;
3055                 unlock_extent(tree, start, end);
3056                 btrfs_start_ordered_extent(inode, ordered, 1);
3057                 btrfs_put_ordered_extent(ordered);
3058         }
3059
3060         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3061                             bio_flags, rw);
3062         return ret;
3063 }
3064
3065 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3066                             get_extent_t *get_extent, int mirror_num)
3067 {
3068         struct bio *bio = NULL;
3069         unsigned long bio_flags = 0;
3070         int ret;
3071
3072         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3073                                       &bio_flags, READ);
3074         if (bio)
3075                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3076         return ret;
3077 }
3078
3079 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3080                                  get_extent_t *get_extent, int mirror_num)
3081 {
3082         struct bio *bio = NULL;
3083         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3084         int ret;
3085
3086         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3087                                       &bio_flags, READ);
3088         if (bio)
3089                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3090         return ret;
3091 }
3092
3093 static noinline void update_nr_written(struct page *page,
3094                                       struct writeback_control *wbc,
3095                                       unsigned long nr_written)
3096 {
3097         wbc->nr_to_write -= nr_written;
3098         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3099             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3100                 page->mapping->writeback_index = page->index + nr_written;
3101 }
3102
3103 /*
3104  * helper for __extent_writepage, doing all of the delayed allocation setup.
3105  *
3106  * This returns 1 if our fill_delalloc function did all the work required
3107  * to write the page (copy into inline extent).  In this case the IO has
3108  * been started and the page is already unlocked.
3109  *
3110  * This returns 0 if all went well (page still locked)
3111  * This returns < 0 if there were errors (page still locked)
3112  */
3113 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3114                               struct page *page, struct writeback_control *wbc,
3115                               struct extent_page_data *epd,
3116                               u64 delalloc_start,
3117                               unsigned long *nr_written)
3118 {
3119         struct extent_io_tree *tree = epd->tree;
3120         u64 page_end = delalloc_start + PAGE_CACHE_SIZE - 1;
3121         u64 nr_delalloc;
3122         u64 delalloc_to_write = 0;
3123         u64 delalloc_end = 0;
3124         int ret;
3125         int page_started = 0;
3126
3127         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3128                 return 0;
3129
3130         while (delalloc_end < page_end) {
3131                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3132                                                page,
3133                                                &delalloc_start,
3134                                                &delalloc_end,
3135                                                128 * 1024 * 1024);
3136                 if (nr_delalloc == 0) {
3137                         delalloc_start = delalloc_end + 1;
3138                         continue;
3139                 }
3140                 ret = tree->ops->fill_delalloc(inode, page,
3141                                                delalloc_start,
3142                                                delalloc_end,
3143                                                &page_started,
3144                                                nr_written);
3145                 /* File system has been set read-only */
3146                 if (ret) {
3147                         SetPageError(page);
3148                         /* fill_delalloc should be return < 0 for error
3149                          * but just in case, we use > 0 here meaning the
3150                          * IO is started, so we don't want to return > 0
3151                          * unless things are going well.
3152                          */
3153                         ret = ret < 0 ? ret : -EIO;
3154                         goto done;
3155                 }
3156                 /*
3157                  * delalloc_end is already one less than the total
3158                  * length, so we don't subtract one from
3159                  * PAGE_CACHE_SIZE
3160                  */
3161                 delalloc_to_write += (delalloc_end - delalloc_start +
3162                                       PAGE_CACHE_SIZE) >>
3163                                       PAGE_CACHE_SHIFT;
3164                 delalloc_start = delalloc_end + 1;
3165         }
3166         if (wbc->nr_to_write < delalloc_to_write) {
3167                 int thresh = 8192;
3168
3169                 if (delalloc_to_write < thresh * 2)
3170                         thresh = delalloc_to_write;
3171                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3172                                          thresh);
3173         }
3174
3175         /* did the fill delalloc function already unlock and start
3176          * the IO?
3177          */
3178         if (page_started) {
3179                 /*
3180                  * we've unlocked the page, so we can't update
3181                  * the mapping's writeback index, just update
3182                  * nr_to_write.
3183                  */
3184                 wbc->nr_to_write -= *nr_written;
3185                 return 1;
3186         }
3187
3188         ret = 0;
3189
3190 done:
3191         return ret;
3192 }
3193
3194 /*
3195  * helper for __extent_writepage.  This calls the writepage start hooks,
3196  * and does the loop to map the page into extents and bios.
3197  *
3198  * We return 1 if the IO is started and the page is unlocked,
3199  * 0 if all went well (page still locked)
3200  * < 0 if there were errors (page still locked)
3201  */
3202 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3203                                  struct page *page,
3204                                  struct writeback_control *wbc,
3205                                  struct extent_page_data *epd,
3206                                  loff_t i_size,
3207                                  unsigned long nr_written,
3208                                  int write_flags, int *nr_ret)
3209 {
3210         struct extent_io_tree *tree = epd->tree;
3211         u64 start = page_offset(page);
3212         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3213         u64 end;
3214         u64 cur = start;
3215         u64 extent_offset;
3216         u64 block_start;
3217         u64 iosize;
3218         sector_t sector;
3219         struct extent_state *cached_state = NULL;
3220         struct extent_map *em;
3221         struct block_device *bdev;
3222         size_t pg_offset = 0;
3223         size_t blocksize;
3224         int ret = 0;
3225         int nr = 0;
3226         bool compressed;
3227
3228         if (tree->ops && tree->ops->writepage_start_hook) {
3229                 ret = tree->ops->writepage_start_hook(page, start,
3230                                                       page_end);
3231                 if (ret) {
3232                         /* Fixup worker will requeue */
3233                         if (ret == -EBUSY)
3234                                 wbc->pages_skipped++;
3235                         else
3236                                 redirty_page_for_writepage(wbc, page);
3237
3238                         update_nr_written(page, wbc, nr_written);
3239                         unlock_page(page);
3240                         ret = 1;
3241                         goto done_unlocked;
3242                 }
3243         }
3244
3245         /*
3246          * we don't want to touch the inode after unlocking the page,
3247          * so we update the mapping writeback index now
3248          */
3249         update_nr_written(page, wbc, nr_written + 1);
3250
3251         end = page_end;
3252         if (i_size <= start) {
3253                 if (tree->ops && tree->ops->writepage_end_io_hook)
3254                         tree->ops->writepage_end_io_hook(page, start,
3255                                                          page_end, NULL, 1);
3256                 goto done;
3257         }
3258
3259         blocksize = inode->i_sb->s_blocksize;
3260
3261         while (cur <= end) {
3262                 u64 em_end;
3263                 if (cur >= i_size) {
3264                         if (tree->ops && tree->ops->writepage_end_io_hook)
3265                                 tree->ops->writepage_end_io_hook(page, cur,
3266                                                          page_end, NULL, 1);
3267                         break;
3268                 }
3269                 em = epd->get_extent(inode, page, pg_offset, cur,
3270                                      end - cur + 1, 1);
3271                 if (IS_ERR_OR_NULL(em)) {
3272                         SetPageError(page);
3273                         ret = PTR_ERR_OR_ZERO(em);
3274                         break;
3275                 }
3276
3277                 extent_offset = cur - em->start;
3278                 em_end = extent_map_end(em);
3279                 BUG_ON(em_end <= cur);
3280                 BUG_ON(end < cur);
3281                 iosize = min(em_end - cur, end - cur + 1);
3282                 iosize = ALIGN(iosize, blocksize);
3283                 sector = (em->block_start + extent_offset) >> 9;
3284                 bdev = em->bdev;
3285                 block_start = em->block_start;
3286                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3287                 free_extent_map(em);
3288                 em = NULL;
3289
3290                 /*
3291                  * compressed and inline extents are written through other
3292                  * paths in the FS
3293                  */
3294                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3295                     block_start == EXTENT_MAP_INLINE) {
3296                         /*
3297                          * end_io notification does not happen here for
3298                          * compressed extents
3299                          */
3300                         if (!compressed && tree->ops &&
3301                             tree->ops->writepage_end_io_hook)
3302                                 tree->ops->writepage_end_io_hook(page, cur,
3303                                                          cur + iosize - 1,
3304                                                          NULL, 1);
3305                         else if (compressed) {
3306                                 /* we don't want to end_page_writeback on
3307                                  * a compressed extent.  this happens
3308                                  * elsewhere
3309                                  */
3310                                 nr++;
3311                         }
3312
3313                         cur += iosize;
3314                         pg_offset += iosize;
3315                         continue;
3316                 }
3317
3318                 if (tree->ops && tree->ops->writepage_io_hook) {
3319                         ret = tree->ops->writepage_io_hook(page, cur,
3320                                                 cur + iosize - 1);
3321                 } else {
3322                         ret = 0;
3323                 }
3324                 if (ret) {
3325                         SetPageError(page);
3326                 } else {
3327                         unsigned long max_nr = (i_size >> PAGE_CACHE_SHIFT) + 1;
3328
3329                         set_range_writeback(tree, cur, cur + iosize - 1);
3330                         if (!PageWriteback(page)) {
3331                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
3332                                            "page %lu not writeback, cur %llu end %llu",
3333                                        page->index, cur, end);
3334                         }
3335
3336                         ret = submit_extent_page(write_flags, tree, page,
3337                                                  sector, iosize, pg_offset,
3338                                                  bdev, &epd->bio, max_nr,
3339                                                  end_bio_extent_writepage,
3340                                                  0, 0, 0);
3341                         if (ret)
3342                                 SetPageError(page);
3343                 }
3344                 cur = cur + iosize;
3345                 pg_offset += iosize;
3346                 nr++;
3347         }
3348 done:
3349         *nr_ret = nr;
3350
3351 done_unlocked:
3352
3353         /* drop our reference on any cached states */
3354         free_extent_state(cached_state);
3355         return ret;
3356 }
3357
3358 /*
3359  * the writepage semantics are similar to regular writepage.  extent
3360  * records are inserted to lock ranges in the tree, and as dirty areas
3361  * are found, they are marked writeback.  Then the lock bits are removed
3362  * and the end_io handler clears the writeback ranges
3363  */
3364 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3365                               void *data)
3366 {
3367         struct inode *inode = page->mapping->host;
3368         struct extent_page_data *epd = data;
3369         u64 start = page_offset(page);
3370         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3371         int ret;
3372         int nr = 0;
3373         size_t pg_offset = 0;
3374         loff_t i_size = i_size_read(inode);
3375         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3376         int write_flags;
3377         unsigned long nr_written = 0;
3378
3379         if (wbc->sync_mode == WB_SYNC_ALL)
3380                 write_flags = WRITE_SYNC;
3381         else
3382                 write_flags = WRITE;
3383
3384         trace___extent_writepage(page, inode, wbc);
3385
3386         WARN_ON(!PageLocked(page));
3387
3388         ClearPageError(page);
3389
3390         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3391         if (page->index > end_index ||
3392            (page->index == end_index && !pg_offset)) {
3393                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3394                 unlock_page(page);
3395                 return 0;
3396         }
3397
3398         if (page->index == end_index) {
3399                 char *userpage;
3400
3401                 userpage = kmap_atomic(page);
3402                 memset(userpage + pg_offset, 0,
3403                        PAGE_CACHE_SIZE - pg_offset);
3404                 kunmap_atomic(userpage);
3405                 flush_dcache_page(page);
3406         }
3407
3408         pg_offset = 0;
3409
3410         set_page_extent_mapped(page);
3411
3412         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3413         if (ret == 1)
3414                 goto done_unlocked;
3415         if (ret)
3416                 goto done;
3417
3418         ret = __extent_writepage_io(inode, page, wbc, epd,
3419                                     i_size, nr_written, write_flags, &nr);
3420         if (ret == 1)
3421                 goto done_unlocked;
3422
3423 done:
3424         if (nr == 0) {
3425                 /* make sure the mapping tag for page dirty gets cleared */
3426                 set_page_writeback(page);
3427                 end_page_writeback(page);
3428         }
3429         if (PageError(page)) {
3430                 ret = ret < 0 ? ret : -EIO;
3431                 end_extent_writepage(page, ret, start, page_end);
3432         }
3433         unlock_page(page);
3434         return ret;
3435
3436 done_unlocked:
3437         return 0;
3438 }
3439
3440 static int eb_wait(void *word)
3441 {
3442         io_schedule();
3443         return 0;
3444 }
3445
3446 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3447 {
3448         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3449                     TASK_UNINTERRUPTIBLE);
3450 }
3451
3452 static noinline_for_stack int
3453 lock_extent_buffer_for_io(struct extent_buffer *eb,
3454                           struct btrfs_fs_info *fs_info,
3455                           struct extent_page_data *epd)
3456 {
3457         unsigned long i, num_pages;
3458         int flush = 0;
3459         int ret = 0;
3460
3461         if (!btrfs_try_tree_write_lock(eb)) {
3462                 flush = 1;
3463                 flush_write_bio(epd);
3464                 btrfs_tree_lock(eb);
3465         }
3466
3467         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3468                 btrfs_tree_unlock(eb);
3469                 if (!epd->sync_io)
3470                         return 0;
3471                 if (!flush) {
3472                         flush_write_bio(epd);
3473                         flush = 1;
3474                 }
3475                 while (1) {
3476                         wait_on_extent_buffer_writeback(eb);
3477                         btrfs_tree_lock(eb);
3478                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3479                                 break;
3480                         btrfs_tree_unlock(eb);
3481                 }
3482         }
3483
3484         /*
3485          * We need to do this to prevent races in people who check if the eb is
3486          * under IO since we can end up having no IO bits set for a short period
3487          * of time.
3488          */
3489         spin_lock(&eb->refs_lock);
3490         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3491                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3492                 spin_unlock(&eb->refs_lock);
3493                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3494                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3495                                      -eb->len,
3496                                      fs_info->dirty_metadata_batch);
3497                 ret = 1;
3498         } else {
3499                 spin_unlock(&eb->refs_lock);
3500         }
3501
3502         btrfs_tree_unlock(eb);
3503
3504         if (!ret)
3505                 return ret;
3506
3507         num_pages = num_extent_pages(eb->start, eb->len);
3508         for (i = 0; i < num_pages; i++) {
3509                 struct page *p = extent_buffer_page(eb, i);
3510
3511                 if (!trylock_page(p)) {
3512                         if (!flush) {
3513                                 flush_write_bio(epd);
3514                                 flush = 1;
3515                         }
3516                         lock_page(p);
3517                 }
3518         }
3519
3520         return ret;
3521 }
3522
3523 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3524 {
3525         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3526         smp_mb__after_atomic();
3527         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3528 }
3529
3530 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3531 {
3532         struct bio_vec *bvec;
3533         struct extent_buffer *eb;
3534         int i, done;
3535
3536         bio_for_each_segment_all(bvec, bio, i) {
3537                 struct page *page = bvec->bv_page;
3538
3539                 eb = (struct extent_buffer *)page->private;
3540                 BUG_ON(!eb);
3541                 done = atomic_dec_and_test(&eb->io_pages);
3542
3543                 if (err || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3544                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3545                         ClearPageUptodate(page);
3546                         SetPageError(page);
3547                 }
3548
3549                 end_page_writeback(page);
3550
3551                 if (!done)
3552                         continue;
3553
3554                 end_extent_buffer_writeback(eb);
3555         }
3556
3557         bio_put(bio);
3558 }
3559
3560 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3561                         struct btrfs_fs_info *fs_info,
3562                         struct writeback_control *wbc,
3563                         struct extent_page_data *epd)
3564 {
3565         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3566         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3567         u64 offset = eb->start;
3568         unsigned long i, num_pages;
3569         unsigned long bio_flags = 0;
3570         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3571         int ret = 0;
3572
3573         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3574         num_pages = num_extent_pages(eb->start, eb->len);
3575         atomic_set(&eb->io_pages, num_pages);
3576         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3577                 bio_flags = EXTENT_BIO_TREE_LOG;
3578
3579         for (i = 0; i < num_pages; i++) {
3580                 struct page *p = extent_buffer_page(eb, i);
3581
3582                 clear_page_dirty_for_io(p);
3583                 set_page_writeback(p);
3584                 ret = submit_extent_page(rw, tree, p, offset >> 9,
3585                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3586                                          -1, end_bio_extent_buffer_writepage,
3587                                          0, epd->bio_flags, bio_flags);
3588                 epd->bio_flags = bio_flags;
3589                 if (ret) {
3590                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3591                         SetPageError(p);
3592                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3593                                 end_extent_buffer_writeback(eb);
3594                         ret = -EIO;
3595                         break;
3596                 }
3597                 offset += PAGE_CACHE_SIZE;
3598                 update_nr_written(p, wbc, 1);
3599                 unlock_page(p);
3600         }
3601
3602         if (unlikely(ret)) {
3603                 for (; i < num_pages; i++) {
3604                         struct page *p = extent_buffer_page(eb, i);
3605                         unlock_page(p);
3606                 }
3607         }
3608
3609         return ret;
3610 }
3611
3612 int btree_write_cache_pages(struct address_space *mapping,
3613                                    struct writeback_control *wbc)
3614 {
3615         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3616         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3617         struct extent_buffer *eb, *prev_eb = NULL;
3618         struct extent_page_data epd = {
3619                 .bio = NULL,
3620                 .tree = tree,
3621                 .extent_locked = 0,
3622                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3623                 .bio_flags = 0,
3624         };
3625         int ret = 0;
3626         int done = 0;
3627         int nr_to_write_done = 0;
3628         struct pagevec pvec;
3629         int nr_pages;
3630         pgoff_t index;
3631         pgoff_t end;            /* Inclusive */
3632         int scanned = 0;
3633         int tag;
3634
3635         pagevec_init(&pvec, 0);
3636         if (wbc->range_cyclic) {
3637                 index = mapping->writeback_index; /* Start from prev offset */
3638                 end = -1;
3639         } else {
3640                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3641                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3642                 scanned = 1;
3643         }
3644         if (wbc->sync_mode == WB_SYNC_ALL)
3645                 tag = PAGECACHE_TAG_TOWRITE;
3646         else
3647                 tag = PAGECACHE_TAG_DIRTY;
3648 retry:
3649         if (wbc->sync_mode == WB_SYNC_ALL)
3650                 tag_pages_for_writeback(mapping, index, end);
3651         while (!done && !nr_to_write_done && (index <= end) &&
3652                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3653                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3654                 unsigned i;
3655
3656                 scanned = 1;
3657                 for (i = 0; i < nr_pages; i++) {
3658                         struct page *page = pvec.pages[i];
3659
3660                         if (!PagePrivate(page))
3661                                 continue;
3662
3663                         if (!wbc->range_cyclic && page->index > end) {
3664                                 done = 1;
3665                                 break;
3666                         }
3667
3668                         spin_lock(&mapping->private_lock);
3669                         if (!PagePrivate(page)) {
3670                                 spin_unlock(&mapping->private_lock);
3671                                 continue;
3672                         }
3673
3674                         eb = (struct extent_buffer *)page->private;
3675
3676                         /*
3677                          * Shouldn't happen and normally this would be a BUG_ON
3678                          * but no sense in crashing the users box for something
3679                          * we can survive anyway.
3680                          */
3681                         if (WARN_ON(!eb)) {
3682                                 spin_unlock(&mapping->private_lock);
3683                                 continue;
3684                         }
3685
3686                         if (eb == prev_eb) {
3687                                 spin_unlock(&mapping->private_lock);
3688                                 continue;
3689                         }
3690
3691                         ret = atomic_inc_not_zero(&eb->refs);
3692                         spin_unlock(&mapping->private_lock);
3693                         if (!ret)
3694                                 continue;
3695
3696                         prev_eb = eb;
3697                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3698                         if (!ret) {
3699                                 free_extent_buffer(eb);
3700                                 continue;
3701                         }
3702
3703                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3704                         if (ret) {
3705                                 done = 1;
3706                                 free_extent_buffer(eb);
3707                                 break;
3708                         }
3709                         free_extent_buffer(eb);
3710
3711                         /*
3712                          * the filesystem may choose to bump up nr_to_write.
3713                          * We have to make sure to honor the new nr_to_write
3714                          * at any time
3715                          */
3716                         nr_to_write_done = wbc->nr_to_write <= 0;
3717                 }
3718                 pagevec_release(&pvec);
3719                 cond_resched();
3720         }
3721         if (!scanned && !done) {
3722                 /*
3723                  * We hit the last page and there is more work to be done: wrap
3724                  * back to the start of the file
3725                  */
3726                 scanned = 1;
3727                 index = 0;
3728                 goto retry;
3729         }
3730         flush_write_bio(&epd);
3731         return ret;
3732 }
3733
3734 /**
3735  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3736  * @mapping: address space structure to write
3737  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3738  * @writepage: function called for each page
3739  * @data: data passed to writepage function
3740  *
3741  * If a page is already under I/O, write_cache_pages() skips it, even
3742  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3743  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3744  * and msync() need to guarantee that all the data which was dirty at the time
3745  * the call was made get new I/O started against them.  If wbc->sync_mode is
3746  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3747  * existing IO to complete.
3748  */
3749 static int extent_write_cache_pages(struct extent_io_tree *tree,
3750                              struct address_space *mapping,
3751                              struct writeback_control *wbc,
3752                              writepage_t writepage, void *data,
3753                              void (*flush_fn)(void *))
3754 {
3755         struct inode *inode = mapping->host;
3756         int ret = 0;
3757         int done = 0;
3758         int err = 0;
3759         int nr_to_write_done = 0;
3760         struct pagevec pvec;
3761         int nr_pages;
3762         pgoff_t index;
3763         pgoff_t end;            /* Inclusive */
3764         int scanned = 0;
3765         int tag;
3766
3767         /*
3768          * We have to hold onto the inode so that ordered extents can do their
3769          * work when the IO finishes.  The alternative to this is failing to add
3770          * an ordered extent if the igrab() fails there and that is a huge pain
3771          * to deal with, so instead just hold onto the inode throughout the
3772          * writepages operation.  If it fails here we are freeing up the inode
3773          * anyway and we'd rather not waste our time writing out stuff that is
3774          * going to be truncated anyway.
3775          */
3776         if (!igrab(inode))
3777                 return 0;
3778
3779         pagevec_init(&pvec, 0);
3780         if (wbc->range_cyclic) {
3781                 index = mapping->writeback_index; /* Start from prev offset */
3782                 end = -1;
3783         } else {
3784                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3785                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3786                 scanned = 1;
3787         }
3788         if (wbc->sync_mode == WB_SYNC_ALL)
3789                 tag = PAGECACHE_TAG_TOWRITE;
3790         else
3791                 tag = PAGECACHE_TAG_DIRTY;
3792 retry:
3793         if (wbc->sync_mode == WB_SYNC_ALL)
3794                 tag_pages_for_writeback(mapping, index, end);
3795         while (!done && !nr_to_write_done && (index <= end) &&
3796                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3797                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3798                 unsigned i;
3799
3800                 scanned = 1;
3801                 for (i = 0; i < nr_pages; i++) {
3802                         struct page *page = pvec.pages[i];
3803
3804                         /*
3805                          * At this point we hold neither mapping->tree_lock nor
3806                          * lock on the page itself: the page may be truncated or
3807                          * invalidated (changing page->mapping to NULL), or even
3808                          * swizzled back from swapper_space to tmpfs file
3809                          * mapping
3810                          */
3811                         if (!trylock_page(page)) {
3812                                 flush_fn(data);
3813                                 lock_page(page);
3814                         }
3815
3816                         if (unlikely(page->mapping != mapping)) {
3817                                 unlock_page(page);
3818                                 continue;
3819                         }
3820
3821                         if (!wbc->range_cyclic && page->index > end) {
3822                                 done = 1;
3823                                 unlock_page(page);
3824                                 continue;
3825                         }
3826
3827                         if (wbc->sync_mode != WB_SYNC_NONE) {
3828                                 if (PageWriteback(page))
3829                                         flush_fn(data);
3830                                 wait_on_page_writeback(page);
3831                         }
3832
3833                         if (PageWriteback(page) ||
3834                             !clear_page_dirty_for_io(page)) {
3835                                 unlock_page(page);
3836                                 continue;
3837                         }
3838
3839                         ret = (*writepage)(page, wbc, data);
3840
3841                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3842                                 unlock_page(page);
3843                                 ret = 0;
3844                         }
3845                         if (!err && ret < 0)
3846                                 err = ret;
3847
3848                         /*
3849                          * the filesystem may choose to bump up nr_to_write.
3850                          * We have to make sure to honor the new nr_to_write
3851                          * at any time
3852                          */
3853                         nr_to_write_done = wbc->nr_to_write <= 0;
3854                 }
3855                 pagevec_release(&pvec);
3856                 cond_resched();
3857         }
3858         if (!scanned && !done && !err) {
3859                 /*
3860                  * We hit the last page and there is more work to be done: wrap
3861                  * back to the start of the file
3862                  */
3863                 scanned = 1;
3864                 index = 0;
3865                 goto retry;
3866         }
3867         btrfs_add_delayed_iput(inode);
3868         return err;
3869 }
3870
3871 static void flush_epd_write_bio(struct extent_page_data *epd)
3872 {
3873         if (epd->bio) {
3874                 int rw = WRITE;
3875                 int ret;
3876
3877                 if (epd->sync_io)
3878                         rw = WRITE_SYNC;
3879
3880                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
3881                 BUG_ON(ret < 0); /* -ENOMEM */
3882                 epd->bio = NULL;
3883         }
3884 }
3885
3886 static noinline void flush_write_bio(void *data)
3887 {
3888         struct extent_page_data *epd = data;
3889         flush_epd_write_bio(epd);
3890 }
3891
3892 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3893                           get_extent_t *get_extent,
3894                           struct writeback_control *wbc)
3895 {
3896         int ret;
3897         struct extent_page_data epd = {
3898                 .bio = NULL,
3899                 .tree = tree,
3900                 .get_extent = get_extent,
3901                 .extent_locked = 0,
3902                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3903                 .bio_flags = 0,
3904         };
3905
3906         ret = __extent_writepage(page, wbc, &epd);
3907
3908         flush_epd_write_bio(&epd);
3909         return ret;
3910 }
3911
3912 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3913                               u64 start, u64 end, get_extent_t *get_extent,
3914                               int mode)
3915 {
3916         int ret = 0;
3917         struct address_space *mapping = inode->i_mapping;
3918         struct page *page;
3919         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3920                 PAGE_CACHE_SHIFT;
3921
3922         struct extent_page_data epd = {
3923                 .bio = NULL,
3924                 .tree = tree,
3925                 .get_extent = get_extent,
3926                 .extent_locked = 1,
3927                 .sync_io = mode == WB_SYNC_ALL,
3928                 .bio_flags = 0,
3929         };
3930         struct writeback_control wbc_writepages = {
3931                 .sync_mode      = mode,
3932                 .nr_to_write    = nr_pages * 2,
3933                 .range_start    = start,
3934                 .range_end      = end + 1,
3935         };
3936
3937         while (start <= end) {
3938                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3939                 if (clear_page_dirty_for_io(page))
3940                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3941                 else {
3942                         if (tree->ops && tree->ops->writepage_end_io_hook)
3943                                 tree->ops->writepage_end_io_hook(page, start,
3944                                                  start + PAGE_CACHE_SIZE - 1,
3945                                                  NULL, 1);
3946                         unlock_page(page);
3947                 }
3948                 page_cache_release(page);
3949                 start += PAGE_CACHE_SIZE;
3950         }
3951
3952         flush_epd_write_bio(&epd);
3953         return ret;
3954 }
3955
3956 int extent_writepages(struct extent_io_tree *tree,
3957                       struct address_space *mapping,
3958                       get_extent_t *get_extent,
3959                       struct writeback_control *wbc)
3960 {
3961         int ret = 0;
3962         struct extent_page_data epd = {
3963                 .bio = NULL,
3964                 .tree = tree,
3965                 .get_extent = get_extent,
3966                 .extent_locked = 0,
3967                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3968                 .bio_flags = 0,
3969         };
3970
3971         ret = extent_write_cache_pages(tree, mapping, wbc,
3972                                        __extent_writepage, &epd,
3973                                        flush_write_bio);
3974         flush_epd_write_bio(&epd);
3975         return ret;
3976 }
3977
3978 int extent_readpages(struct extent_io_tree *tree,
3979                      struct address_space *mapping,
3980                      struct list_head *pages, unsigned nr_pages,
3981                      get_extent_t get_extent)
3982 {
3983         struct bio *bio = NULL;
3984         unsigned page_idx;
3985         unsigned long bio_flags = 0;
3986         struct page *pagepool[16];
3987         struct page *page;
3988         struct extent_map *em_cached = NULL;
3989         int nr = 0;
3990
3991         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3992                 page = list_entry(pages->prev, struct page, lru);
3993
3994                 prefetchw(&page->flags);
3995                 list_del(&page->lru);
3996                 if (add_to_page_cache_lru(page, mapping,
3997                                         page->index, GFP_NOFS)) {
3998                         page_cache_release(page);
3999                         continue;
4000                 }
4001
4002                 pagepool[nr++] = page;
4003                 if (nr < ARRAY_SIZE(pagepool))
4004                         continue;
4005                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4006                                    &bio, 0, &bio_flags, READ);
4007                 nr = 0;
4008         }
4009         if (nr)
4010                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4011                                    &bio, 0, &bio_flags, READ);
4012
4013         if (em_cached)
4014                 free_extent_map(em_cached);
4015
4016         BUG_ON(!list_empty(pages));
4017         if (bio)
4018                 return submit_one_bio(READ, bio, 0, bio_flags);
4019         return 0;
4020 }
4021
4022 /*
4023  * basic invalidatepage code, this waits on any locked or writeback
4024  * ranges corresponding to the page, and then deletes any extent state
4025  * records from the tree
4026  */
4027 int extent_invalidatepage(struct extent_io_tree *tree,
4028                           struct page *page, unsigned long offset)
4029 {
4030         struct extent_state *cached_state = NULL;
4031         u64 start = page_offset(page);
4032         u64 end = start + PAGE_CACHE_SIZE - 1;
4033         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4034
4035         start += ALIGN(offset, blocksize);
4036         if (start > end)
4037                 return 0;
4038
4039         lock_extent_bits(tree, start, end, 0, &cached_state);
4040         wait_on_page_writeback(page);
4041         clear_extent_bit(tree, start, end,
4042                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4043                          EXTENT_DO_ACCOUNTING,
4044                          1, 1, &cached_state, GFP_NOFS);
4045         return 0;
4046 }
4047
4048 /*
4049  * a helper for releasepage, this tests for areas of the page that
4050  * are locked or under IO and drops the related state bits if it is safe
4051  * to drop the page.
4052  */
4053 static int try_release_extent_state(struct extent_map_tree *map,
4054                                     struct extent_io_tree *tree,
4055                                     struct page *page, gfp_t mask)
4056 {
4057         u64 start = page_offset(page);
4058         u64 end = start + PAGE_CACHE_SIZE - 1;
4059         int ret = 1;
4060
4061         if (test_range_bit(tree, start, end,
4062                            EXTENT_IOBITS, 0, NULL))
4063                 ret = 0;
4064         else {
4065                 if ((mask & GFP_NOFS) == GFP_NOFS)
4066                         mask = GFP_NOFS;
4067                 /*
4068                  * at this point we can safely clear everything except the
4069                  * locked bit and the nodatasum bit
4070                  */
4071                 ret = clear_extent_bit(tree, start, end,
4072                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4073                                  0, 0, NULL, mask);
4074
4075                 /* if clear_extent_bit failed for enomem reasons,
4076                  * we can't allow the release to continue.
4077                  */
4078                 if (ret < 0)
4079                         ret = 0;
4080                 else
4081                         ret = 1;
4082         }
4083         return ret;
4084 }
4085
4086 /*
4087  * a helper for releasepage.  As long as there are no locked extents
4088  * in the range corresponding to the page, both state records and extent
4089  * map records are removed
4090  */
4091 int try_release_extent_mapping(struct extent_map_tree *map,
4092                                struct extent_io_tree *tree, struct page *page,
4093                                gfp_t mask)
4094 {
4095         struct extent_map *em;
4096         u64 start = page_offset(page);
4097         u64 end = start + PAGE_CACHE_SIZE - 1;
4098
4099         if ((mask & __GFP_WAIT) &&
4100             page->mapping->host->i_size > 16 * 1024 * 1024) {
4101                 u64 len;
4102                 while (start <= end) {
4103                         len = end - start + 1;
4104                         write_lock(&map->lock);
4105                         em = lookup_extent_mapping(map, start, len);
4106                         if (!em) {
4107                                 write_unlock(&map->lock);
4108                                 break;
4109                         }
4110                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4111                             em->start != start) {
4112                                 write_unlock(&map->lock);
4113                                 free_extent_map(em);
4114                                 break;
4115                         }
4116                         if (!test_range_bit(tree, em->start,
4117                                             extent_map_end(em) - 1,
4118                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4119                                             0, NULL)) {
4120                                 remove_extent_mapping(map, em);
4121                                 /* once for the rb tree */
4122                                 free_extent_map(em);
4123                         }
4124                         start = extent_map_end(em);
4125                         write_unlock(&map->lock);
4126
4127                         /* once for us */
4128                         free_extent_map(em);
4129                 }
4130         }
4131         return try_release_extent_state(map, tree, page, mask);
4132 }
4133
4134 /*
4135  * helper function for fiemap, which doesn't want to see any holes.
4136  * This maps until we find something past 'last'
4137  */
4138 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4139                                                 u64 offset,
4140                                                 u64 last,
4141                                                 get_extent_t *get_extent)
4142 {
4143         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4144         struct extent_map *em;
4145         u64 len;
4146
4147         if (offset >= last)
4148                 return NULL;
4149
4150         while (1) {
4151                 len = last - offset;
4152                 if (len == 0)
4153                         break;
4154                 len = ALIGN(len, sectorsize);
4155                 em = get_extent(inode, NULL, 0, offset, len, 0);
4156                 if (IS_ERR_OR_NULL(em))
4157                         return em;
4158
4159                 /* if this isn't a hole return it */
4160                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4161                     em->block_start != EXTENT_MAP_HOLE) {
4162                         return em;
4163                 }
4164
4165                 /* this is a hole, advance to the next extent */
4166                 offset = extent_map_end(em);
4167                 free_extent_map(em);
4168                 if (offset >= last)
4169                         break;
4170         }
4171         return NULL;
4172 }
4173
4174 static noinline int count_ext_ref(u64 inum, u64 offset, u64 root_id, void *ctx)
4175 {
4176         unsigned long cnt = *((unsigned long *)ctx);
4177
4178         cnt++;
4179         *((unsigned long *)ctx) = cnt;
4180
4181         /* Now we're sure that the extent is shared. */
4182         if (cnt > 1)
4183                 return 1;
4184         return 0;
4185 }
4186
4187 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4188                 __u64 start, __u64 len, get_extent_t *get_extent)
4189 {
4190         int ret = 0;
4191         u64 off = start;
4192         u64 max = start + len;
4193         u32 flags = 0;
4194         u32 found_type;
4195         u64 last;
4196         u64 last_for_get_extent = 0;
4197         u64 disko = 0;
4198         u64 isize = i_size_read(inode);
4199         struct btrfs_key found_key;
4200         struct extent_map *em = NULL;
4201         struct extent_state *cached_state = NULL;
4202         struct btrfs_path *path;
4203         int end = 0;
4204         u64 em_start = 0;
4205         u64 em_len = 0;
4206         u64 em_end = 0;
4207
4208         if (len == 0)
4209                 return -EINVAL;
4210
4211         path = btrfs_alloc_path();
4212         if (!path)
4213                 return -ENOMEM;
4214         path->leave_spinning = 1;
4215
4216         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
4217         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
4218
4219         /*
4220          * lookup the last file extent.  We're not using i_size here
4221          * because there might be preallocation past i_size
4222          */
4223         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
4224                                        path, btrfs_ino(inode), -1, 0);
4225         if (ret < 0) {
4226                 btrfs_free_path(path);
4227                 return ret;
4228         }
4229         WARN_ON(!ret);
4230         path->slots[0]--;
4231         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4232         found_type = btrfs_key_type(&found_key);
4233
4234         /* No extents, but there might be delalloc bits */
4235         if (found_key.objectid != btrfs_ino(inode) ||
4236             found_type != BTRFS_EXTENT_DATA_KEY) {
4237                 /* have to trust i_size as the end */
4238                 last = (u64)-1;
4239                 last_for_get_extent = isize;
4240         } else {
4241                 /*
4242                  * remember the start of the last extent.  There are a
4243                  * bunch of different factors that go into the length of the
4244                  * extent, so its much less complex to remember where it started
4245                  */
4246                 last = found_key.offset;
4247                 last_for_get_extent = last + 1;
4248         }
4249         btrfs_release_path(path);
4250
4251         /*
4252          * we might have some extents allocated but more delalloc past those
4253          * extents.  so, we trust isize unless the start of the last extent is
4254          * beyond isize
4255          */
4256         if (last < isize) {
4257                 last = (u64)-1;
4258                 last_for_get_extent = isize;
4259         }
4260
4261         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4262                          &cached_state);
4263
4264         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4265                                    get_extent);
4266         if (!em)
4267                 goto out;
4268         if (IS_ERR(em)) {
4269                 ret = PTR_ERR(em);
4270                 goto out;
4271         }
4272
4273         while (!end) {
4274                 u64 offset_in_extent = 0;
4275
4276                 /* break if the extent we found is outside the range */
4277                 if (em->start >= max || extent_map_end(em) < off)
4278                         break;
4279
4280                 /*
4281                  * get_extent may return an extent that starts before our
4282                  * requested range.  We have to make sure the ranges
4283                  * we return to fiemap always move forward and don't
4284                  * overlap, so adjust the offsets here
4285                  */
4286                 em_start = max(em->start, off);
4287
4288                 /*
4289                  * record the offset from the start of the extent
4290                  * for adjusting the disk offset below.  Only do this if the
4291                  * extent isn't compressed since our in ram offset may be past
4292                  * what we have actually allocated on disk.
4293                  */
4294                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4295                         offset_in_extent = em_start - em->start;
4296                 em_end = extent_map_end(em);
4297                 em_len = em_end - em_start;
4298                 disko = 0;
4299                 flags = 0;
4300
4301                 /*
4302                  * bump off for our next call to get_extent
4303                  */
4304                 off = extent_map_end(em);
4305                 if (off >= max)
4306                         end = 1;
4307
4308                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4309                         end = 1;
4310                         flags |= FIEMAP_EXTENT_LAST;
4311                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4312                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4313                                   FIEMAP_EXTENT_NOT_ALIGNED);
4314                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4315                         flags |= (FIEMAP_EXTENT_DELALLOC |
4316                                   FIEMAP_EXTENT_UNKNOWN);
4317                 } else {
4318                         unsigned long ref_cnt = 0;
4319
4320                         disko = em->block_start + offset_in_extent;
4321
4322                         /*
4323                          * As btrfs supports shared space, this information
4324                          * can be exported to userspace tools via
4325                          * flag FIEMAP_EXTENT_SHARED.
4326                          */
4327                         ret = iterate_inodes_from_logical(
4328                                         em->block_start,
4329                                         BTRFS_I(inode)->root->fs_info,
4330                                         path, count_ext_ref, &ref_cnt);
4331                         if (ret < 0 && ret != -ENOENT)
4332                                 goto out_free;
4333
4334                         if (ref_cnt > 1)
4335                                 flags |= FIEMAP_EXTENT_SHARED;
4336                 }
4337                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4338                         flags |= FIEMAP_EXTENT_ENCODED;
4339
4340                 free_extent_map(em);
4341                 em = NULL;
4342                 if ((em_start >= last) || em_len == (u64)-1 ||
4343                    (last == (u64)-1 && isize <= em_end)) {
4344                         flags |= FIEMAP_EXTENT_LAST;
4345                         end = 1;
4346                 }
4347
4348                 /* now scan forward to see if this is really the last extent. */
4349                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4350                                            get_extent);
4351                 if (IS_ERR(em)) {
4352                         ret = PTR_ERR(em);
4353                         goto out;
4354                 }
4355                 if (!em) {
4356                         flags |= FIEMAP_EXTENT_LAST;
4357                         end = 1;
4358                 }
4359                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4360                                               em_len, flags);
4361                 if (ret)
4362                         goto out_free;
4363         }
4364 out_free:
4365         free_extent_map(em);
4366 out:
4367         btrfs_free_path(path);
4368         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4369                              &cached_state, GFP_NOFS);
4370         return ret;
4371 }
4372
4373 static void __free_extent_buffer(struct extent_buffer *eb)
4374 {
4375         btrfs_leak_debug_del(&eb->leak_list);
4376         kmem_cache_free(extent_buffer_cache, eb);
4377 }
4378
4379 int extent_buffer_under_io(struct extent_buffer *eb)
4380 {
4381         return (atomic_read(&eb->io_pages) ||
4382                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4383                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4384 }
4385
4386 /*
4387  * Helper for releasing extent buffer page.
4388  */
4389 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4390                                                 unsigned long start_idx)
4391 {
4392         unsigned long index;
4393         unsigned long num_pages;
4394         struct page *page;
4395         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4396
4397         BUG_ON(extent_buffer_under_io(eb));
4398
4399         num_pages = num_extent_pages(eb->start, eb->len);
4400         index = start_idx + num_pages;
4401         if (start_idx >= index)
4402                 return;
4403
4404         do {
4405                 index--;
4406                 page = extent_buffer_page(eb, index);
4407                 if (page && mapped) {
4408                         spin_lock(&page->mapping->private_lock);
4409                         /*
4410                          * We do this since we'll remove the pages after we've
4411                          * removed the eb from the radix tree, so we could race
4412                          * and have this page now attached to the new eb.  So
4413                          * only clear page_private if it's still connected to
4414                          * this eb.
4415                          */
4416                         if (PagePrivate(page) &&
4417                             page->private == (unsigned long)eb) {
4418                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4419                                 BUG_ON(PageDirty(page));
4420                                 BUG_ON(PageWriteback(page));
4421                                 /*
4422                                  * We need to make sure we haven't be attached
4423                                  * to a new eb.
4424                                  */
4425                                 ClearPagePrivate(page);
4426                                 set_page_private(page, 0);
4427                                 /* One for the page private */
4428                                 page_cache_release(page);
4429                         }
4430                         spin_unlock(&page->mapping->private_lock);
4431
4432                 }
4433                 if (page) {
4434                         /* One for when we alloced the page */
4435                         page_cache_release(page);
4436                 }
4437         } while (index != start_idx);
4438 }
4439
4440 /*
4441  * Helper for releasing the extent buffer.
4442  */
4443 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4444 {
4445         btrfs_release_extent_buffer_page(eb, 0);
4446         __free_extent_buffer(eb);
4447 }
4448
4449 static struct extent_buffer *
4450 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4451                       unsigned long len, gfp_t mask)
4452 {
4453         struct extent_buffer *eb = NULL;
4454
4455         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
4456         if (eb == NULL)
4457                 return NULL;
4458         eb->start = start;
4459         eb->len = len;
4460         eb->fs_info = fs_info;
4461         eb->bflags = 0;
4462         rwlock_init(&eb->lock);
4463         atomic_set(&eb->write_locks, 0);
4464         atomic_set(&eb->read_locks, 0);
4465         atomic_set(&eb->blocking_readers, 0);
4466         atomic_set(&eb->blocking_writers, 0);
4467         atomic_set(&eb->spinning_readers, 0);
4468         atomic_set(&eb->spinning_writers, 0);
4469         eb->lock_nested = 0;
4470         init_waitqueue_head(&eb->write_lock_wq);
4471         init_waitqueue_head(&eb->read_lock_wq);
4472
4473         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4474
4475         spin_lock_init(&eb->refs_lock);
4476         atomic_set(&eb->refs, 1);
4477         atomic_set(&eb->io_pages, 0);
4478
4479         /*
4480          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4481          */
4482         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4483                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4484         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4485
4486         return eb;
4487 }
4488
4489 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4490 {
4491         unsigned long i;
4492         struct page *p;
4493         struct extent_buffer *new;
4494         unsigned long num_pages = num_extent_pages(src->start, src->len);
4495
4496         new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_NOFS);
4497         if (new == NULL)
4498                 return NULL;
4499
4500         for (i = 0; i < num_pages; i++) {
4501                 p = alloc_page(GFP_NOFS);
4502                 if (!p) {
4503                         btrfs_release_extent_buffer(new);
4504                         return NULL;
4505                 }
4506                 attach_extent_buffer_page(new, p);
4507                 WARN_ON(PageDirty(p));
4508                 SetPageUptodate(p);
4509                 new->pages[i] = p;
4510         }
4511
4512         copy_extent_buffer(new, src, 0, 0, src->len);
4513         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4514         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4515
4516         return new;
4517 }
4518
4519 struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4520 {
4521         struct extent_buffer *eb;
4522         unsigned long num_pages = num_extent_pages(0, len);
4523         unsigned long i;
4524
4525         eb = __alloc_extent_buffer(NULL, start, len, GFP_NOFS);
4526         if (!eb)
4527                 return NULL;
4528
4529         for (i = 0; i < num_pages; i++) {
4530                 eb->pages[i] = alloc_page(GFP_NOFS);
4531                 if (!eb->pages[i])
4532                         goto err;
4533         }
4534         set_extent_buffer_uptodate(eb);
4535         btrfs_set_header_nritems(eb, 0);
4536         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4537
4538         return eb;
4539 err:
4540         for (; i > 0; i--)
4541                 __free_page(eb->pages[i - 1]);
4542         __free_extent_buffer(eb);
4543         return NULL;
4544 }
4545
4546 static void check_buffer_tree_ref(struct extent_buffer *eb)
4547 {
4548         int refs;
4549         /* the ref bit is tricky.  We have to make sure it is set
4550          * if we have the buffer dirty.   Otherwise the
4551          * code to free a buffer can end up dropping a dirty
4552          * page
4553          *
4554          * Once the ref bit is set, it won't go away while the
4555          * buffer is dirty or in writeback, and it also won't
4556          * go away while we have the reference count on the
4557          * eb bumped.
4558          *
4559          * We can't just set the ref bit without bumping the
4560          * ref on the eb because free_extent_buffer might
4561          * see the ref bit and try to clear it.  If this happens
4562          * free_extent_buffer might end up dropping our original
4563          * ref by mistake and freeing the page before we are able
4564          * to add one more ref.
4565          *
4566          * So bump the ref count first, then set the bit.  If someone
4567          * beat us to it, drop the ref we added.
4568          */
4569         refs = atomic_read(&eb->refs);
4570         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4571                 return;
4572
4573         spin_lock(&eb->refs_lock);
4574         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4575                 atomic_inc(&eb->refs);
4576         spin_unlock(&eb->refs_lock);
4577 }
4578
4579 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4580                 struct page *accessed)
4581 {
4582         unsigned long num_pages, i;
4583
4584         check_buffer_tree_ref(eb);
4585
4586         num_pages = num_extent_pages(eb->start, eb->len);
4587         for (i = 0; i < num_pages; i++) {
4588                 struct page *p = extent_buffer_page(eb, i);
4589                 if (p != accessed)
4590                         mark_page_accessed(p);
4591         }
4592 }
4593
4594 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4595                                          u64 start)
4596 {
4597         struct extent_buffer *eb;
4598
4599         rcu_read_lock();
4600         eb = radix_tree_lookup(&fs_info->buffer_radix,
4601                                start >> PAGE_CACHE_SHIFT);
4602         if (eb && atomic_inc_not_zero(&eb->refs)) {
4603                 rcu_read_unlock();
4604                 mark_extent_buffer_accessed(eb, NULL);
4605                 return eb;
4606         }
4607         rcu_read_unlock();
4608
4609         return NULL;
4610 }
4611
4612 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4613 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4614                                                u64 start, unsigned long len)
4615 {
4616         struct extent_buffer *eb, *exists = NULL;
4617         int ret;
4618
4619         eb = find_extent_buffer(fs_info, start);
4620         if (eb)
4621                 return eb;
4622         eb = alloc_dummy_extent_buffer(start, len);
4623         if (!eb)
4624                 return NULL;
4625         eb->fs_info = fs_info;
4626 again:
4627         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4628         if (ret)
4629                 goto free_eb;
4630         spin_lock(&fs_info->buffer_lock);
4631         ret = radix_tree_insert(&fs_info->buffer_radix,
4632                                 start >> PAGE_CACHE_SHIFT, eb);
4633         spin_unlock(&fs_info->buffer_lock);
4634         radix_tree_preload_end();
4635         if (ret == -EEXIST) {
4636                 exists = find_extent_buffer(fs_info, start);
4637                 if (exists)
4638                         goto free_eb;
4639                 else
4640                         goto again;
4641         }
4642         check_buffer_tree_ref(eb);
4643         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4644
4645         /*
4646          * We will free dummy extent buffer's if they come into
4647          * free_extent_buffer with a ref count of 2, but if we are using this we
4648          * want the buffers to stay in memory until we're done with them, so
4649          * bump the ref count again.
4650          */
4651         atomic_inc(&eb->refs);
4652         return eb;
4653 free_eb:
4654         btrfs_release_extent_buffer(eb);
4655         return exists;
4656 }
4657 #endif
4658
4659 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4660                                           u64 start, unsigned long len)
4661 {
4662         unsigned long num_pages = num_extent_pages(start, len);
4663         unsigned long i;
4664         unsigned long index = start >> PAGE_CACHE_SHIFT;
4665         struct extent_buffer *eb;
4666         struct extent_buffer *exists = NULL;
4667         struct page *p;
4668         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4669         int uptodate = 1;
4670         int ret;
4671
4672         eb = find_extent_buffer(fs_info, start);
4673         if (eb)
4674                 return eb;
4675
4676         eb = __alloc_extent_buffer(fs_info, start, len, GFP_NOFS);
4677         if (!eb)
4678                 return NULL;
4679
4680         for (i = 0; i < num_pages; i++, index++) {
4681                 p = find_or_create_page(mapping, index, GFP_NOFS);
4682                 if (!p)
4683                         goto free_eb;
4684
4685                 spin_lock(&mapping->private_lock);
4686                 if (PagePrivate(p)) {
4687                         /*
4688                          * We could have already allocated an eb for this page
4689                          * and attached one so lets see if we can get a ref on
4690                          * the existing eb, and if we can we know it's good and
4691                          * we can just return that one, else we know we can just
4692                          * overwrite page->private.
4693                          */
4694                         exists = (struct extent_buffer *)p->private;
4695                         if (atomic_inc_not_zero(&exists->refs)) {
4696                                 spin_unlock(&mapping->private_lock);
4697                                 unlock_page(p);
4698                                 page_cache_release(p);
4699                                 mark_extent_buffer_accessed(exists, p);
4700                                 goto free_eb;
4701                         }
4702
4703                         /*
4704                          * Do this so attach doesn't complain and we need to
4705                          * drop the ref the old guy had.
4706                          */
4707                         ClearPagePrivate(p);
4708                         WARN_ON(PageDirty(p));
4709                         page_cache_release(p);
4710                 }
4711                 attach_extent_buffer_page(eb, p);
4712                 spin_unlock(&mapping->private_lock);
4713                 WARN_ON(PageDirty(p));
4714                 eb->pages[i] = p;
4715                 if (!PageUptodate(p))
4716                         uptodate = 0;
4717
4718                 /*
4719                  * see below about how we avoid a nasty race with release page
4720                  * and why we unlock later
4721                  */
4722         }
4723         if (uptodate)
4724                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4725 again:
4726         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4727         if (ret)
4728                 goto free_eb;
4729
4730         spin_lock(&fs_info->buffer_lock);
4731         ret = radix_tree_insert(&fs_info->buffer_radix,
4732                                 start >> PAGE_CACHE_SHIFT, eb);
4733         spin_unlock(&fs_info->buffer_lock);
4734         radix_tree_preload_end();
4735         if (ret == -EEXIST) {
4736                 exists = find_extent_buffer(fs_info, start);
4737                 if (exists)
4738                         goto free_eb;
4739                 else
4740                         goto again;
4741         }
4742         /* add one reference for the tree */
4743         check_buffer_tree_ref(eb);
4744         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4745
4746         /*
4747          * there is a race where release page may have
4748          * tried to find this extent buffer in the radix
4749          * but failed.  It will tell the VM it is safe to
4750          * reclaim the, and it will clear the page private bit.
4751          * We must make sure to set the page private bit properly
4752          * after the extent buffer is in the radix tree so
4753          * it doesn't get lost
4754          */
4755         SetPageChecked(eb->pages[0]);
4756         for (i = 1; i < num_pages; i++) {
4757                 p = extent_buffer_page(eb, i);
4758                 ClearPageChecked(p);
4759                 unlock_page(p);
4760         }
4761         unlock_page(eb->pages[0]);
4762         return eb;
4763
4764 free_eb:
4765         for (i = 0; i < num_pages; i++) {
4766                 if (eb->pages[i])
4767                         unlock_page(eb->pages[i]);
4768         }
4769
4770         WARN_ON(!atomic_dec_and_test(&eb->refs));
4771         btrfs_release_extent_buffer(eb);
4772         return exists;
4773 }
4774
4775 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4776 {
4777         struct extent_buffer *eb =
4778                         container_of(head, struct extent_buffer, rcu_head);
4779
4780         __free_extent_buffer(eb);
4781 }
4782
4783 /* Expects to have eb->eb_lock already held */
4784 static int release_extent_buffer(struct extent_buffer *eb)
4785 {
4786         WARN_ON(atomic_read(&eb->refs) == 0);
4787         if (atomic_dec_and_test(&eb->refs)) {
4788                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4789                         struct btrfs_fs_info *fs_info = eb->fs_info;
4790
4791                         spin_unlock(&eb->refs_lock);
4792
4793                         spin_lock(&fs_info->buffer_lock);
4794                         radix_tree_delete(&fs_info->buffer_radix,
4795                                           eb->start >> PAGE_CACHE_SHIFT);
4796                         spin_unlock(&fs_info->buffer_lock);
4797                 } else {
4798                         spin_unlock(&eb->refs_lock);
4799                 }
4800
4801                 /* Should be safe to release our pages at this point */
4802                 btrfs_release_extent_buffer_page(eb, 0);
4803                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4804                 return 1;
4805         }
4806         spin_unlock(&eb->refs_lock);
4807
4808         return 0;
4809 }
4810
4811 void free_extent_buffer(struct extent_buffer *eb)
4812 {
4813         int refs;
4814         int old;
4815         if (!eb)
4816                 return;
4817
4818         while (1) {
4819                 refs = atomic_read(&eb->refs);
4820                 if (refs <= 3)
4821                         break;
4822                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
4823                 if (old == refs)
4824                         return;
4825         }
4826
4827         spin_lock(&eb->refs_lock);
4828         if (atomic_read(&eb->refs) == 2 &&
4829             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4830                 atomic_dec(&eb->refs);
4831
4832         if (atomic_read(&eb->refs) == 2 &&
4833             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4834             !extent_buffer_under_io(eb) &&
4835             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4836                 atomic_dec(&eb->refs);
4837
4838         /*
4839          * I know this is terrible, but it's temporary until we stop tracking
4840          * the uptodate bits and such for the extent buffers.
4841          */
4842         release_extent_buffer(eb);
4843 }
4844
4845 void free_extent_buffer_stale(struct extent_buffer *eb)
4846 {
4847         if (!eb)
4848                 return;
4849
4850         spin_lock(&eb->refs_lock);
4851         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4852
4853         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4854             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4855                 atomic_dec(&eb->refs);
4856         release_extent_buffer(eb);
4857 }
4858
4859 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4860 {
4861         unsigned long i;
4862         unsigned long num_pages;
4863         struct page *page;
4864
4865         num_pages = num_extent_pages(eb->start, eb->len);
4866
4867         for (i = 0; i < num_pages; i++) {
4868                 page = extent_buffer_page(eb, i);
4869                 if (!PageDirty(page))
4870                         continue;
4871
4872                 lock_page(page);
4873                 WARN_ON(!PagePrivate(page));
4874
4875                 clear_page_dirty_for_io(page);
4876                 spin_lock_irq(&page->mapping->tree_lock);
4877                 if (!PageDirty(page)) {
4878                         radix_tree_tag_clear(&page->mapping->page_tree,
4879                                                 page_index(page),
4880                                                 PAGECACHE_TAG_DIRTY);
4881                 }
4882                 spin_unlock_irq(&page->mapping->tree_lock);
4883                 ClearPageError(page);
4884                 unlock_page(page);
4885         }
4886         WARN_ON(atomic_read(&eb->refs) == 0);
4887 }
4888
4889 int set_extent_buffer_dirty(struct extent_buffer *eb)
4890 {
4891         unsigned long i;
4892         unsigned long num_pages;
4893         int was_dirty = 0;
4894
4895         check_buffer_tree_ref(eb);
4896
4897         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4898
4899         num_pages = num_extent_pages(eb->start, eb->len);
4900         WARN_ON(atomic_read(&eb->refs) == 0);
4901         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4902
4903         for (i = 0; i < num_pages; i++)
4904                 set_page_dirty(extent_buffer_page(eb, i));
4905         return was_dirty;
4906 }
4907
4908 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4909 {
4910         unsigned long i;
4911         struct page *page;
4912         unsigned long num_pages;
4913
4914         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4915         num_pages = num_extent_pages(eb->start, eb->len);
4916         for (i = 0; i < num_pages; i++) {
4917                 page = extent_buffer_page(eb, i);
4918                 if (page)
4919                         ClearPageUptodate(page);
4920         }
4921         return 0;
4922 }
4923
4924 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4925 {
4926         unsigned long i;
4927         struct page *page;
4928         unsigned long num_pages;
4929
4930         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4931         num_pages = num_extent_pages(eb->start, eb->len);
4932         for (i = 0; i < num_pages; i++) {
4933                 page = extent_buffer_page(eb, i);
4934                 SetPageUptodate(page);
4935         }
4936         return 0;
4937 }
4938
4939 int extent_buffer_uptodate(struct extent_buffer *eb)
4940 {
4941         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4942 }
4943
4944 int read_extent_buffer_pages(struct extent_io_tree *tree,
4945                              struct extent_buffer *eb, u64 start, int wait,
4946                              get_extent_t *get_extent, int mirror_num)
4947 {
4948         unsigned long i;
4949         unsigned long start_i;
4950         struct page *page;
4951         int err;
4952         int ret = 0;
4953         int locked_pages = 0;
4954         int all_uptodate = 1;
4955         unsigned long num_pages;
4956         unsigned long num_reads = 0;
4957         struct bio *bio = NULL;
4958         unsigned long bio_flags = 0;
4959
4960         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4961                 return 0;
4962
4963         if (start) {
4964                 WARN_ON(start < eb->start);
4965                 start_i = (start >> PAGE_CACHE_SHIFT) -
4966                         (eb->start >> PAGE_CACHE_SHIFT);
4967         } else {
4968                 start_i = 0;
4969         }
4970
4971         num_pages = num_extent_pages(eb->start, eb->len);
4972         for (i = start_i; i < num_pages; i++) {
4973                 page = extent_buffer_page(eb, i);
4974                 if (wait == WAIT_NONE) {
4975                         if (!trylock_page(page))
4976                                 goto unlock_exit;
4977                 } else {
4978                         lock_page(page);
4979                 }
4980                 locked_pages++;
4981                 if (!PageUptodate(page)) {
4982                         num_reads++;
4983                         all_uptodate = 0;
4984                 }
4985         }
4986         if (all_uptodate) {
4987                 if (start_i == 0)
4988                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4989                 goto unlock_exit;
4990         }
4991
4992         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4993         eb->read_mirror = 0;
4994         atomic_set(&eb->io_pages, num_reads);
4995         for (i = start_i; i < num_pages; i++) {
4996                 page = extent_buffer_page(eb, i);
4997                 if (!PageUptodate(page)) {
4998                         ClearPageError(page);
4999                         err = __extent_read_full_page(tree, page,
5000                                                       get_extent, &bio,
5001                                                       mirror_num, &bio_flags,
5002                                                       READ | REQ_META);
5003                         if (err)
5004                                 ret = err;
5005                 } else {
5006                         unlock_page(page);
5007                 }
5008         }
5009
5010         if (bio) {
5011                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
5012                                      bio_flags);
5013                 if (err)
5014                         return err;
5015         }
5016
5017         if (ret || wait != WAIT_COMPLETE)
5018                 return ret;
5019
5020         for (i = start_i; i < num_pages; i++) {
5021                 page = extent_buffer_page(eb, i);
5022                 wait_on_page_locked(page);
5023                 if (!PageUptodate(page))
5024                         ret = -EIO;
5025         }
5026
5027         return ret;
5028
5029 unlock_exit:
5030         i = start_i;
5031         while (locked_pages > 0) {
5032                 page = extent_buffer_page(eb, i);
5033                 i++;
5034                 unlock_page(page);
5035                 locked_pages--;
5036         }
5037         return ret;
5038 }
5039
5040 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5041                         unsigned long start,
5042                         unsigned long len)
5043 {
5044         size_t cur;
5045         size_t offset;
5046         struct page *page;
5047         char *kaddr;
5048         char *dst = (char *)dstv;
5049         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5050         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5051
5052         WARN_ON(start > eb->len);
5053         WARN_ON(start + len > eb->start + eb->len);
5054
5055         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5056
5057         while (len > 0) {
5058                 page = extent_buffer_page(eb, i);
5059
5060                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5061                 kaddr = page_address(page);
5062                 memcpy(dst, kaddr + offset, cur);
5063
5064                 dst += cur;
5065                 len -= cur;
5066                 offset = 0;
5067                 i++;
5068         }
5069 }
5070
5071 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5072                                unsigned long min_len, char **map,
5073                                unsigned long *map_start,
5074                                unsigned long *map_len)
5075 {
5076         size_t offset = start & (PAGE_CACHE_SIZE - 1);
5077         char *kaddr;
5078         struct page *p;
5079         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5080         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5081         unsigned long end_i = (start_offset + start + min_len - 1) >>
5082                 PAGE_CACHE_SHIFT;
5083
5084         if (i != end_i)
5085                 return -EINVAL;
5086
5087         if (i == 0) {
5088                 offset = start_offset;
5089                 *map_start = 0;
5090         } else {
5091                 offset = 0;
5092                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
5093         }
5094
5095         if (start + min_len > eb->len) {
5096                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
5097                        "wanted %lu %lu\n",
5098                        eb->start, eb->len, start, min_len);
5099                 return -EINVAL;
5100         }
5101
5102         p = extent_buffer_page(eb, i);
5103         kaddr = page_address(p);
5104         *map = kaddr + offset;
5105         *map_len = PAGE_CACHE_SIZE - offset;
5106         return 0;
5107 }
5108
5109 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5110                           unsigned long start,
5111                           unsigned long len)
5112 {
5113         size_t cur;
5114         size_t offset;
5115         struct page *page;
5116         char *kaddr;
5117         char *ptr = (char *)ptrv;
5118         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5119         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5120         int ret = 0;
5121
5122         WARN_ON(start > eb->len);
5123         WARN_ON(start + len > eb->start + eb->len);
5124
5125         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5126
5127         while (len > 0) {
5128                 page = extent_buffer_page(eb, i);
5129
5130                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5131
5132                 kaddr = page_address(page);
5133                 ret = memcmp(ptr, kaddr + offset, cur);
5134                 if (ret)
5135                         break;
5136
5137                 ptr += cur;
5138                 len -= cur;
5139                 offset = 0;
5140                 i++;
5141         }
5142         return ret;
5143 }
5144
5145 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5146                          unsigned long start, unsigned long len)
5147 {
5148         size_t cur;
5149         size_t offset;
5150         struct page *page;
5151         char *kaddr;
5152         char *src = (char *)srcv;
5153         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5154         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5155
5156         WARN_ON(start > eb->len);
5157         WARN_ON(start + len > eb->start + eb->len);
5158
5159         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5160
5161         while (len > 0) {
5162                 page = extent_buffer_page(eb, i);
5163                 WARN_ON(!PageUptodate(page));
5164
5165                 cur = min(len, PAGE_CACHE_SIZE - offset);
5166                 kaddr = page_address(page);
5167                 memcpy(kaddr + offset, src, cur);
5168
5169                 src += cur;
5170                 len -= cur;
5171                 offset = 0;
5172                 i++;
5173         }
5174 }
5175
5176 void memset_extent_buffer(struct extent_buffer *eb, char c,
5177                           unsigned long start, unsigned long len)
5178 {
5179         size_t cur;
5180         size_t offset;
5181         struct page *page;
5182         char *kaddr;
5183         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5184         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5185
5186         WARN_ON(start > eb->len);
5187         WARN_ON(start + len > eb->start + eb->len);
5188
5189         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5190
5191         while (len > 0) {
5192                 page = extent_buffer_page(eb, i);
5193                 WARN_ON(!PageUptodate(page));
5194
5195                 cur = min(len, PAGE_CACHE_SIZE - offset);
5196                 kaddr = page_address(page);
5197                 memset(kaddr + offset, c, cur);
5198
5199                 len -= cur;
5200                 offset = 0;
5201                 i++;
5202         }
5203 }
5204
5205 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5206                         unsigned long dst_offset, unsigned long src_offset,
5207                         unsigned long len)
5208 {
5209         u64 dst_len = dst->len;
5210         size_t cur;
5211         size_t offset;
5212         struct page *page;
5213         char *kaddr;
5214         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5215         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5216
5217         WARN_ON(src->len != dst_len);
5218
5219         offset = (start_offset + dst_offset) &
5220                 (PAGE_CACHE_SIZE - 1);
5221
5222         while (len > 0) {
5223                 page = extent_buffer_page(dst, i);
5224                 WARN_ON(!PageUptodate(page));
5225
5226                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5227
5228                 kaddr = page_address(page);
5229                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5230
5231                 src_offset += cur;
5232                 len -= cur;
5233                 offset = 0;
5234                 i++;
5235         }
5236 }
5237
5238 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5239 {
5240         unsigned long distance = (src > dst) ? src - dst : dst - src;
5241         return distance < len;
5242 }
5243
5244 static void copy_pages(struct page *dst_page, struct page *src_page,
5245                        unsigned long dst_off, unsigned long src_off,
5246                        unsigned long len)
5247 {
5248         char *dst_kaddr = page_address(dst_page);
5249         char *src_kaddr;
5250         int must_memmove = 0;
5251
5252         if (dst_page != src_page) {
5253                 src_kaddr = page_address(src_page);
5254         } else {
5255                 src_kaddr = dst_kaddr;
5256                 if (areas_overlap(src_off, dst_off, len))
5257                         must_memmove = 1;
5258         }
5259
5260         if (must_memmove)
5261                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5262         else
5263                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5264 }
5265
5266 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5267                            unsigned long src_offset, unsigned long len)
5268 {
5269         size_t cur;
5270         size_t dst_off_in_page;
5271         size_t src_off_in_page;
5272         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5273         unsigned long dst_i;
5274         unsigned long src_i;
5275
5276         if (src_offset + len > dst->len) {
5277                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5278                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5279                 BUG_ON(1);
5280         }
5281         if (dst_offset + len > dst->len) {
5282                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5283                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5284                 BUG_ON(1);
5285         }
5286
5287         while (len > 0) {
5288                 dst_off_in_page = (start_offset + dst_offset) &
5289                         (PAGE_CACHE_SIZE - 1);
5290                 src_off_in_page = (start_offset + src_offset) &
5291                         (PAGE_CACHE_SIZE - 1);
5292
5293                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5294                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5295
5296                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5297                                                src_off_in_page));
5298                 cur = min_t(unsigned long, cur,
5299                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5300
5301                 copy_pages(extent_buffer_page(dst, dst_i),
5302                            extent_buffer_page(dst, src_i),
5303                            dst_off_in_page, src_off_in_page, cur);
5304
5305                 src_offset += cur;
5306                 dst_offset += cur;
5307                 len -= cur;
5308         }
5309 }
5310
5311 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5312                            unsigned long src_offset, unsigned long len)
5313 {
5314         size_t cur;
5315         size_t dst_off_in_page;
5316         size_t src_off_in_page;
5317         unsigned long dst_end = dst_offset + len - 1;
5318         unsigned long src_end = src_offset + len - 1;
5319         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5320         unsigned long dst_i;
5321         unsigned long src_i;
5322
5323         if (src_offset + len > dst->len) {
5324                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5325                        "len %lu len %lu\n", src_offset, len, dst->len);
5326                 BUG_ON(1);
5327         }
5328         if (dst_offset + len > dst->len) {
5329                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5330                        "len %lu len %lu\n", dst_offset, len, dst->len);
5331                 BUG_ON(1);
5332         }
5333         if (dst_offset < src_offset) {
5334                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5335                 return;
5336         }
5337         while (len > 0) {
5338                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5339                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5340
5341                 dst_off_in_page = (start_offset + dst_end) &
5342                         (PAGE_CACHE_SIZE - 1);
5343                 src_off_in_page = (start_offset + src_end) &
5344                         (PAGE_CACHE_SIZE - 1);
5345
5346                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5347                 cur = min(cur, dst_off_in_page + 1);
5348                 copy_pages(extent_buffer_page(dst, dst_i),
5349                            extent_buffer_page(dst, src_i),
5350                            dst_off_in_page - cur + 1,
5351                            src_off_in_page - cur + 1, cur);
5352
5353                 dst_end -= cur;
5354                 src_end -= cur;
5355                 len -= cur;
5356         }
5357 }
5358
5359 int try_release_extent_buffer(struct page *page)
5360 {
5361         struct extent_buffer *eb;
5362
5363         /*
5364          * We need to make sure noboody is attaching this page to an eb right
5365          * now.
5366          */
5367         spin_lock(&page->mapping->private_lock);
5368         if (!PagePrivate(page)) {
5369                 spin_unlock(&page->mapping->private_lock);
5370                 return 1;
5371         }
5372
5373         eb = (struct extent_buffer *)page->private;
5374         BUG_ON(!eb);
5375
5376         /*
5377          * This is a little awful but should be ok, we need to make sure that
5378          * the eb doesn't disappear out from under us while we're looking at
5379          * this page.
5380          */
5381         spin_lock(&eb->refs_lock);
5382         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5383                 spin_unlock(&eb->refs_lock);
5384                 spin_unlock(&page->mapping->private_lock);
5385                 return 0;
5386         }
5387         spin_unlock(&page->mapping->private_lock);
5388
5389         /*
5390          * If tree ref isn't set then we know the ref on this eb is a real ref,
5391          * so just return, this page will likely be freed soon anyway.
5392          */
5393         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5394                 spin_unlock(&eb->refs_lock);
5395                 return 0;
5396         }
5397
5398         return release_extent_buffer(eb);
5399 }