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 = 0;
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                                 offset += len;
2536                                 continue;
2537                         }
2538                 }
2539 readpage_ok:
2540                 if (likely(uptodate)) {
2541                         loff_t i_size = i_size_read(inode);
2542                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2543                         unsigned offset;
2544
2545                         /* Zero out the end if this page straddles i_size */
2546                         offset = i_size & (PAGE_CACHE_SIZE-1);
2547                         if (page->index == end_index && offset)
2548                                 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2549                         SetPageUptodate(page);
2550                 } else {
2551                         ClearPageUptodate(page);
2552                         SetPageError(page);
2553                 }
2554                 unlock_page(page);
2555                 offset += len;
2556
2557                 if (unlikely(!uptodate)) {
2558                         if (extent_len) {
2559                                 endio_readpage_release_extent(tree,
2560                                                               extent_start,
2561                                                               extent_len, 1);
2562                                 extent_start = 0;
2563                                 extent_len = 0;
2564                         }
2565                         endio_readpage_release_extent(tree, start,
2566                                                       end - start + 1, 0);
2567                 } else if (!extent_len) {
2568                         extent_start = start;
2569                         extent_len = end + 1 - start;
2570                 } else if (extent_start + extent_len == start) {
2571                         extent_len += end + 1 - start;
2572                 } else {
2573                         endio_readpage_release_extent(tree, extent_start,
2574                                                       extent_len, uptodate);
2575                         extent_start = start;
2576                         extent_len = end + 1 - start;
2577                 }
2578         }
2579
2580         if (extent_len)
2581                 endio_readpage_release_extent(tree, extent_start, extent_len,
2582                                               uptodate);
2583         if (io_bio->end_io)
2584                 io_bio->end_io(io_bio, err);
2585         bio_put(bio);
2586 }
2587
2588 /*
2589  * this allocates from the btrfs_bioset.  We're returning a bio right now
2590  * but you can call btrfs_io_bio for the appropriate container_of magic
2591  */
2592 struct bio *
2593 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2594                 gfp_t gfp_flags)
2595 {
2596         struct btrfs_io_bio *btrfs_bio;
2597         struct bio *bio;
2598
2599         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2600
2601         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2602                 while (!bio && (nr_vecs /= 2)) {
2603                         bio = bio_alloc_bioset(gfp_flags,
2604                                                nr_vecs, btrfs_bioset);
2605                 }
2606         }
2607
2608         if (bio) {
2609                 bio->bi_bdev = bdev;
2610                 bio->bi_iter.bi_sector = first_sector;
2611                 btrfs_bio = btrfs_io_bio(bio);
2612                 btrfs_bio->csum = NULL;
2613                 btrfs_bio->csum_allocated = NULL;
2614                 btrfs_bio->end_io = NULL;
2615         }
2616         return bio;
2617 }
2618
2619 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2620 {
2621         return bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2622 }
2623
2624
2625 /* this also allocates from the btrfs_bioset */
2626 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2627 {
2628         struct btrfs_io_bio *btrfs_bio;
2629         struct bio *bio;
2630
2631         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2632         if (bio) {
2633                 btrfs_bio = btrfs_io_bio(bio);
2634                 btrfs_bio->csum = NULL;
2635                 btrfs_bio->csum_allocated = NULL;
2636                 btrfs_bio->end_io = NULL;
2637         }
2638         return bio;
2639 }
2640
2641
2642 static int __must_check submit_one_bio(int rw, struct bio *bio,
2643                                        int mirror_num, unsigned long bio_flags)
2644 {
2645         int ret = 0;
2646         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2647         struct page *page = bvec->bv_page;
2648         struct extent_io_tree *tree = bio->bi_private;
2649         u64 start;
2650
2651         start = page_offset(page) + bvec->bv_offset;
2652
2653         bio->bi_private = NULL;
2654
2655         bio_get(bio);
2656
2657         if (tree->ops && tree->ops->submit_bio_hook)
2658                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2659                                            mirror_num, bio_flags, start);
2660         else
2661                 btrfsic_submit_bio(rw, bio);
2662
2663         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2664                 ret = -EOPNOTSUPP;
2665         bio_put(bio);
2666         return ret;
2667 }
2668
2669 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2670                      unsigned long offset, size_t size, struct bio *bio,
2671                      unsigned long bio_flags)
2672 {
2673         int ret = 0;
2674         if (tree->ops && tree->ops->merge_bio_hook)
2675                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2676                                                 bio_flags);
2677         BUG_ON(ret < 0);
2678         return ret;
2679
2680 }
2681
2682 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2683                               struct page *page, sector_t sector,
2684                               size_t size, unsigned long offset,
2685                               struct block_device *bdev,
2686                               struct bio **bio_ret,
2687                               unsigned long max_pages,
2688                               bio_end_io_t end_io_func,
2689                               int mirror_num,
2690                               unsigned long prev_bio_flags,
2691                               unsigned long bio_flags)
2692 {
2693         int ret = 0;
2694         struct bio *bio;
2695         int nr;
2696         int contig = 0;
2697         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2698         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2699         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2700
2701         if (bio_ret && *bio_ret) {
2702                 bio = *bio_ret;
2703                 if (old_compressed)
2704                         contig = bio->bi_iter.bi_sector == sector;
2705                 else
2706                         contig = bio_end_sector(bio) == sector;
2707
2708                 if (prev_bio_flags != bio_flags || !contig ||
2709                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2710                     bio_add_page(bio, page, page_size, offset) < page_size) {
2711                         ret = submit_one_bio(rw, bio, mirror_num,
2712                                              prev_bio_flags);
2713                         if (ret < 0)
2714                                 return ret;
2715                         bio = NULL;
2716                 } else {
2717                         return 0;
2718                 }
2719         }
2720         if (this_compressed)
2721                 nr = BIO_MAX_PAGES;
2722         else
2723                 nr = bio_get_nr_vecs(bdev);
2724
2725         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2726         if (!bio)
2727                 return -ENOMEM;
2728
2729         bio_add_page(bio, page, page_size, offset);
2730         bio->bi_end_io = end_io_func;
2731         bio->bi_private = tree;
2732
2733         if (bio_ret)
2734                 *bio_ret = bio;
2735         else
2736                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2737
2738         return ret;
2739 }
2740
2741 static void attach_extent_buffer_page(struct extent_buffer *eb,
2742                                       struct page *page)
2743 {
2744         if (!PagePrivate(page)) {
2745                 SetPagePrivate(page);
2746                 page_cache_get(page);
2747                 set_page_private(page, (unsigned long)eb);
2748         } else {
2749                 WARN_ON(page->private != (unsigned long)eb);
2750         }
2751 }
2752
2753 void set_page_extent_mapped(struct page *page)
2754 {
2755         if (!PagePrivate(page)) {
2756                 SetPagePrivate(page);
2757                 page_cache_get(page);
2758                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2759         }
2760 }
2761
2762 static struct extent_map *
2763 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2764                  u64 start, u64 len, get_extent_t *get_extent,
2765                  struct extent_map **em_cached)
2766 {
2767         struct extent_map *em;
2768
2769         if (em_cached && *em_cached) {
2770                 em = *em_cached;
2771                 if (extent_map_in_tree(em) && start >= em->start &&
2772                     start < extent_map_end(em)) {
2773                         atomic_inc(&em->refs);
2774                         return em;
2775                 }
2776
2777                 free_extent_map(em);
2778                 *em_cached = NULL;
2779         }
2780
2781         em = get_extent(inode, page, pg_offset, start, len, 0);
2782         if (em_cached && !IS_ERR_OR_NULL(em)) {
2783                 BUG_ON(*em_cached);
2784                 atomic_inc(&em->refs);
2785                 *em_cached = em;
2786         }
2787         return em;
2788 }
2789 /*
2790  * basic readpage implementation.  Locked extent state structs are inserted
2791  * into the tree that are removed when the IO is done (by the end_io
2792  * handlers)
2793  * XXX JDM: This needs looking at to ensure proper page locking
2794  */
2795 static int __do_readpage(struct extent_io_tree *tree,
2796                          struct page *page,
2797                          get_extent_t *get_extent,
2798                          struct extent_map **em_cached,
2799                          struct bio **bio, int mirror_num,
2800                          unsigned long *bio_flags, int rw)
2801 {
2802         struct inode *inode = page->mapping->host;
2803         u64 start = page_offset(page);
2804         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2805         u64 end;
2806         u64 cur = start;
2807         u64 extent_offset;
2808         u64 last_byte = i_size_read(inode);
2809         u64 block_start;
2810         u64 cur_end;
2811         sector_t sector;
2812         struct extent_map *em;
2813         struct block_device *bdev;
2814         int ret;
2815         int nr = 0;
2816         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2817         size_t pg_offset = 0;
2818         size_t iosize;
2819         size_t disk_io_size;
2820         size_t blocksize = inode->i_sb->s_blocksize;
2821         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2822
2823         set_page_extent_mapped(page);
2824
2825         end = page_end;
2826         if (!PageUptodate(page)) {
2827                 if (cleancache_get_page(page) == 0) {
2828                         BUG_ON(blocksize != PAGE_SIZE);
2829                         unlock_extent(tree, start, end);
2830                         goto out;
2831                 }
2832         }
2833
2834         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2835                 char *userpage;
2836                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2837
2838                 if (zero_offset) {
2839                         iosize = PAGE_CACHE_SIZE - zero_offset;
2840                         userpage = kmap_atomic(page);
2841                         memset(userpage + zero_offset, 0, iosize);
2842                         flush_dcache_page(page);
2843                         kunmap_atomic(userpage);
2844                 }
2845         }
2846         while (cur <= end) {
2847                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2848
2849                 if (cur >= last_byte) {
2850                         char *userpage;
2851                         struct extent_state *cached = NULL;
2852
2853                         iosize = PAGE_CACHE_SIZE - pg_offset;
2854                         userpage = kmap_atomic(page);
2855                         memset(userpage + pg_offset, 0, iosize);
2856                         flush_dcache_page(page);
2857                         kunmap_atomic(userpage);
2858                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2859                                             &cached, GFP_NOFS);
2860                         if (!parent_locked)
2861                                 unlock_extent_cached(tree, cur,
2862                                                      cur + iosize - 1,
2863                                                      &cached, GFP_NOFS);
2864                         break;
2865                 }
2866                 em = __get_extent_map(inode, page, pg_offset, cur,
2867                                       end - cur + 1, get_extent, em_cached);
2868                 if (IS_ERR_OR_NULL(em)) {
2869                         SetPageError(page);
2870                         if (!parent_locked)
2871                                 unlock_extent(tree, cur, end);
2872                         break;
2873                 }
2874                 extent_offset = cur - em->start;
2875                 BUG_ON(extent_map_end(em) <= cur);
2876                 BUG_ON(end < cur);
2877
2878                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2879                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2880                         extent_set_compress_type(&this_bio_flag,
2881                                                  em->compress_type);
2882                 }
2883
2884                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2885                 cur_end = min(extent_map_end(em) - 1, end);
2886                 iosize = ALIGN(iosize, blocksize);
2887                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2888                         disk_io_size = em->block_len;
2889                         sector = em->block_start >> 9;
2890                 } else {
2891                         sector = (em->block_start + extent_offset) >> 9;
2892                         disk_io_size = iosize;
2893                 }
2894                 bdev = em->bdev;
2895                 block_start = em->block_start;
2896                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2897                         block_start = EXTENT_MAP_HOLE;
2898                 free_extent_map(em);
2899                 em = NULL;
2900
2901                 /* we've found a hole, just zero and go on */
2902                 if (block_start == EXTENT_MAP_HOLE) {
2903                         char *userpage;
2904                         struct extent_state *cached = NULL;
2905
2906                         userpage = kmap_atomic(page);
2907                         memset(userpage + pg_offset, 0, iosize);
2908                         flush_dcache_page(page);
2909                         kunmap_atomic(userpage);
2910
2911                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2912                                             &cached, GFP_NOFS);
2913                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2914                                              &cached, GFP_NOFS);
2915                         cur = cur + iosize;
2916                         pg_offset += iosize;
2917                         continue;
2918                 }
2919                 /* the get_extent function already copied into the page */
2920                 if (test_range_bit(tree, cur, cur_end,
2921                                    EXTENT_UPTODATE, 1, NULL)) {
2922                         check_page_uptodate(tree, page);
2923                         if (!parent_locked)
2924                                 unlock_extent(tree, cur, cur + iosize - 1);
2925                         cur = cur + iosize;
2926                         pg_offset += iosize;
2927                         continue;
2928                 }
2929                 /* we have an inline extent but it didn't get marked up
2930                  * to date.  Error out
2931                  */
2932                 if (block_start == EXTENT_MAP_INLINE) {
2933                         SetPageError(page);
2934                         if (!parent_locked)
2935                                 unlock_extent(tree, cur, cur + iosize - 1);
2936                         cur = cur + iosize;
2937                         pg_offset += iosize;
2938                         continue;
2939                 }
2940
2941                 pnr -= page->index;
2942                 ret = submit_extent_page(rw, tree, page,
2943                                          sector, disk_io_size, pg_offset,
2944                                          bdev, bio, pnr,
2945                                          end_bio_extent_readpage, mirror_num,
2946                                          *bio_flags,
2947                                          this_bio_flag);
2948                 if (!ret) {
2949                         nr++;
2950                         *bio_flags = this_bio_flag;
2951                 } else {
2952                         SetPageError(page);
2953                         if (!parent_locked)
2954                                 unlock_extent(tree, cur, cur + iosize - 1);
2955                 }
2956                 cur = cur + iosize;
2957                 pg_offset += iosize;
2958         }
2959 out:
2960         if (!nr) {
2961                 if (!PageError(page))
2962                         SetPageUptodate(page);
2963                 unlock_page(page);
2964         }
2965         return 0;
2966 }
2967
2968 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
2969                                              struct page *pages[], int nr_pages,
2970                                              u64 start, u64 end,
2971                                              get_extent_t *get_extent,
2972                                              struct extent_map **em_cached,
2973                                              struct bio **bio, int mirror_num,
2974                                              unsigned long *bio_flags, int rw)
2975 {
2976         struct inode *inode;
2977         struct btrfs_ordered_extent *ordered;
2978         int index;
2979
2980         inode = pages[0]->mapping->host;
2981         while (1) {
2982                 lock_extent(tree, start, end);
2983                 ordered = btrfs_lookup_ordered_range(inode, start,
2984                                                      end - start + 1);
2985                 if (!ordered)
2986                         break;
2987                 unlock_extent(tree, start, end);
2988                 btrfs_start_ordered_extent(inode, ordered, 1);
2989                 btrfs_put_ordered_extent(ordered);
2990         }
2991
2992         for (index = 0; index < nr_pages; index++) {
2993                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
2994                               mirror_num, bio_flags, rw);
2995                 page_cache_release(pages[index]);
2996         }
2997 }
2998
2999 static void __extent_readpages(struct extent_io_tree *tree,
3000                                struct page *pages[],
3001                                int nr_pages, get_extent_t *get_extent,
3002                                struct extent_map **em_cached,
3003                                struct bio **bio, int mirror_num,
3004                                unsigned long *bio_flags, int rw)
3005 {
3006         u64 start = 0;
3007         u64 end = 0;
3008         u64 page_start;
3009         int index;
3010         int first_index = 0;
3011
3012         for (index = 0; index < nr_pages; index++) {
3013                 page_start = page_offset(pages[index]);
3014                 if (!end) {
3015                         start = page_start;
3016                         end = start + PAGE_CACHE_SIZE - 1;
3017                         first_index = index;
3018                 } else if (end + 1 == page_start) {
3019                         end += PAGE_CACHE_SIZE;
3020                 } else {
3021                         __do_contiguous_readpages(tree, &pages[first_index],
3022                                                   index - first_index, start,
3023                                                   end, get_extent, em_cached,
3024                                                   bio, mirror_num, bio_flags,
3025                                                   rw);
3026                         start = page_start;
3027                         end = start + PAGE_CACHE_SIZE - 1;
3028                         first_index = index;
3029                 }
3030         }
3031
3032         if (end)
3033                 __do_contiguous_readpages(tree, &pages[first_index],
3034                                           index - first_index, start,
3035                                           end, get_extent, em_cached, bio,
3036                                           mirror_num, bio_flags, rw);
3037 }
3038
3039 static int __extent_read_full_page(struct extent_io_tree *tree,
3040                                    struct page *page,
3041                                    get_extent_t *get_extent,
3042                                    struct bio **bio, int mirror_num,
3043                                    unsigned long *bio_flags, int rw)
3044 {
3045         struct inode *inode = page->mapping->host;
3046         struct btrfs_ordered_extent *ordered;
3047         u64 start = page_offset(page);
3048         u64 end = start + PAGE_CACHE_SIZE - 1;
3049         int ret;
3050
3051         while (1) {
3052                 lock_extent(tree, start, end);
3053                 ordered = btrfs_lookup_ordered_extent(inode, start);
3054                 if (!ordered)
3055                         break;
3056                 unlock_extent(tree, start, end);
3057                 btrfs_start_ordered_extent(inode, ordered, 1);
3058                 btrfs_put_ordered_extent(ordered);
3059         }
3060
3061         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3062                             bio_flags, rw);
3063         return ret;
3064 }
3065
3066 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3067                             get_extent_t *get_extent, int mirror_num)
3068 {
3069         struct bio *bio = NULL;
3070         unsigned long bio_flags = 0;
3071         int ret;
3072
3073         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3074                                       &bio_flags, READ);
3075         if (bio)
3076                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3077         return ret;
3078 }
3079
3080 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3081                                  get_extent_t *get_extent, int mirror_num)
3082 {
3083         struct bio *bio = NULL;
3084         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3085         int ret;
3086
3087         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3088                                       &bio_flags, READ);
3089         if (bio)
3090                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3091         return ret;
3092 }
3093
3094 static noinline void update_nr_written(struct page *page,
3095                                       struct writeback_control *wbc,
3096                                       unsigned long nr_written)
3097 {
3098         wbc->nr_to_write -= nr_written;
3099         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3100             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3101                 page->mapping->writeback_index = page->index + nr_written;
3102 }
3103
3104 /*
3105  * helper for __extent_writepage, doing all of the delayed allocation setup.
3106  *
3107  * This returns 1 if our fill_delalloc function did all the work required
3108  * to write the page (copy into inline extent).  In this case the IO has
3109  * been started and the page is already unlocked.
3110  *
3111  * This returns 0 if all went well (page still locked)
3112  * This returns < 0 if there were errors (page still locked)
3113  */
3114 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3115                               struct page *page, struct writeback_control *wbc,
3116                               struct extent_page_data *epd,
3117                               u64 delalloc_start,
3118                               unsigned long *nr_written)
3119 {
3120         struct extent_io_tree *tree = epd->tree;
3121         u64 page_end = delalloc_start + PAGE_CACHE_SIZE - 1;
3122         u64 nr_delalloc;
3123         u64 delalloc_to_write = 0;
3124         u64 delalloc_end = 0;
3125         int ret;
3126         int page_started = 0;
3127
3128         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3129                 return 0;
3130
3131         while (delalloc_end < page_end) {
3132                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3133                                                page,
3134                                                &delalloc_start,
3135                                                &delalloc_end,
3136                                                128 * 1024 * 1024);
3137                 if (nr_delalloc == 0) {
3138                         delalloc_start = delalloc_end + 1;
3139                         continue;
3140                 }
3141                 ret = tree->ops->fill_delalloc(inode, page,
3142                                                delalloc_start,
3143                                                delalloc_end,
3144                                                &page_started,
3145                                                nr_written);
3146                 /* File system has been set read-only */
3147                 if (ret) {
3148                         SetPageError(page);
3149                         /* fill_delalloc should be return < 0 for error
3150                          * but just in case, we use > 0 here meaning the
3151                          * IO is started, so we don't want to return > 0
3152                          * unless things are going well.
3153                          */
3154                         ret = ret < 0 ? ret : -EIO;
3155                         goto done;
3156                 }
3157                 /*
3158                  * delalloc_end is already one less than the total
3159                  * length, so we don't subtract one from
3160                  * PAGE_CACHE_SIZE
3161                  */
3162                 delalloc_to_write += (delalloc_end - delalloc_start +
3163                                       PAGE_CACHE_SIZE) >>
3164                                       PAGE_CACHE_SHIFT;
3165                 delalloc_start = delalloc_end + 1;
3166         }
3167         if (wbc->nr_to_write < delalloc_to_write) {
3168                 int thresh = 8192;
3169
3170                 if (delalloc_to_write < thresh * 2)
3171                         thresh = delalloc_to_write;
3172                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3173                                          thresh);
3174         }
3175
3176         /* did the fill delalloc function already unlock and start
3177          * the IO?
3178          */
3179         if (page_started) {
3180                 /*
3181                  * we've unlocked the page, so we can't update
3182                  * the mapping's writeback index, just update
3183                  * nr_to_write.
3184                  */
3185                 wbc->nr_to_write -= *nr_written;
3186                 return 1;
3187         }
3188
3189         ret = 0;
3190
3191 done:
3192         return ret;
3193 }
3194
3195 /*
3196  * helper for __extent_writepage.  This calls the writepage start hooks,
3197  * and does the loop to map the page into extents and bios.
3198  *
3199  * We return 1 if the IO is started and the page is unlocked,
3200  * 0 if all went well (page still locked)
3201  * < 0 if there were errors (page still locked)
3202  */
3203 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3204                                  struct page *page,
3205                                  struct writeback_control *wbc,
3206                                  struct extent_page_data *epd,
3207                                  loff_t i_size,
3208                                  unsigned long nr_written,
3209                                  int write_flags, int *nr_ret)
3210 {
3211         struct extent_io_tree *tree = epd->tree;
3212         u64 start = page_offset(page);
3213         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3214         u64 end;
3215         u64 cur = start;
3216         u64 extent_offset;
3217         u64 block_start;
3218         u64 iosize;
3219         sector_t sector;
3220         struct extent_state *cached_state = NULL;
3221         struct extent_map *em;
3222         struct block_device *bdev;
3223         size_t pg_offset = 0;
3224         size_t blocksize;
3225         int ret = 0;
3226         int nr = 0;
3227         bool compressed;
3228
3229         if (tree->ops && tree->ops->writepage_start_hook) {
3230                 ret = tree->ops->writepage_start_hook(page, start,
3231                                                       page_end);
3232                 if (ret) {
3233                         /* Fixup worker will requeue */
3234                         if (ret == -EBUSY)
3235                                 wbc->pages_skipped++;
3236                         else
3237                                 redirty_page_for_writepage(wbc, page);
3238
3239                         update_nr_written(page, wbc, nr_written);
3240                         unlock_page(page);
3241                         ret = 1;
3242                         goto done_unlocked;
3243                 }
3244         }
3245
3246         /*
3247          * we don't want to touch the inode after unlocking the page,
3248          * so we update the mapping writeback index now
3249          */
3250         update_nr_written(page, wbc, nr_written + 1);
3251
3252         end = page_end;
3253         if (i_size <= start) {
3254                 if (tree->ops && tree->ops->writepage_end_io_hook)
3255                         tree->ops->writepage_end_io_hook(page, start,
3256                                                          page_end, NULL, 1);
3257                 goto done;
3258         }
3259
3260         blocksize = inode->i_sb->s_blocksize;
3261
3262         while (cur <= end) {
3263                 u64 em_end;
3264                 if (cur >= i_size) {
3265                         if (tree->ops && tree->ops->writepage_end_io_hook)
3266                                 tree->ops->writepage_end_io_hook(page, cur,
3267                                                          page_end, NULL, 1);
3268                         break;
3269                 }
3270                 em = epd->get_extent(inode, page, pg_offset, cur,
3271                                      end - cur + 1, 1);
3272                 if (IS_ERR_OR_NULL(em)) {
3273                         SetPageError(page);
3274                         ret = PTR_ERR_OR_ZERO(em);
3275                         break;
3276                 }
3277
3278                 extent_offset = cur - em->start;
3279                 em_end = extent_map_end(em);
3280                 BUG_ON(em_end <= cur);
3281                 BUG_ON(end < cur);
3282                 iosize = min(em_end - cur, end - cur + 1);
3283                 iosize = ALIGN(iosize, blocksize);
3284                 sector = (em->block_start + extent_offset) >> 9;
3285                 bdev = em->bdev;
3286                 block_start = em->block_start;
3287                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3288                 free_extent_map(em);
3289                 em = NULL;
3290
3291                 /*
3292                  * compressed and inline extents are written through other
3293                  * paths in the FS
3294                  */
3295                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3296                     block_start == EXTENT_MAP_INLINE) {
3297                         /*
3298                          * end_io notification does not happen here for
3299                          * compressed extents
3300                          */
3301                         if (!compressed && tree->ops &&
3302                             tree->ops->writepage_end_io_hook)
3303                                 tree->ops->writepage_end_io_hook(page, cur,
3304                                                          cur + iosize - 1,
3305                                                          NULL, 1);
3306                         else if (compressed) {
3307                                 /* we don't want to end_page_writeback on
3308                                  * a compressed extent.  this happens
3309                                  * elsewhere
3310                                  */
3311                                 nr++;
3312                         }
3313
3314                         cur += iosize;
3315                         pg_offset += iosize;
3316                         continue;
3317                 }
3318
3319                 if (tree->ops && tree->ops->writepage_io_hook) {
3320                         ret = tree->ops->writepage_io_hook(page, cur,
3321                                                 cur + iosize - 1);
3322                 } else {
3323                         ret = 0;
3324                 }
3325                 if (ret) {
3326                         SetPageError(page);
3327                 } else {
3328                         unsigned long max_nr = (i_size >> PAGE_CACHE_SHIFT) + 1;
3329
3330                         set_range_writeback(tree, cur, cur + iosize - 1);
3331                         if (!PageWriteback(page)) {
3332                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
3333                                            "page %lu not writeback, cur %llu end %llu",
3334                                        page->index, cur, end);
3335                         }
3336
3337                         ret = submit_extent_page(write_flags, tree, page,
3338                                                  sector, iosize, pg_offset,
3339                                                  bdev, &epd->bio, max_nr,
3340                                                  end_bio_extent_writepage,
3341                                                  0, 0, 0);
3342                         if (ret)
3343                                 SetPageError(page);
3344                 }
3345                 cur = cur + iosize;
3346                 pg_offset += iosize;
3347                 nr++;
3348         }
3349 done:
3350         *nr_ret = nr;
3351
3352 done_unlocked:
3353
3354         /* drop our reference on any cached states */
3355         free_extent_state(cached_state);
3356         return ret;
3357 }
3358
3359 /*
3360  * the writepage semantics are similar to regular writepage.  extent
3361  * records are inserted to lock ranges in the tree, and as dirty areas
3362  * are found, they are marked writeback.  Then the lock bits are removed
3363  * and the end_io handler clears the writeback ranges
3364  */
3365 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3366                               void *data)
3367 {
3368         struct inode *inode = page->mapping->host;
3369         struct extent_page_data *epd = data;
3370         u64 start = page_offset(page);
3371         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3372         int ret;
3373         int nr = 0;
3374         size_t pg_offset = 0;
3375         loff_t i_size = i_size_read(inode);
3376         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3377         int write_flags;
3378         unsigned long nr_written = 0;
3379
3380         if (wbc->sync_mode == WB_SYNC_ALL)
3381                 write_flags = WRITE_SYNC;
3382         else
3383                 write_flags = WRITE;
3384
3385         trace___extent_writepage(page, inode, wbc);
3386
3387         WARN_ON(!PageLocked(page));
3388
3389         ClearPageError(page);
3390
3391         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3392         if (page->index > end_index ||
3393            (page->index == end_index && !pg_offset)) {
3394                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3395                 unlock_page(page);
3396                 return 0;
3397         }
3398
3399         if (page->index == end_index) {
3400                 char *userpage;
3401
3402                 userpage = kmap_atomic(page);
3403                 memset(userpage + pg_offset, 0,
3404                        PAGE_CACHE_SIZE - pg_offset);
3405                 kunmap_atomic(userpage);
3406                 flush_dcache_page(page);
3407         }
3408
3409         pg_offset = 0;
3410
3411         set_page_extent_mapped(page);
3412
3413         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3414         if (ret == 1)
3415                 goto done_unlocked;
3416         if (ret)
3417                 goto done;
3418
3419         ret = __extent_writepage_io(inode, page, wbc, epd,
3420                                     i_size, nr_written, write_flags, &nr);
3421         if (ret == 1)
3422                 goto done_unlocked;
3423
3424 done:
3425         if (nr == 0) {
3426                 /* make sure the mapping tag for page dirty gets cleared */
3427                 set_page_writeback(page);
3428                 end_page_writeback(page);
3429         }
3430         if (PageError(page)) {
3431                 ret = ret < 0 ? ret : -EIO;
3432                 end_extent_writepage(page, ret, start, page_end);
3433         }
3434         unlock_page(page);
3435         return ret;
3436
3437 done_unlocked:
3438         return 0;
3439 }
3440
3441 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3442 {
3443         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3444                        TASK_UNINTERRUPTIBLE);
3445 }
3446
3447 static noinline_for_stack int
3448 lock_extent_buffer_for_io(struct extent_buffer *eb,
3449                           struct btrfs_fs_info *fs_info,
3450                           struct extent_page_data *epd)
3451 {
3452         unsigned long i, num_pages;
3453         int flush = 0;
3454         int ret = 0;
3455
3456         if (!btrfs_try_tree_write_lock(eb)) {
3457                 flush = 1;
3458                 flush_write_bio(epd);
3459                 btrfs_tree_lock(eb);
3460         }
3461
3462         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3463                 btrfs_tree_unlock(eb);
3464                 if (!epd->sync_io)
3465                         return 0;
3466                 if (!flush) {
3467                         flush_write_bio(epd);
3468                         flush = 1;
3469                 }
3470                 while (1) {
3471                         wait_on_extent_buffer_writeback(eb);
3472                         btrfs_tree_lock(eb);
3473                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3474                                 break;
3475                         btrfs_tree_unlock(eb);
3476                 }
3477         }
3478
3479         /*
3480          * We need to do this to prevent races in people who check if the eb is
3481          * under IO since we can end up having no IO bits set for a short period
3482          * of time.
3483          */
3484         spin_lock(&eb->refs_lock);
3485         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3486                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3487                 spin_unlock(&eb->refs_lock);
3488                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3489                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3490                                      -eb->len,
3491                                      fs_info->dirty_metadata_batch);
3492                 ret = 1;
3493         } else {
3494                 spin_unlock(&eb->refs_lock);
3495         }
3496
3497         btrfs_tree_unlock(eb);
3498
3499         if (!ret)
3500                 return ret;
3501
3502         num_pages = num_extent_pages(eb->start, eb->len);
3503         for (i = 0; i < num_pages; i++) {
3504                 struct page *p = extent_buffer_page(eb, i);
3505
3506                 if (!trylock_page(p)) {
3507                         if (!flush) {
3508                                 flush_write_bio(epd);
3509                                 flush = 1;
3510                         }
3511                         lock_page(p);
3512                 }
3513         }
3514
3515         return ret;
3516 }
3517
3518 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3519 {
3520         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3521         smp_mb__after_atomic();
3522         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3523 }
3524
3525 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3526 {
3527         struct bio_vec *bvec;
3528         struct extent_buffer *eb;
3529         int i, done;
3530
3531         bio_for_each_segment_all(bvec, bio, i) {
3532                 struct page *page = bvec->bv_page;
3533
3534                 eb = (struct extent_buffer *)page->private;
3535                 BUG_ON(!eb);
3536                 done = atomic_dec_and_test(&eb->io_pages);
3537
3538                 if (err || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3539                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3540                         ClearPageUptodate(page);
3541                         SetPageError(page);
3542                 }
3543
3544                 end_page_writeback(page);
3545
3546                 if (!done)
3547                         continue;
3548
3549                 end_extent_buffer_writeback(eb);
3550         }
3551
3552         bio_put(bio);
3553 }
3554
3555 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3556                         struct btrfs_fs_info *fs_info,
3557                         struct writeback_control *wbc,
3558                         struct extent_page_data *epd)
3559 {
3560         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3561         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3562         u64 offset = eb->start;
3563         unsigned long i, num_pages;
3564         unsigned long bio_flags = 0;
3565         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3566         int ret = 0;
3567
3568         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3569         num_pages = num_extent_pages(eb->start, eb->len);
3570         atomic_set(&eb->io_pages, num_pages);
3571         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3572                 bio_flags = EXTENT_BIO_TREE_LOG;
3573
3574         for (i = 0; i < num_pages; i++) {
3575                 struct page *p = extent_buffer_page(eb, i);
3576
3577                 clear_page_dirty_for_io(p);
3578                 set_page_writeback(p);
3579                 ret = submit_extent_page(rw, tree, p, offset >> 9,
3580                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3581                                          -1, end_bio_extent_buffer_writepage,
3582                                          0, epd->bio_flags, bio_flags);
3583                 epd->bio_flags = bio_flags;
3584                 if (ret) {
3585                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3586                         SetPageError(p);
3587                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3588                                 end_extent_buffer_writeback(eb);
3589                         ret = -EIO;
3590                         break;
3591                 }
3592                 offset += PAGE_CACHE_SIZE;
3593                 update_nr_written(p, wbc, 1);
3594                 unlock_page(p);
3595         }
3596
3597         if (unlikely(ret)) {
3598                 for (; i < num_pages; i++) {
3599                         struct page *p = extent_buffer_page(eb, i);
3600                         unlock_page(p);
3601                 }
3602         }
3603
3604         return ret;
3605 }
3606
3607 int btree_write_cache_pages(struct address_space *mapping,
3608                                    struct writeback_control *wbc)
3609 {
3610         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3611         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3612         struct extent_buffer *eb, *prev_eb = NULL;
3613         struct extent_page_data epd = {
3614                 .bio = NULL,
3615                 .tree = tree,
3616                 .extent_locked = 0,
3617                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3618                 .bio_flags = 0,
3619         };
3620         int ret = 0;
3621         int done = 0;
3622         int nr_to_write_done = 0;
3623         struct pagevec pvec;
3624         int nr_pages;
3625         pgoff_t index;
3626         pgoff_t end;            /* Inclusive */
3627         int scanned = 0;
3628         int tag;
3629
3630         pagevec_init(&pvec, 0);
3631         if (wbc->range_cyclic) {
3632                 index = mapping->writeback_index; /* Start from prev offset */
3633                 end = -1;
3634         } else {
3635                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3636                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3637                 scanned = 1;
3638         }
3639         if (wbc->sync_mode == WB_SYNC_ALL)
3640                 tag = PAGECACHE_TAG_TOWRITE;
3641         else
3642                 tag = PAGECACHE_TAG_DIRTY;
3643 retry:
3644         if (wbc->sync_mode == WB_SYNC_ALL)
3645                 tag_pages_for_writeback(mapping, index, end);
3646         while (!done && !nr_to_write_done && (index <= end) &&
3647                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3648                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3649                 unsigned i;
3650
3651                 scanned = 1;
3652                 for (i = 0; i < nr_pages; i++) {
3653                         struct page *page = pvec.pages[i];
3654
3655                         if (!PagePrivate(page))
3656                                 continue;
3657
3658                         if (!wbc->range_cyclic && page->index > end) {
3659                                 done = 1;
3660                                 break;
3661                         }
3662
3663                         spin_lock(&mapping->private_lock);
3664                         if (!PagePrivate(page)) {
3665                                 spin_unlock(&mapping->private_lock);
3666                                 continue;
3667                         }
3668
3669                         eb = (struct extent_buffer *)page->private;
3670
3671                         /*
3672                          * Shouldn't happen and normally this would be a BUG_ON
3673                          * but no sense in crashing the users box for something
3674                          * we can survive anyway.
3675                          */
3676                         if (WARN_ON(!eb)) {
3677                                 spin_unlock(&mapping->private_lock);
3678                                 continue;
3679                         }
3680
3681                         if (eb == prev_eb) {
3682                                 spin_unlock(&mapping->private_lock);
3683                                 continue;
3684                         }
3685
3686                         ret = atomic_inc_not_zero(&eb->refs);
3687                         spin_unlock(&mapping->private_lock);
3688                         if (!ret)
3689                                 continue;
3690
3691                         prev_eb = eb;
3692                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3693                         if (!ret) {
3694                                 free_extent_buffer(eb);
3695                                 continue;
3696                         }
3697
3698                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3699                         if (ret) {
3700                                 done = 1;
3701                                 free_extent_buffer(eb);
3702                                 break;
3703                         }
3704                         free_extent_buffer(eb);
3705
3706                         /*
3707                          * the filesystem may choose to bump up nr_to_write.
3708                          * We have to make sure to honor the new nr_to_write
3709                          * at any time
3710                          */
3711                         nr_to_write_done = wbc->nr_to_write <= 0;
3712                 }
3713                 pagevec_release(&pvec);
3714                 cond_resched();
3715         }
3716         if (!scanned && !done) {
3717                 /*
3718                  * We hit the last page and there is more work to be done: wrap
3719                  * back to the start of the file
3720                  */
3721                 scanned = 1;
3722                 index = 0;
3723                 goto retry;
3724         }
3725         flush_write_bio(&epd);
3726         return ret;
3727 }
3728
3729 /**
3730  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3731  * @mapping: address space structure to write
3732  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3733  * @writepage: function called for each page
3734  * @data: data passed to writepage function
3735  *
3736  * If a page is already under I/O, write_cache_pages() skips it, even
3737  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3738  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3739  * and msync() need to guarantee that all the data which was dirty at the time
3740  * the call was made get new I/O started against them.  If wbc->sync_mode is
3741  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3742  * existing IO to complete.
3743  */
3744 static int extent_write_cache_pages(struct extent_io_tree *tree,
3745                              struct address_space *mapping,
3746                              struct writeback_control *wbc,
3747                              writepage_t writepage, void *data,
3748                              void (*flush_fn)(void *))
3749 {
3750         struct inode *inode = mapping->host;
3751         int ret = 0;
3752         int done = 0;
3753         int err = 0;
3754         int nr_to_write_done = 0;
3755         struct pagevec pvec;
3756         int nr_pages;
3757         pgoff_t index;
3758         pgoff_t end;            /* Inclusive */
3759         int scanned = 0;
3760         int tag;
3761
3762         /*
3763          * We have to hold onto the inode so that ordered extents can do their
3764          * work when the IO finishes.  The alternative to this is failing to add
3765          * an ordered extent if the igrab() fails there and that is a huge pain
3766          * to deal with, so instead just hold onto the inode throughout the
3767          * writepages operation.  If it fails here we are freeing up the inode
3768          * anyway and we'd rather not waste our time writing out stuff that is
3769          * going to be truncated anyway.
3770          */
3771         if (!igrab(inode))
3772                 return 0;
3773
3774         pagevec_init(&pvec, 0);
3775         if (wbc->range_cyclic) {
3776                 index = mapping->writeback_index; /* Start from prev offset */
3777                 end = -1;
3778         } else {
3779                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3780                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3781                 scanned = 1;
3782         }
3783         if (wbc->sync_mode == WB_SYNC_ALL)
3784                 tag = PAGECACHE_TAG_TOWRITE;
3785         else
3786                 tag = PAGECACHE_TAG_DIRTY;
3787 retry:
3788         if (wbc->sync_mode == WB_SYNC_ALL)
3789                 tag_pages_for_writeback(mapping, index, end);
3790         while (!done && !nr_to_write_done && (index <= end) &&
3791                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3792                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3793                 unsigned i;
3794
3795                 scanned = 1;
3796                 for (i = 0; i < nr_pages; i++) {
3797                         struct page *page = pvec.pages[i];
3798
3799                         /*
3800                          * At this point we hold neither mapping->tree_lock nor
3801                          * lock on the page itself: the page may be truncated or
3802                          * invalidated (changing page->mapping to NULL), or even
3803                          * swizzled back from swapper_space to tmpfs file
3804                          * mapping
3805                          */
3806                         if (!trylock_page(page)) {
3807                                 flush_fn(data);
3808                                 lock_page(page);
3809                         }
3810
3811                         if (unlikely(page->mapping != mapping)) {
3812                                 unlock_page(page);
3813                                 continue;
3814                         }
3815
3816                         if (!wbc->range_cyclic && page->index > end) {
3817                                 done = 1;
3818                                 unlock_page(page);
3819                                 continue;
3820                         }
3821
3822                         if (wbc->sync_mode != WB_SYNC_NONE) {
3823                                 if (PageWriteback(page))
3824                                         flush_fn(data);
3825                                 wait_on_page_writeback(page);
3826                         }
3827
3828                         if (PageWriteback(page) ||
3829                             !clear_page_dirty_for_io(page)) {
3830                                 unlock_page(page);
3831                                 continue;
3832                         }
3833
3834                         ret = (*writepage)(page, wbc, data);
3835
3836                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3837                                 unlock_page(page);
3838                                 ret = 0;
3839                         }
3840                         if (!err && ret < 0)
3841                                 err = ret;
3842
3843                         /*
3844                          * the filesystem may choose to bump up nr_to_write.
3845                          * We have to make sure to honor the new nr_to_write
3846                          * at any time
3847                          */
3848                         nr_to_write_done = wbc->nr_to_write <= 0;
3849                 }
3850                 pagevec_release(&pvec);
3851                 cond_resched();
3852         }
3853         if (!scanned && !done && !err) {
3854                 /*
3855                  * We hit the last page and there is more work to be done: wrap
3856                  * back to the start of the file
3857                  */
3858                 scanned = 1;
3859                 index = 0;
3860                 goto retry;
3861         }
3862         btrfs_add_delayed_iput(inode);
3863         return err;
3864 }
3865
3866 static void flush_epd_write_bio(struct extent_page_data *epd)
3867 {
3868         if (epd->bio) {
3869                 int rw = WRITE;
3870                 int ret;
3871
3872                 if (epd->sync_io)
3873                         rw = WRITE_SYNC;
3874
3875                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
3876                 BUG_ON(ret < 0); /* -ENOMEM */
3877                 epd->bio = NULL;
3878         }
3879 }
3880
3881 static noinline void flush_write_bio(void *data)
3882 {
3883         struct extent_page_data *epd = data;
3884         flush_epd_write_bio(epd);
3885 }
3886
3887 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3888                           get_extent_t *get_extent,
3889                           struct writeback_control *wbc)
3890 {
3891         int ret;
3892         struct extent_page_data epd = {
3893                 .bio = NULL,
3894                 .tree = tree,
3895                 .get_extent = get_extent,
3896                 .extent_locked = 0,
3897                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3898                 .bio_flags = 0,
3899         };
3900
3901         ret = __extent_writepage(page, wbc, &epd);
3902
3903         flush_epd_write_bio(&epd);
3904         return ret;
3905 }
3906
3907 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3908                               u64 start, u64 end, get_extent_t *get_extent,
3909                               int mode)
3910 {
3911         int ret = 0;
3912         struct address_space *mapping = inode->i_mapping;
3913         struct page *page;
3914         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3915                 PAGE_CACHE_SHIFT;
3916
3917         struct extent_page_data epd = {
3918                 .bio = NULL,
3919                 .tree = tree,
3920                 .get_extent = get_extent,
3921                 .extent_locked = 1,
3922                 .sync_io = mode == WB_SYNC_ALL,
3923                 .bio_flags = 0,
3924         };
3925         struct writeback_control wbc_writepages = {
3926                 .sync_mode      = mode,
3927                 .nr_to_write    = nr_pages * 2,
3928                 .range_start    = start,
3929                 .range_end      = end + 1,
3930         };
3931
3932         while (start <= end) {
3933                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3934                 if (clear_page_dirty_for_io(page))
3935                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3936                 else {
3937                         if (tree->ops && tree->ops->writepage_end_io_hook)
3938                                 tree->ops->writepage_end_io_hook(page, start,
3939                                                  start + PAGE_CACHE_SIZE - 1,
3940                                                  NULL, 1);
3941                         unlock_page(page);
3942                 }
3943                 page_cache_release(page);
3944                 start += PAGE_CACHE_SIZE;
3945         }
3946
3947         flush_epd_write_bio(&epd);
3948         return ret;
3949 }
3950
3951 int extent_writepages(struct extent_io_tree *tree,
3952                       struct address_space *mapping,
3953                       get_extent_t *get_extent,
3954                       struct writeback_control *wbc)
3955 {
3956         int ret = 0;
3957         struct extent_page_data epd = {
3958                 .bio = NULL,
3959                 .tree = tree,
3960                 .get_extent = get_extent,
3961                 .extent_locked = 0,
3962                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3963                 .bio_flags = 0,
3964         };
3965
3966         ret = extent_write_cache_pages(tree, mapping, wbc,
3967                                        __extent_writepage, &epd,
3968                                        flush_write_bio);
3969         flush_epd_write_bio(&epd);
3970         return ret;
3971 }
3972
3973 int extent_readpages(struct extent_io_tree *tree,
3974                      struct address_space *mapping,
3975                      struct list_head *pages, unsigned nr_pages,
3976                      get_extent_t get_extent)
3977 {
3978         struct bio *bio = NULL;
3979         unsigned page_idx;
3980         unsigned long bio_flags = 0;
3981         struct page *pagepool[16];
3982         struct page *page;
3983         struct extent_map *em_cached = NULL;
3984         int nr = 0;
3985
3986         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3987                 page = list_entry(pages->prev, struct page, lru);
3988
3989                 prefetchw(&page->flags);
3990                 list_del(&page->lru);
3991                 if (add_to_page_cache_lru(page, mapping,
3992                                         page->index, GFP_NOFS)) {
3993                         page_cache_release(page);
3994                         continue;
3995                 }
3996
3997                 pagepool[nr++] = page;
3998                 if (nr < ARRAY_SIZE(pagepool))
3999                         continue;
4000                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4001                                    &bio, 0, &bio_flags, READ);
4002                 nr = 0;
4003         }
4004         if (nr)
4005                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4006                                    &bio, 0, &bio_flags, READ);
4007
4008         if (em_cached)
4009                 free_extent_map(em_cached);
4010
4011         BUG_ON(!list_empty(pages));
4012         if (bio)
4013                 return submit_one_bio(READ, bio, 0, bio_flags);
4014         return 0;
4015 }
4016
4017 /*
4018  * basic invalidatepage code, this waits on any locked or writeback
4019  * ranges corresponding to the page, and then deletes any extent state
4020  * records from the tree
4021  */
4022 int extent_invalidatepage(struct extent_io_tree *tree,
4023                           struct page *page, unsigned long offset)
4024 {
4025         struct extent_state *cached_state = NULL;
4026         u64 start = page_offset(page);
4027         u64 end = start + PAGE_CACHE_SIZE - 1;
4028         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4029
4030         start += ALIGN(offset, blocksize);
4031         if (start > end)
4032                 return 0;
4033
4034         lock_extent_bits(tree, start, end, 0, &cached_state);
4035         wait_on_page_writeback(page);
4036         clear_extent_bit(tree, start, end,
4037                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4038                          EXTENT_DO_ACCOUNTING,
4039                          1, 1, &cached_state, GFP_NOFS);
4040         return 0;
4041 }
4042
4043 /*
4044  * a helper for releasepage, this tests for areas of the page that
4045  * are locked or under IO and drops the related state bits if it is safe
4046  * to drop the page.
4047  */
4048 static int try_release_extent_state(struct extent_map_tree *map,
4049                                     struct extent_io_tree *tree,
4050                                     struct page *page, gfp_t mask)
4051 {
4052         u64 start = page_offset(page);
4053         u64 end = start + PAGE_CACHE_SIZE - 1;
4054         int ret = 1;
4055
4056         if (test_range_bit(tree, start, end,
4057                            EXTENT_IOBITS, 0, NULL))
4058                 ret = 0;
4059         else {
4060                 if ((mask & GFP_NOFS) == GFP_NOFS)
4061                         mask = GFP_NOFS;
4062                 /*
4063                  * at this point we can safely clear everything except the
4064                  * locked bit and the nodatasum bit
4065                  */
4066                 ret = clear_extent_bit(tree, start, end,
4067                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4068                                  0, 0, NULL, mask);
4069
4070                 /* if clear_extent_bit failed for enomem reasons,
4071                  * we can't allow the release to continue.
4072                  */
4073                 if (ret < 0)
4074                         ret = 0;
4075                 else
4076                         ret = 1;
4077         }
4078         return ret;
4079 }
4080
4081 /*
4082  * a helper for releasepage.  As long as there are no locked extents
4083  * in the range corresponding to the page, both state records and extent
4084  * map records are removed
4085  */
4086 int try_release_extent_mapping(struct extent_map_tree *map,
4087                                struct extent_io_tree *tree, struct page *page,
4088                                gfp_t mask)
4089 {
4090         struct extent_map *em;
4091         u64 start = page_offset(page);
4092         u64 end = start + PAGE_CACHE_SIZE - 1;
4093
4094         if ((mask & __GFP_WAIT) &&
4095             page->mapping->host->i_size > 16 * 1024 * 1024) {
4096                 u64 len;
4097                 while (start <= end) {
4098                         len = end - start + 1;
4099                         write_lock(&map->lock);
4100                         em = lookup_extent_mapping(map, start, len);
4101                         if (!em) {
4102                                 write_unlock(&map->lock);
4103                                 break;
4104                         }
4105                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4106                             em->start != start) {
4107                                 write_unlock(&map->lock);
4108                                 free_extent_map(em);
4109                                 break;
4110                         }
4111                         if (!test_range_bit(tree, em->start,
4112                                             extent_map_end(em) - 1,
4113                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4114                                             0, NULL)) {
4115                                 remove_extent_mapping(map, em);
4116                                 /* once for the rb tree */
4117                                 free_extent_map(em);
4118                         }
4119                         start = extent_map_end(em);
4120                         write_unlock(&map->lock);
4121
4122                         /* once for us */
4123                         free_extent_map(em);
4124                 }
4125         }
4126         return try_release_extent_state(map, tree, page, mask);
4127 }
4128
4129 /*
4130  * helper function for fiemap, which doesn't want to see any holes.
4131  * This maps until we find something past 'last'
4132  */
4133 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4134                                                 u64 offset,
4135                                                 u64 last,
4136                                                 get_extent_t *get_extent)
4137 {
4138         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4139         struct extent_map *em;
4140         u64 len;
4141
4142         if (offset >= last)
4143                 return NULL;
4144
4145         while (1) {
4146                 len = last - offset;
4147                 if (len == 0)
4148                         break;
4149                 len = ALIGN(len, sectorsize);
4150                 em = get_extent(inode, NULL, 0, offset, len, 0);
4151                 if (IS_ERR_OR_NULL(em))
4152                         return em;
4153
4154                 /* if this isn't a hole return it */
4155                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4156                     em->block_start != EXTENT_MAP_HOLE) {
4157                         return em;
4158                 }
4159
4160                 /* this is a hole, advance to the next extent */
4161                 offset = extent_map_end(em);
4162                 free_extent_map(em);
4163                 if (offset >= last)
4164                         break;
4165         }
4166         return NULL;
4167 }
4168
4169 static noinline int count_ext_ref(u64 inum, u64 offset, u64 root_id, void *ctx)
4170 {
4171         unsigned long cnt = *((unsigned long *)ctx);
4172
4173         cnt++;
4174         *((unsigned long *)ctx) = cnt;
4175
4176         /* Now we're sure that the extent is shared. */
4177         if (cnt > 1)
4178                 return 1;
4179         return 0;
4180 }
4181
4182 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4183                 __u64 start, __u64 len, get_extent_t *get_extent)
4184 {
4185         int ret = 0;
4186         u64 off = start;
4187         u64 max = start + len;
4188         u32 flags = 0;
4189         u32 found_type;
4190         u64 last;
4191         u64 last_for_get_extent = 0;
4192         u64 disko = 0;
4193         u64 isize = i_size_read(inode);
4194         struct btrfs_key found_key;
4195         struct extent_map *em = NULL;
4196         struct extent_state *cached_state = NULL;
4197         struct btrfs_path *path;
4198         int end = 0;
4199         u64 em_start = 0;
4200         u64 em_len = 0;
4201         u64 em_end = 0;
4202
4203         if (len == 0)
4204                 return -EINVAL;
4205
4206         path = btrfs_alloc_path();
4207         if (!path)
4208                 return -ENOMEM;
4209         path->leave_spinning = 1;
4210
4211         start = round_down(start, BTRFS_I(inode)->root->sectorsize);
4212         len = round_up(max, BTRFS_I(inode)->root->sectorsize) - start;
4213
4214         /*
4215          * lookup the last file extent.  We're not using i_size here
4216          * because there might be preallocation past i_size
4217          */
4218         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
4219                                        path, btrfs_ino(inode), -1, 0);
4220         if (ret < 0) {
4221                 btrfs_free_path(path);
4222                 return ret;
4223         }
4224         WARN_ON(!ret);
4225         path->slots[0]--;
4226         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4227         found_type = btrfs_key_type(&found_key);
4228
4229         /* No extents, but there might be delalloc bits */
4230         if (found_key.objectid != btrfs_ino(inode) ||
4231             found_type != BTRFS_EXTENT_DATA_KEY) {
4232                 /* have to trust i_size as the end */
4233                 last = (u64)-1;
4234                 last_for_get_extent = isize;
4235         } else {
4236                 /*
4237                  * remember the start of the last extent.  There are a
4238                  * bunch of different factors that go into the length of the
4239                  * extent, so its much less complex to remember where it started
4240                  */
4241                 last = found_key.offset;
4242                 last_for_get_extent = last + 1;
4243         }
4244         btrfs_release_path(path);
4245
4246         /*
4247          * we might have some extents allocated but more delalloc past those
4248          * extents.  so, we trust isize unless the start of the last extent is
4249          * beyond isize
4250          */
4251         if (last < isize) {
4252                 last = (u64)-1;
4253                 last_for_get_extent = isize;
4254         }
4255
4256         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4257                          &cached_state);
4258
4259         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4260                                    get_extent);
4261         if (!em)
4262                 goto out;
4263         if (IS_ERR(em)) {
4264                 ret = PTR_ERR(em);
4265                 goto out;
4266         }
4267
4268         while (!end) {
4269                 u64 offset_in_extent = 0;
4270
4271                 /* break if the extent we found is outside the range */
4272                 if (em->start >= max || extent_map_end(em) < off)
4273                         break;
4274
4275                 /*
4276                  * get_extent may return an extent that starts before our
4277                  * requested range.  We have to make sure the ranges
4278                  * we return to fiemap always move forward and don't
4279                  * overlap, so adjust the offsets here
4280                  */
4281                 em_start = max(em->start, off);
4282
4283                 /*
4284                  * record the offset from the start of the extent
4285                  * for adjusting the disk offset below.  Only do this if the
4286                  * extent isn't compressed since our in ram offset may be past
4287                  * what we have actually allocated on disk.
4288                  */
4289                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4290                         offset_in_extent = em_start - em->start;
4291                 em_end = extent_map_end(em);
4292                 em_len = em_end - em_start;
4293                 disko = 0;
4294                 flags = 0;
4295
4296                 /*
4297                  * bump off for our next call to get_extent
4298                  */
4299                 off = extent_map_end(em);
4300                 if (off >= max)
4301                         end = 1;
4302
4303                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4304                         end = 1;
4305                         flags |= FIEMAP_EXTENT_LAST;
4306                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4307                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4308                                   FIEMAP_EXTENT_NOT_ALIGNED);
4309                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4310                         flags |= (FIEMAP_EXTENT_DELALLOC |
4311                                   FIEMAP_EXTENT_UNKNOWN);
4312                 } else {
4313                         unsigned long ref_cnt = 0;
4314
4315                         disko = em->block_start + offset_in_extent;
4316
4317                         /*
4318                          * As btrfs supports shared space, this information
4319                          * can be exported to userspace tools via
4320                          * flag FIEMAP_EXTENT_SHARED.
4321                          */
4322                         ret = iterate_inodes_from_logical(
4323                                         em->block_start,
4324                                         BTRFS_I(inode)->root->fs_info,
4325                                         path, count_ext_ref, &ref_cnt);
4326                         if (ret < 0 && ret != -ENOENT)
4327                                 goto out_free;
4328
4329                         if (ref_cnt > 1)
4330                                 flags |= FIEMAP_EXTENT_SHARED;
4331                 }
4332                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4333                         flags |= FIEMAP_EXTENT_ENCODED;
4334
4335                 free_extent_map(em);
4336                 em = NULL;
4337                 if ((em_start >= last) || em_len == (u64)-1 ||
4338                    (last == (u64)-1 && isize <= em_end)) {
4339                         flags |= FIEMAP_EXTENT_LAST;
4340                         end = 1;
4341                 }
4342
4343                 /* now scan forward to see if this is really the last extent. */
4344                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4345                                            get_extent);
4346                 if (IS_ERR(em)) {
4347                         ret = PTR_ERR(em);
4348                         goto out;
4349                 }
4350                 if (!em) {
4351                         flags |= FIEMAP_EXTENT_LAST;
4352                         end = 1;
4353                 }
4354                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4355                                               em_len, flags);
4356                 if (ret)
4357                         goto out_free;
4358         }
4359 out_free:
4360         free_extent_map(em);
4361 out:
4362         btrfs_free_path(path);
4363         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4364                              &cached_state, GFP_NOFS);
4365         return ret;
4366 }
4367
4368 static void __free_extent_buffer(struct extent_buffer *eb)
4369 {
4370         btrfs_leak_debug_del(&eb->leak_list);
4371         kmem_cache_free(extent_buffer_cache, eb);
4372 }
4373
4374 int extent_buffer_under_io(struct extent_buffer *eb)
4375 {
4376         return (atomic_read(&eb->io_pages) ||
4377                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4378                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4379 }
4380
4381 /*
4382  * Helper for releasing extent buffer page.
4383  */
4384 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4385                                                 unsigned long start_idx)
4386 {
4387         unsigned long index;
4388         unsigned long num_pages;
4389         struct page *page;
4390         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4391
4392         BUG_ON(extent_buffer_under_io(eb));
4393
4394         num_pages = num_extent_pages(eb->start, eb->len);
4395         index = start_idx + num_pages;
4396         if (start_idx >= index)
4397                 return;
4398
4399         do {
4400                 index--;
4401                 page = extent_buffer_page(eb, index);
4402                 if (page && mapped) {
4403                         spin_lock(&page->mapping->private_lock);
4404                         /*
4405                          * We do this since we'll remove the pages after we've
4406                          * removed the eb from the radix tree, so we could race
4407                          * and have this page now attached to the new eb.  So
4408                          * only clear page_private if it's still connected to
4409                          * this eb.
4410                          */
4411                         if (PagePrivate(page) &&
4412                             page->private == (unsigned long)eb) {
4413                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4414                                 BUG_ON(PageDirty(page));
4415                                 BUG_ON(PageWriteback(page));
4416                                 /*
4417                                  * We need to make sure we haven't be attached
4418                                  * to a new eb.
4419                                  */
4420                                 ClearPagePrivate(page);
4421                                 set_page_private(page, 0);
4422                                 /* One for the page private */
4423                                 page_cache_release(page);
4424                         }
4425                         spin_unlock(&page->mapping->private_lock);
4426
4427                 }
4428                 if (page) {
4429                         /* One for when we alloced the page */
4430                         page_cache_release(page);
4431                 }
4432         } while (index != start_idx);
4433 }
4434
4435 /*
4436  * Helper for releasing the extent buffer.
4437  */
4438 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4439 {
4440         btrfs_release_extent_buffer_page(eb, 0);
4441         __free_extent_buffer(eb);
4442 }
4443
4444 static struct extent_buffer *
4445 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4446                       unsigned long len, gfp_t mask)
4447 {
4448         struct extent_buffer *eb = NULL;
4449
4450         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
4451         if (eb == NULL)
4452                 return NULL;
4453         eb->start = start;
4454         eb->len = len;
4455         eb->fs_info = fs_info;
4456         eb->bflags = 0;
4457         rwlock_init(&eb->lock);
4458         atomic_set(&eb->write_locks, 0);
4459         atomic_set(&eb->read_locks, 0);
4460         atomic_set(&eb->blocking_readers, 0);
4461         atomic_set(&eb->blocking_writers, 0);
4462         atomic_set(&eb->spinning_readers, 0);
4463         atomic_set(&eb->spinning_writers, 0);
4464         eb->lock_nested = 0;
4465         init_waitqueue_head(&eb->write_lock_wq);
4466         init_waitqueue_head(&eb->read_lock_wq);
4467
4468         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4469
4470         spin_lock_init(&eb->refs_lock);
4471         atomic_set(&eb->refs, 1);
4472         atomic_set(&eb->io_pages, 0);
4473
4474         /*
4475          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4476          */
4477         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4478                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4479         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4480
4481         return eb;
4482 }
4483
4484 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4485 {
4486         unsigned long i;
4487         struct page *p;
4488         struct extent_buffer *new;
4489         unsigned long num_pages = num_extent_pages(src->start, src->len);
4490
4491         new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_NOFS);
4492         if (new == NULL)
4493                 return NULL;
4494
4495         for (i = 0; i < num_pages; i++) {
4496                 p = alloc_page(GFP_NOFS);
4497                 if (!p) {
4498                         btrfs_release_extent_buffer(new);
4499                         return NULL;
4500                 }
4501                 attach_extent_buffer_page(new, p);
4502                 WARN_ON(PageDirty(p));
4503                 SetPageUptodate(p);
4504                 new->pages[i] = p;
4505         }
4506
4507         copy_extent_buffer(new, src, 0, 0, src->len);
4508         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4509         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4510
4511         return new;
4512 }
4513
4514 struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4515 {
4516         struct extent_buffer *eb;
4517         unsigned long num_pages = num_extent_pages(0, len);
4518         unsigned long i;
4519
4520         eb = __alloc_extent_buffer(NULL, start, len, GFP_NOFS);
4521         if (!eb)
4522                 return NULL;
4523
4524         for (i = 0; i < num_pages; i++) {
4525                 eb->pages[i] = alloc_page(GFP_NOFS);
4526                 if (!eb->pages[i])
4527                         goto err;
4528         }
4529         set_extent_buffer_uptodate(eb);
4530         btrfs_set_header_nritems(eb, 0);
4531         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4532
4533         return eb;
4534 err:
4535         for (; i > 0; i--)
4536                 __free_page(eb->pages[i - 1]);
4537         __free_extent_buffer(eb);
4538         return NULL;
4539 }
4540
4541 static void check_buffer_tree_ref(struct extent_buffer *eb)
4542 {
4543         int refs;
4544         /* the ref bit is tricky.  We have to make sure it is set
4545          * if we have the buffer dirty.   Otherwise the
4546          * code to free a buffer can end up dropping a dirty
4547          * page
4548          *
4549          * Once the ref bit is set, it won't go away while the
4550          * buffer is dirty or in writeback, and it also won't
4551          * go away while we have the reference count on the
4552          * eb bumped.
4553          *
4554          * We can't just set the ref bit without bumping the
4555          * ref on the eb because free_extent_buffer might
4556          * see the ref bit and try to clear it.  If this happens
4557          * free_extent_buffer might end up dropping our original
4558          * ref by mistake and freeing the page before we are able
4559          * to add one more ref.
4560          *
4561          * So bump the ref count first, then set the bit.  If someone
4562          * beat us to it, drop the ref we added.
4563          */
4564         refs = atomic_read(&eb->refs);
4565         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4566                 return;
4567
4568         spin_lock(&eb->refs_lock);
4569         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4570                 atomic_inc(&eb->refs);
4571         spin_unlock(&eb->refs_lock);
4572 }
4573
4574 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4575                 struct page *accessed)
4576 {
4577         unsigned long num_pages, i;
4578
4579         check_buffer_tree_ref(eb);
4580
4581         num_pages = num_extent_pages(eb->start, eb->len);
4582         for (i = 0; i < num_pages; i++) {
4583                 struct page *p = extent_buffer_page(eb, i);
4584                 if (p != accessed)
4585                         mark_page_accessed(p);
4586         }
4587 }
4588
4589 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4590                                          u64 start)
4591 {
4592         struct extent_buffer *eb;
4593
4594         rcu_read_lock();
4595         eb = radix_tree_lookup(&fs_info->buffer_radix,
4596                                start >> PAGE_CACHE_SHIFT);
4597         if (eb && atomic_inc_not_zero(&eb->refs)) {
4598                 rcu_read_unlock();
4599                 mark_extent_buffer_accessed(eb, NULL);
4600                 return eb;
4601         }
4602         rcu_read_unlock();
4603
4604         return NULL;
4605 }
4606
4607 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4608 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4609                                                u64 start, unsigned long len)
4610 {
4611         struct extent_buffer *eb, *exists = NULL;
4612         int ret;
4613
4614         eb = find_extent_buffer(fs_info, start);
4615         if (eb)
4616                 return eb;
4617         eb = alloc_dummy_extent_buffer(start, len);
4618         if (!eb)
4619                 return NULL;
4620         eb->fs_info = fs_info;
4621 again:
4622         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4623         if (ret)
4624                 goto free_eb;
4625         spin_lock(&fs_info->buffer_lock);
4626         ret = radix_tree_insert(&fs_info->buffer_radix,
4627                                 start >> PAGE_CACHE_SHIFT, eb);
4628         spin_unlock(&fs_info->buffer_lock);
4629         radix_tree_preload_end();
4630         if (ret == -EEXIST) {
4631                 exists = find_extent_buffer(fs_info, start);
4632                 if (exists)
4633                         goto free_eb;
4634                 else
4635                         goto again;
4636         }
4637         check_buffer_tree_ref(eb);
4638         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4639
4640         /*
4641          * We will free dummy extent buffer's if they come into
4642          * free_extent_buffer with a ref count of 2, but if we are using this we
4643          * want the buffers to stay in memory until we're done with them, so
4644          * bump the ref count again.
4645          */
4646         atomic_inc(&eb->refs);
4647         return eb;
4648 free_eb:
4649         btrfs_release_extent_buffer(eb);
4650         return exists;
4651 }
4652 #endif
4653
4654 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4655                                           u64 start, unsigned long len)
4656 {
4657         unsigned long num_pages = num_extent_pages(start, len);
4658         unsigned long i;
4659         unsigned long index = start >> PAGE_CACHE_SHIFT;
4660         struct extent_buffer *eb;
4661         struct extent_buffer *exists = NULL;
4662         struct page *p;
4663         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4664         int uptodate = 1;
4665         int ret;
4666
4667         eb = find_extent_buffer(fs_info, start);
4668         if (eb)
4669                 return eb;
4670
4671         eb = __alloc_extent_buffer(fs_info, start, len, GFP_NOFS);
4672         if (!eb)
4673                 return NULL;
4674
4675         for (i = 0; i < num_pages; i++, index++) {
4676                 p = find_or_create_page(mapping, index, GFP_NOFS);
4677                 if (!p)
4678                         goto free_eb;
4679
4680                 spin_lock(&mapping->private_lock);
4681                 if (PagePrivate(p)) {
4682                         /*
4683                          * We could have already allocated an eb for this page
4684                          * and attached one so lets see if we can get a ref on
4685                          * the existing eb, and if we can we know it's good and
4686                          * we can just return that one, else we know we can just
4687                          * overwrite page->private.
4688                          */
4689                         exists = (struct extent_buffer *)p->private;
4690                         if (atomic_inc_not_zero(&exists->refs)) {
4691                                 spin_unlock(&mapping->private_lock);
4692                                 unlock_page(p);
4693                                 page_cache_release(p);
4694                                 mark_extent_buffer_accessed(exists, p);
4695                                 goto free_eb;
4696                         }
4697
4698                         /*
4699                          * Do this so attach doesn't complain and we need to
4700                          * drop the ref the old guy had.
4701                          */
4702                         ClearPagePrivate(p);
4703                         WARN_ON(PageDirty(p));
4704                         page_cache_release(p);
4705                 }
4706                 attach_extent_buffer_page(eb, p);
4707                 spin_unlock(&mapping->private_lock);
4708                 WARN_ON(PageDirty(p));
4709                 eb->pages[i] = p;
4710                 if (!PageUptodate(p))
4711                         uptodate = 0;
4712
4713                 /*
4714                  * see below about how we avoid a nasty race with release page
4715                  * and why we unlock later
4716                  */
4717         }
4718         if (uptodate)
4719                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4720 again:
4721         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4722         if (ret)
4723                 goto free_eb;
4724
4725         spin_lock(&fs_info->buffer_lock);
4726         ret = radix_tree_insert(&fs_info->buffer_radix,
4727                                 start >> PAGE_CACHE_SHIFT, eb);
4728         spin_unlock(&fs_info->buffer_lock);
4729         radix_tree_preload_end();
4730         if (ret == -EEXIST) {
4731                 exists = find_extent_buffer(fs_info, start);
4732                 if (exists)
4733                         goto free_eb;
4734                 else
4735                         goto again;
4736         }
4737         /* add one reference for the tree */
4738         check_buffer_tree_ref(eb);
4739         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4740
4741         /*
4742          * there is a race where release page may have
4743          * tried to find this extent buffer in the radix
4744          * but failed.  It will tell the VM it is safe to
4745          * reclaim the, and it will clear the page private bit.
4746          * We must make sure to set the page private bit properly
4747          * after the extent buffer is in the radix tree so
4748          * it doesn't get lost
4749          */
4750         SetPageChecked(eb->pages[0]);
4751         for (i = 1; i < num_pages; i++) {
4752                 p = extent_buffer_page(eb, i);
4753                 ClearPageChecked(p);
4754                 unlock_page(p);
4755         }
4756         unlock_page(eb->pages[0]);
4757         return eb;
4758
4759 free_eb:
4760         for (i = 0; i < num_pages; i++) {
4761                 if (eb->pages[i])
4762                         unlock_page(eb->pages[i]);
4763         }
4764
4765         WARN_ON(!atomic_dec_and_test(&eb->refs));
4766         btrfs_release_extent_buffer(eb);
4767         return exists;
4768 }
4769
4770 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4771 {
4772         struct extent_buffer *eb =
4773                         container_of(head, struct extent_buffer, rcu_head);
4774
4775         __free_extent_buffer(eb);
4776 }
4777
4778 /* Expects to have eb->eb_lock already held */
4779 static int release_extent_buffer(struct extent_buffer *eb)
4780 {
4781         WARN_ON(atomic_read(&eb->refs) == 0);
4782         if (atomic_dec_and_test(&eb->refs)) {
4783                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4784                         struct btrfs_fs_info *fs_info = eb->fs_info;
4785
4786                         spin_unlock(&eb->refs_lock);
4787
4788                         spin_lock(&fs_info->buffer_lock);
4789                         radix_tree_delete(&fs_info->buffer_radix,
4790                                           eb->start >> PAGE_CACHE_SHIFT);
4791                         spin_unlock(&fs_info->buffer_lock);
4792                 } else {
4793                         spin_unlock(&eb->refs_lock);
4794                 }
4795
4796                 /* Should be safe to release our pages at this point */
4797                 btrfs_release_extent_buffer_page(eb, 0);
4798                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4799                 return 1;
4800         }
4801         spin_unlock(&eb->refs_lock);
4802
4803         return 0;
4804 }
4805
4806 void free_extent_buffer(struct extent_buffer *eb)
4807 {
4808         int refs;
4809         int old;
4810         if (!eb)
4811                 return;
4812
4813         while (1) {
4814                 refs = atomic_read(&eb->refs);
4815                 if (refs <= 3)
4816                         break;
4817                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
4818                 if (old == refs)
4819                         return;
4820         }
4821
4822         spin_lock(&eb->refs_lock);
4823         if (atomic_read(&eb->refs) == 2 &&
4824             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4825                 atomic_dec(&eb->refs);
4826
4827         if (atomic_read(&eb->refs) == 2 &&
4828             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4829             !extent_buffer_under_io(eb) &&
4830             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4831                 atomic_dec(&eb->refs);
4832
4833         /*
4834          * I know this is terrible, but it's temporary until we stop tracking
4835          * the uptodate bits and such for the extent buffers.
4836          */
4837         release_extent_buffer(eb);
4838 }
4839
4840 void free_extent_buffer_stale(struct extent_buffer *eb)
4841 {
4842         if (!eb)
4843                 return;
4844
4845         spin_lock(&eb->refs_lock);
4846         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4847
4848         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4849             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4850                 atomic_dec(&eb->refs);
4851         release_extent_buffer(eb);
4852 }
4853
4854 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4855 {
4856         unsigned long i;
4857         unsigned long num_pages;
4858         struct page *page;
4859
4860         num_pages = num_extent_pages(eb->start, eb->len);
4861
4862         for (i = 0; i < num_pages; i++) {
4863                 page = extent_buffer_page(eb, i);
4864                 if (!PageDirty(page))
4865                         continue;
4866
4867                 lock_page(page);
4868                 WARN_ON(!PagePrivate(page));
4869
4870                 clear_page_dirty_for_io(page);
4871                 spin_lock_irq(&page->mapping->tree_lock);
4872                 if (!PageDirty(page)) {
4873                         radix_tree_tag_clear(&page->mapping->page_tree,
4874                                                 page_index(page),
4875                                                 PAGECACHE_TAG_DIRTY);
4876                 }
4877                 spin_unlock_irq(&page->mapping->tree_lock);
4878                 ClearPageError(page);
4879                 unlock_page(page);
4880         }
4881         WARN_ON(atomic_read(&eb->refs) == 0);
4882 }
4883
4884 int set_extent_buffer_dirty(struct extent_buffer *eb)
4885 {
4886         unsigned long i;
4887         unsigned long num_pages;
4888         int was_dirty = 0;
4889
4890         check_buffer_tree_ref(eb);
4891
4892         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4893
4894         num_pages = num_extent_pages(eb->start, eb->len);
4895         WARN_ON(atomic_read(&eb->refs) == 0);
4896         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4897
4898         for (i = 0; i < num_pages; i++)
4899                 set_page_dirty(extent_buffer_page(eb, i));
4900         return was_dirty;
4901 }
4902
4903 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4904 {
4905         unsigned long i;
4906         struct page *page;
4907         unsigned long num_pages;
4908
4909         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4910         num_pages = num_extent_pages(eb->start, eb->len);
4911         for (i = 0; i < num_pages; i++) {
4912                 page = extent_buffer_page(eb, i);
4913                 if (page)
4914                         ClearPageUptodate(page);
4915         }
4916         return 0;
4917 }
4918
4919 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4920 {
4921         unsigned long i;
4922         struct page *page;
4923         unsigned long num_pages;
4924
4925         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4926         num_pages = num_extent_pages(eb->start, eb->len);
4927         for (i = 0; i < num_pages; i++) {
4928                 page = extent_buffer_page(eb, i);
4929                 SetPageUptodate(page);
4930         }
4931         return 0;
4932 }
4933
4934 int extent_buffer_uptodate(struct extent_buffer *eb)
4935 {
4936         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4937 }
4938
4939 int read_extent_buffer_pages(struct extent_io_tree *tree,
4940                              struct extent_buffer *eb, u64 start, int wait,
4941                              get_extent_t *get_extent, int mirror_num)
4942 {
4943         unsigned long i;
4944         unsigned long start_i;
4945         struct page *page;
4946         int err;
4947         int ret = 0;
4948         int locked_pages = 0;
4949         int all_uptodate = 1;
4950         unsigned long num_pages;
4951         unsigned long num_reads = 0;
4952         struct bio *bio = NULL;
4953         unsigned long bio_flags = 0;
4954
4955         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4956                 return 0;
4957
4958         if (start) {
4959                 WARN_ON(start < eb->start);
4960                 start_i = (start >> PAGE_CACHE_SHIFT) -
4961                         (eb->start >> PAGE_CACHE_SHIFT);
4962         } else {
4963                 start_i = 0;
4964         }
4965
4966         num_pages = num_extent_pages(eb->start, eb->len);
4967         for (i = start_i; i < num_pages; i++) {
4968                 page = extent_buffer_page(eb, i);
4969                 if (wait == WAIT_NONE) {
4970                         if (!trylock_page(page))
4971                                 goto unlock_exit;
4972                 } else {
4973                         lock_page(page);
4974                 }
4975                 locked_pages++;
4976                 if (!PageUptodate(page)) {
4977                         num_reads++;
4978                         all_uptodate = 0;
4979                 }
4980         }
4981         if (all_uptodate) {
4982                 if (start_i == 0)
4983                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4984                 goto unlock_exit;
4985         }
4986
4987         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4988         eb->read_mirror = 0;
4989         atomic_set(&eb->io_pages, num_reads);
4990         for (i = start_i; i < num_pages; i++) {
4991                 page = extent_buffer_page(eb, i);
4992                 if (!PageUptodate(page)) {
4993                         ClearPageError(page);
4994                         err = __extent_read_full_page(tree, page,
4995                                                       get_extent, &bio,
4996                                                       mirror_num, &bio_flags,
4997                                                       READ | REQ_META);
4998                         if (err)
4999                                 ret = err;
5000                 } else {
5001                         unlock_page(page);
5002                 }
5003         }
5004
5005         if (bio) {
5006                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
5007                                      bio_flags);
5008                 if (err)
5009                         return err;
5010         }
5011
5012         if (ret || wait != WAIT_COMPLETE)
5013                 return ret;
5014
5015         for (i = start_i; i < num_pages; i++) {
5016                 page = extent_buffer_page(eb, i);
5017                 wait_on_page_locked(page);
5018                 if (!PageUptodate(page))
5019                         ret = -EIO;
5020         }
5021
5022         return ret;
5023
5024 unlock_exit:
5025         i = start_i;
5026         while (locked_pages > 0) {
5027                 page = extent_buffer_page(eb, i);
5028                 i++;
5029                 unlock_page(page);
5030                 locked_pages--;
5031         }
5032         return ret;
5033 }
5034
5035 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5036                         unsigned long start,
5037                         unsigned long len)
5038 {
5039         size_t cur;
5040         size_t offset;
5041         struct page *page;
5042         char *kaddr;
5043         char *dst = (char *)dstv;
5044         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5045         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5046
5047         WARN_ON(start > eb->len);
5048         WARN_ON(start + len > eb->start + eb->len);
5049
5050         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5051
5052         while (len > 0) {
5053                 page = extent_buffer_page(eb, i);
5054
5055                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5056                 kaddr = page_address(page);
5057                 memcpy(dst, kaddr + offset, cur);
5058
5059                 dst += cur;
5060                 len -= cur;
5061                 offset = 0;
5062                 i++;
5063         }
5064 }
5065
5066 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5067                         unsigned long start,
5068                         unsigned long len)
5069 {
5070         size_t cur;
5071         size_t offset;
5072         struct page *page;
5073         char *kaddr;
5074         char __user *dst = (char __user *)dstv;
5075         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5076         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5077         int ret = 0;
5078
5079         WARN_ON(start > eb->len);
5080         WARN_ON(start + len > eb->start + eb->len);
5081
5082         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5083
5084         while (len > 0) {
5085                 page = extent_buffer_page(eb, i);
5086
5087                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5088                 kaddr = page_address(page);
5089                 if (copy_to_user(dst, kaddr + offset, cur)) {
5090                         ret = -EFAULT;
5091                         break;
5092                 }
5093
5094                 dst += cur;
5095                 len -= cur;
5096                 offset = 0;
5097                 i++;
5098         }
5099
5100         return ret;
5101 }
5102
5103 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5104                                unsigned long min_len, char **map,
5105                                unsigned long *map_start,
5106                                unsigned long *map_len)
5107 {
5108         size_t offset = start & (PAGE_CACHE_SIZE - 1);
5109         char *kaddr;
5110         struct page *p;
5111         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5112         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5113         unsigned long end_i = (start_offset + start + min_len - 1) >>
5114                 PAGE_CACHE_SHIFT;
5115
5116         if (i != end_i)
5117                 return -EINVAL;
5118
5119         if (i == 0) {
5120                 offset = start_offset;
5121                 *map_start = 0;
5122         } else {
5123                 offset = 0;
5124                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
5125         }
5126
5127         if (start + min_len > eb->len) {
5128                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
5129                        "wanted %lu %lu\n",
5130                        eb->start, eb->len, start, min_len);
5131                 return -EINVAL;
5132         }
5133
5134         p = extent_buffer_page(eb, i);
5135         kaddr = page_address(p);
5136         *map = kaddr + offset;
5137         *map_len = PAGE_CACHE_SIZE - offset;
5138         return 0;
5139 }
5140
5141 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5142                           unsigned long start,
5143                           unsigned long len)
5144 {
5145         size_t cur;
5146         size_t offset;
5147         struct page *page;
5148         char *kaddr;
5149         char *ptr = (char *)ptrv;
5150         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5151         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5152         int ret = 0;
5153
5154         WARN_ON(start > eb->len);
5155         WARN_ON(start + len > eb->start + eb->len);
5156
5157         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5158
5159         while (len > 0) {
5160                 page = extent_buffer_page(eb, i);
5161
5162                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5163
5164                 kaddr = page_address(page);
5165                 ret = memcmp(ptr, kaddr + offset, cur);
5166                 if (ret)
5167                         break;
5168
5169                 ptr += cur;
5170                 len -= cur;
5171                 offset = 0;
5172                 i++;
5173         }
5174         return ret;
5175 }
5176
5177 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5178                          unsigned long start, unsigned long len)
5179 {
5180         size_t cur;
5181         size_t offset;
5182         struct page *page;
5183         char *kaddr;
5184         char *src = (char *)srcv;
5185         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5186         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5187
5188         WARN_ON(start > eb->len);
5189         WARN_ON(start + len > eb->start + eb->len);
5190
5191         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5192
5193         while (len > 0) {
5194                 page = extent_buffer_page(eb, i);
5195                 WARN_ON(!PageUptodate(page));
5196
5197                 cur = min(len, PAGE_CACHE_SIZE - offset);
5198                 kaddr = page_address(page);
5199                 memcpy(kaddr + offset, src, cur);
5200
5201                 src += cur;
5202                 len -= cur;
5203                 offset = 0;
5204                 i++;
5205         }
5206 }
5207
5208 void memset_extent_buffer(struct extent_buffer *eb, char c,
5209                           unsigned long start, unsigned long len)
5210 {
5211         size_t cur;
5212         size_t offset;
5213         struct page *page;
5214         char *kaddr;
5215         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5216         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5217
5218         WARN_ON(start > eb->len);
5219         WARN_ON(start + len > eb->start + eb->len);
5220
5221         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5222
5223         while (len > 0) {
5224                 page = extent_buffer_page(eb, i);
5225                 WARN_ON(!PageUptodate(page));
5226
5227                 cur = min(len, PAGE_CACHE_SIZE - offset);
5228                 kaddr = page_address(page);
5229                 memset(kaddr + offset, c, cur);
5230
5231                 len -= cur;
5232                 offset = 0;
5233                 i++;
5234         }
5235 }
5236
5237 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5238                         unsigned long dst_offset, unsigned long src_offset,
5239                         unsigned long len)
5240 {
5241         u64 dst_len = dst->len;
5242         size_t cur;
5243         size_t offset;
5244         struct page *page;
5245         char *kaddr;
5246         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5247         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5248
5249         WARN_ON(src->len != dst_len);
5250
5251         offset = (start_offset + dst_offset) &
5252                 (PAGE_CACHE_SIZE - 1);
5253
5254         while (len > 0) {
5255                 page = extent_buffer_page(dst, i);
5256                 WARN_ON(!PageUptodate(page));
5257
5258                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5259
5260                 kaddr = page_address(page);
5261                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5262
5263                 src_offset += cur;
5264                 len -= cur;
5265                 offset = 0;
5266                 i++;
5267         }
5268 }
5269
5270 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5271 {
5272         unsigned long distance = (src > dst) ? src - dst : dst - src;
5273         return distance < len;
5274 }
5275
5276 static void copy_pages(struct page *dst_page, struct page *src_page,
5277                        unsigned long dst_off, unsigned long src_off,
5278                        unsigned long len)
5279 {
5280         char *dst_kaddr = page_address(dst_page);
5281         char *src_kaddr;
5282         int must_memmove = 0;
5283
5284         if (dst_page != src_page) {
5285                 src_kaddr = page_address(src_page);
5286         } else {
5287                 src_kaddr = dst_kaddr;
5288                 if (areas_overlap(src_off, dst_off, len))
5289                         must_memmove = 1;
5290         }
5291
5292         if (must_memmove)
5293                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5294         else
5295                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5296 }
5297
5298 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5299                            unsigned long src_offset, unsigned long len)
5300 {
5301         size_t cur;
5302         size_t dst_off_in_page;
5303         size_t src_off_in_page;
5304         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5305         unsigned long dst_i;
5306         unsigned long src_i;
5307
5308         if (src_offset + len > dst->len) {
5309                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5310                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5311                 BUG_ON(1);
5312         }
5313         if (dst_offset + len > dst->len) {
5314                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5315                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5316                 BUG_ON(1);
5317         }
5318
5319         while (len > 0) {
5320                 dst_off_in_page = (start_offset + dst_offset) &
5321                         (PAGE_CACHE_SIZE - 1);
5322                 src_off_in_page = (start_offset + src_offset) &
5323                         (PAGE_CACHE_SIZE - 1);
5324
5325                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5326                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5327
5328                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5329                                                src_off_in_page));
5330                 cur = min_t(unsigned long, cur,
5331                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5332
5333                 copy_pages(extent_buffer_page(dst, dst_i),
5334                            extent_buffer_page(dst, src_i),
5335                            dst_off_in_page, src_off_in_page, cur);
5336
5337                 src_offset += cur;
5338                 dst_offset += cur;
5339                 len -= cur;
5340         }
5341 }
5342
5343 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5344                            unsigned long src_offset, unsigned long len)
5345 {
5346         size_t cur;
5347         size_t dst_off_in_page;
5348         size_t src_off_in_page;
5349         unsigned long dst_end = dst_offset + len - 1;
5350         unsigned long src_end = src_offset + len - 1;
5351         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5352         unsigned long dst_i;
5353         unsigned long src_i;
5354
5355         if (src_offset + len > dst->len) {
5356                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5357                        "len %lu len %lu\n", src_offset, len, dst->len);
5358                 BUG_ON(1);
5359         }
5360         if (dst_offset + len > dst->len) {
5361                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5362                        "len %lu len %lu\n", dst_offset, len, dst->len);
5363                 BUG_ON(1);
5364         }
5365         if (dst_offset < src_offset) {
5366                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5367                 return;
5368         }
5369         while (len > 0) {
5370                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5371                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5372
5373                 dst_off_in_page = (start_offset + dst_end) &
5374                         (PAGE_CACHE_SIZE - 1);
5375                 src_off_in_page = (start_offset + src_end) &
5376                         (PAGE_CACHE_SIZE - 1);
5377
5378                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5379                 cur = min(cur, dst_off_in_page + 1);
5380                 copy_pages(extent_buffer_page(dst, dst_i),
5381                            extent_buffer_page(dst, src_i),
5382                            dst_off_in_page - cur + 1,
5383                            src_off_in_page - cur + 1, cur);
5384
5385                 dst_end -= cur;
5386                 src_end -= cur;
5387                 len -= cur;
5388         }
5389 }
5390
5391 int try_release_extent_buffer(struct page *page)
5392 {
5393         struct extent_buffer *eb;
5394
5395         /*
5396          * We need to make sure noboody is attaching this page to an eb right
5397          * now.
5398          */
5399         spin_lock(&page->mapping->private_lock);
5400         if (!PagePrivate(page)) {
5401                 spin_unlock(&page->mapping->private_lock);
5402                 return 1;
5403         }
5404
5405         eb = (struct extent_buffer *)page->private;
5406         BUG_ON(!eb);
5407
5408         /*
5409          * This is a little awful but should be ok, we need to make sure that
5410          * the eb doesn't disappear out from under us while we're looking at
5411          * this page.
5412          */
5413         spin_lock(&eb->refs_lock);
5414         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5415                 spin_unlock(&eb->refs_lock);
5416                 spin_unlock(&page->mapping->private_lock);
5417                 return 0;
5418         }
5419         spin_unlock(&page->mapping->private_lock);
5420
5421         /*
5422          * If tree ref isn't set then we know the ref on this eb is a real ref,
5423          * so just return, this page will likely be freed soon anyway.
5424          */
5425         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5426                 spin_unlock(&eb->refs_lock);
5427                 return 0;
5428         }
5429
5430         return release_extent_buffer(eb);
5431 }