linux/fs/btrfs/extent_io.c
<<
>>
Prefs
   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/module.h>
   8#include <linux/spinlock.h>
   9#include <linux/blkdev.h>
  10#include <linux/swap.h>
  11#include <linux/writeback.h>
  12#include <linux/pagevec.h>
  13#include <linux/prefetch.h>
  14#include <linux/cleancache.h>
  15#include "extent_io.h"
  16#include "extent_map.h"
  17#include "compat.h"
  18#include "ctree.h"
  19#include "btrfs_inode.h"
  20#include "volumes.h"
  21#include "check-integrity.h"
  22#include "locking.h"
  23#include "rcu-string.h"
  24
  25static struct kmem_cache *extent_state_cache;
  26static struct kmem_cache *extent_buffer_cache;
  27
  28static LIST_HEAD(buffers);
  29static LIST_HEAD(states);
  30
  31#define LEAK_DEBUG 0
  32#if LEAK_DEBUG
  33static DEFINE_SPINLOCK(leak_lock);
  34#endif
  35
  36#define BUFFER_LRU_MAX 64
  37
  38struct tree_entry {
  39        u64 start;
  40        u64 end;
  41        struct rb_node rb_node;
  42};
  43
  44struct extent_page_data {
  45        struct bio *bio;
  46        struct extent_io_tree *tree;
  47        get_extent_t *get_extent;
  48        unsigned long bio_flags;
  49
  50        /* tells writepage not to lock the state bits for this range
  51         * it still does the unlocking
  52         */
  53        unsigned int extent_locked:1;
  54
  55        /* tells the submit_bio code to use a WRITE_SYNC */
  56        unsigned int sync_io:1;
  57};
  58
  59static noinline void flush_write_bio(void *data);
  60static inline struct btrfs_fs_info *
  61tree_fs_info(struct extent_io_tree *tree)
  62{
  63        return btrfs_sb(tree->mapping->host->i_sb);
  64}
  65
  66int __init extent_io_init(void)
  67{
  68        extent_state_cache = kmem_cache_create("btrfs_extent_state",
  69                        sizeof(struct extent_state), 0,
  70                        SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  71        if (!extent_state_cache)
  72                return -ENOMEM;
  73
  74        extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
  75                        sizeof(struct extent_buffer), 0,
  76                        SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  77        if (!extent_buffer_cache)
  78                goto free_state_cache;
  79        return 0;
  80
  81free_state_cache:
  82        kmem_cache_destroy(extent_state_cache);
  83        return -ENOMEM;
  84}
  85
  86void extent_io_exit(void)
  87{
  88        struct extent_state *state;
  89        struct extent_buffer *eb;
  90
  91        while (!list_empty(&states)) {
  92                state = list_entry(states.next, struct extent_state, leak_list);
  93                printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  94                       "state %lu in tree %p refs %d\n",
  95                       (unsigned long long)state->start,
  96                       (unsigned long long)state->end,
  97                       state->state, state->tree, atomic_read(&state->refs));
  98                list_del(&state->leak_list);
  99                kmem_cache_free(extent_state_cache, state);
 100
 101        }
 102
 103        while (!list_empty(&buffers)) {
 104                eb = list_entry(buffers.next, struct extent_buffer, leak_list);
 105                printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
 106                       "refs %d\n", (unsigned long long)eb->start,
 107                       eb->len, atomic_read(&eb->refs));
 108                list_del(&eb->leak_list);
 109                kmem_cache_free(extent_buffer_cache, eb);
 110        }
 111
 112        /*
 113         * Make sure all delayed rcu free are flushed before we
 114         * destroy caches.
 115         */
 116        rcu_barrier();
 117        if (extent_state_cache)
 118                kmem_cache_destroy(extent_state_cache);
 119        if (extent_buffer_cache)
 120                kmem_cache_destroy(extent_buffer_cache);
 121}
 122
 123void extent_io_tree_init(struct extent_io_tree *tree,
 124                         struct address_space *mapping)
 125{
 126        tree->state = RB_ROOT;
 127        INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 128        tree->ops = NULL;
 129        tree->dirty_bytes = 0;
 130        spin_lock_init(&tree->lock);
 131        spin_lock_init(&tree->buffer_lock);
 132        tree->mapping = mapping;
 133}
 134
 135static struct extent_state *alloc_extent_state(gfp_t mask)
 136{
 137        struct extent_state *state;
 138#if LEAK_DEBUG
 139        unsigned long flags;
 140#endif
 141
 142        state = kmem_cache_alloc(extent_state_cache, mask);
 143        if (!state)
 144                return state;
 145        state->state = 0;
 146        state->private = 0;
 147        state->tree = NULL;
 148#if LEAK_DEBUG
 149        spin_lock_irqsave(&leak_lock, flags);
 150        list_add(&state->leak_list, &states);
 151        spin_unlock_irqrestore(&leak_lock, flags);
 152#endif
 153        atomic_set(&state->refs, 1);
 154        init_waitqueue_head(&state->wq);
 155        trace_alloc_extent_state(state, mask, _RET_IP_);
 156        return state;
 157}
 158
 159void free_extent_state(struct extent_state *state)
 160{
 161        if (!state)
 162                return;
 163        if (atomic_dec_and_test(&state->refs)) {
 164#if LEAK_DEBUG
 165                unsigned long flags;
 166#endif
 167                WARN_ON(state->tree);
 168#if LEAK_DEBUG
 169                spin_lock_irqsave(&leak_lock, flags);
 170                list_del(&state->leak_list);
 171                spin_unlock_irqrestore(&leak_lock, flags);
 172#endif
 173                trace_free_extent_state(state, _RET_IP_);
 174                kmem_cache_free(extent_state_cache, state);
 175        }
 176}
 177
 178static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 179                                   struct rb_node *node)
 180{
 181        struct rb_node **p = &root->rb_node;
 182        struct rb_node *parent = NULL;
 183        struct tree_entry *entry;
 184
 185        while (*p) {
 186                parent = *p;
 187                entry = rb_entry(parent, struct tree_entry, rb_node);
 188
 189                if (offset < entry->start)
 190                        p = &(*p)->rb_left;
 191                else if (offset > entry->end)
 192                        p = &(*p)->rb_right;
 193                else
 194                        return parent;
 195        }
 196
 197        rb_link_node(node, parent, p);
 198        rb_insert_color(node, root);
 199        return NULL;
 200}
 201
 202static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 203                                     struct rb_node **prev_ret,
 204                                     struct rb_node **next_ret)
 205{
 206        struct rb_root *root = &tree->state;
 207        struct rb_node *n = root->rb_node;
 208        struct rb_node *prev = NULL;
 209        struct rb_node *orig_prev = NULL;
 210        struct tree_entry *entry;
 211        struct tree_entry *prev_entry = NULL;
 212
 213        while (n) {
 214                entry = rb_entry(n, struct tree_entry, rb_node);
 215                prev = n;
 216                prev_entry = entry;
 217
 218                if (offset < entry->start)
 219                        n = n->rb_left;
 220                else if (offset > entry->end)
 221                        n = n->rb_right;
 222                else
 223                        return n;
 224        }
 225
 226        if (prev_ret) {
 227                orig_prev = prev;
 228                while (prev && offset > prev_entry->end) {
 229                        prev = rb_next(prev);
 230                        prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 231                }
 232                *prev_ret = prev;
 233                prev = orig_prev;
 234        }
 235
 236        if (next_ret) {
 237                prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 238                while (prev && offset < prev_entry->start) {
 239                        prev = rb_prev(prev);
 240                        prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 241                }
 242                *next_ret = prev;
 243        }
 244        return NULL;
 245}
 246
 247static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 248                                          u64 offset)
 249{
 250        struct rb_node *prev = NULL;
 251        struct rb_node *ret;
 252
 253        ret = __etree_search(tree, offset, &prev, NULL);
 254        if (!ret)
 255                return prev;
 256        return ret;
 257}
 258
 259static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 260                     struct extent_state *other)
 261{
 262        if (tree->ops && tree->ops->merge_extent_hook)
 263                tree->ops->merge_extent_hook(tree->mapping->host, new,
 264                                             other);
 265}
 266
 267/*
 268 * utility function to look for merge candidates inside a given range.
 269 * Any extents with matching state are merged together into a single
 270 * extent in the tree.  Extents with EXTENT_IO in their state field
 271 * are not merged because the end_io handlers need to be able to do
 272 * operations on them without sleeping (or doing allocations/splits).
 273 *
 274 * This should be called with the tree lock held.
 275 */
 276static void merge_state(struct extent_io_tree *tree,
 277                        struct extent_state *state)
 278{
 279        struct extent_state *other;
 280        struct rb_node *other_node;
 281
 282        if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 283                return;
 284
 285        other_node = rb_prev(&state->rb_node);
 286        if (other_node) {
 287                other = rb_entry(other_node, struct extent_state, rb_node);
 288                if (other->end == state->start - 1 &&
 289                    other->state == state->state) {
 290                        merge_cb(tree, state, other);
 291                        state->start = other->start;
 292                        other->tree = NULL;
 293                        rb_erase(&other->rb_node, &tree->state);
 294                        free_extent_state(other);
 295                }
 296        }
 297        other_node = rb_next(&state->rb_node);
 298        if (other_node) {
 299                other = rb_entry(other_node, struct extent_state, rb_node);
 300                if (other->start == state->end + 1 &&
 301                    other->state == state->state) {
 302                        merge_cb(tree, state, other);
 303                        state->end = other->end;
 304                        other->tree = NULL;
 305                        rb_erase(&other->rb_node, &tree->state);
 306                        free_extent_state(other);
 307                }
 308        }
 309}
 310
 311static void set_state_cb(struct extent_io_tree *tree,
 312                         struct extent_state *state, int *bits)
 313{
 314        if (tree->ops && tree->ops->set_bit_hook)
 315                tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 316}
 317
 318static void clear_state_cb(struct extent_io_tree *tree,
 319                           struct extent_state *state, int *bits)
 320{
 321        if (tree->ops && tree->ops->clear_bit_hook)
 322                tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 323}
 324
 325static void set_state_bits(struct extent_io_tree *tree,
 326                           struct extent_state *state, int *bits);
 327
 328/*
 329 * insert an extent_state struct into the tree.  'bits' are set on the
 330 * struct before it is inserted.
 331 *
 332 * This may return -EEXIST if the extent is already there, in which case the
 333 * state struct is freed.
 334 *
 335 * The tree lock is not taken internally.  This is a utility function and
 336 * probably isn't what you want to call (see set/clear_extent_bit).
 337 */
 338static int insert_state(struct extent_io_tree *tree,
 339                        struct extent_state *state, u64 start, u64 end,
 340                        int *bits)
 341{
 342        struct rb_node *node;
 343
 344        if (end < start) {
 345                printk(KERN_ERR "btrfs end < start %llu %llu\n",
 346                       (unsigned long long)end,
 347                       (unsigned long long)start);
 348                WARN_ON(1);
 349        }
 350        state->start = start;
 351        state->end = end;
 352
 353        set_state_bits(tree, state, bits);
 354
 355        node = tree_insert(&tree->state, end, &state->rb_node);
 356        if (node) {
 357                struct extent_state *found;
 358                found = rb_entry(node, struct extent_state, rb_node);
 359                printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 360                       "%llu %llu\n", (unsigned long long)found->start,
 361                       (unsigned long long)found->end,
 362                       (unsigned long long)start, (unsigned long long)end);
 363                return -EEXIST;
 364        }
 365        state->tree = tree;
 366        merge_state(tree, state);
 367        return 0;
 368}
 369
 370static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 371                     u64 split)
 372{
 373        if (tree->ops && tree->ops->split_extent_hook)
 374                tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 375}
 376
 377/*
 378 * split a given extent state struct in two, inserting the preallocated
 379 * struct 'prealloc' as the newly created second half.  'split' indicates an
 380 * offset inside 'orig' where it should be split.
 381 *
 382 * Before calling,
 383 * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 384 * are two extent state structs in the tree:
 385 * prealloc: [orig->start, split - 1]
 386 * orig: [ split, orig->end ]
 387 *
 388 * The tree locks are not taken by this function. They need to be held
 389 * by the caller.
 390 */
 391static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 392                       struct extent_state *prealloc, u64 split)
 393{
 394        struct rb_node *node;
 395
 396        split_cb(tree, orig, split);
 397
 398        prealloc->start = orig->start;
 399        prealloc->end = split - 1;
 400        prealloc->state = orig->state;
 401        orig->start = split;
 402
 403        node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 404        if (node) {
 405                free_extent_state(prealloc);
 406                return -EEXIST;
 407        }
 408        prealloc->tree = tree;
 409        return 0;
 410}
 411
 412static struct extent_state *next_state(struct extent_state *state)
 413{
 414        struct rb_node *next = rb_next(&state->rb_node);
 415        if (next)
 416                return rb_entry(next, struct extent_state, rb_node);
 417        else
 418                return NULL;
 419}
 420
 421/*
 422 * utility function to clear some bits in an extent state struct.
 423 * it will optionally wake up any one waiting on this state (wake == 1).
 424 *
 425 * If no bits are set on the state struct after clearing things, the
 426 * struct is freed and removed from the tree
 427 */
 428static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 429                                            struct extent_state *state,
 430                                            int *bits, int wake)
 431{
 432        struct extent_state *next;
 433        int bits_to_clear = *bits & ~EXTENT_CTLBITS;
 434
 435        if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 436                u64 range = state->end - state->start + 1;
 437                WARN_ON(range > tree->dirty_bytes);
 438                tree->dirty_bytes -= range;
 439        }
 440        clear_state_cb(tree, state, bits);
 441        state->state &= ~bits_to_clear;
 442        if (wake)
 443                wake_up(&state->wq);
 444        if (state->state == 0) {
 445                next = next_state(state);
 446                if (state->tree) {
 447                        rb_erase(&state->rb_node, &tree->state);
 448                        state->tree = NULL;
 449                        free_extent_state(state);
 450                } else {
 451                        WARN_ON(1);
 452                }
 453        } else {
 454                merge_state(tree, state);
 455                next = next_state(state);
 456        }
 457        return next;
 458}
 459
 460static struct extent_state *
 461alloc_extent_state_atomic(struct extent_state *prealloc)
 462{
 463        if (!prealloc)
 464                prealloc = alloc_extent_state(GFP_ATOMIC);
 465
 466        return prealloc;
 467}
 468
 469void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 470{
 471        btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 472                    "Extent tree was modified by another "
 473                    "thread while locked.");
 474}
 475
 476/*
 477 * clear some bits on a range in the tree.  This may require splitting
 478 * or inserting elements in the tree, so the gfp mask is used to
 479 * indicate which allocations or sleeping are allowed.
 480 *
 481 * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 482 * the given range from the tree regardless of state (ie for truncate).
 483 *
 484 * the range [start, end] is inclusive.
 485 *
 486 * This takes the tree lock, and returns 0 on success and < 0 on error.
 487 */
 488int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 489                     int bits, int wake, int delete,
 490                     struct extent_state **cached_state,
 491                     gfp_t mask)
 492{
 493        struct extent_state *state;
 494        struct extent_state *cached;
 495        struct extent_state *prealloc = NULL;
 496        struct rb_node *node;
 497        u64 last_end;
 498        int err;
 499        int clear = 0;
 500
 501        if (delete)
 502                bits |= ~EXTENT_CTLBITS;
 503        bits |= EXTENT_FIRST_DELALLOC;
 504
 505        if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 506                clear = 1;
 507again:
 508        if (!prealloc && (mask & __GFP_WAIT)) {
 509                prealloc = alloc_extent_state(mask);
 510                if (!prealloc)
 511                        return -ENOMEM;
 512        }
 513
 514        spin_lock(&tree->lock);
 515        if (cached_state) {
 516                cached = *cached_state;
 517
 518                if (clear) {
 519                        *cached_state = NULL;
 520                        cached_state = NULL;
 521                }
 522
 523                if (cached && cached->tree && cached->start <= start &&
 524                    cached->end > start) {
 525                        if (clear)
 526                                atomic_dec(&cached->refs);
 527                        state = cached;
 528                        goto hit_next;
 529                }
 530                if (clear)
 531                        free_extent_state(cached);
 532        }
 533        /*
 534         * this search will find the extents that end after
 535         * our range starts
 536         */
 537        node = tree_search(tree, start);
 538        if (!node)
 539                goto out;
 540        state = rb_entry(node, struct extent_state, rb_node);
 541hit_next:
 542        if (state->start > end)
 543                goto out;
 544        WARN_ON(state->end < start);
 545        last_end = state->end;
 546
 547        /* the state doesn't have the wanted bits, go ahead */
 548        if (!(state->state & bits)) {
 549                state = next_state(state);
 550                goto next;
 551        }
 552
 553        /*
 554         *     | ---- desired range ---- |
 555         *  | state | or
 556         *  | ------------- state -------------- |
 557         *
 558         * We need to split the extent we found, and may flip
 559         * bits on second half.
 560         *
 561         * If the extent we found extends past our range, we
 562         * just split and search again.  It'll get split again
 563         * the next time though.
 564         *
 565         * If the extent we found is inside our range, we clear
 566         * the desired bit on it.
 567         */
 568
 569        if (state->start < start) {
 570                prealloc = alloc_extent_state_atomic(prealloc);
 571                BUG_ON(!prealloc);
 572                err = split_state(tree, state, prealloc, start);
 573                if (err)
 574                        extent_io_tree_panic(tree, err);
 575
 576                prealloc = NULL;
 577                if (err)
 578                        goto out;
 579                if (state->end <= end) {
 580                        state = clear_state_bit(tree, state, &bits, wake);
 581                        goto next;
 582                }
 583                goto search_again;
 584        }
 585        /*
 586         * | ---- desired range ---- |
 587         *                        | state |
 588         * We need to split the extent, and clear the bit
 589         * on the first half
 590         */
 591        if (state->start <= end && state->end > end) {
 592                prealloc = alloc_extent_state_atomic(prealloc);
 593                BUG_ON(!prealloc);
 594                err = split_state(tree, state, prealloc, end + 1);
 595                if (err)
 596                        extent_io_tree_panic(tree, err);
 597
 598                if (wake)
 599                        wake_up(&state->wq);
 600
 601                clear_state_bit(tree, prealloc, &bits, wake);
 602
 603                prealloc = NULL;
 604                goto out;
 605        }
 606
 607        state = clear_state_bit(tree, state, &bits, wake);
 608next:
 609        if (last_end == (u64)-1)
 610                goto out;
 611        start = last_end + 1;
 612        if (start <= end && state && !need_resched())
 613                goto hit_next;
 614        goto search_again;
 615
 616out:
 617        spin_unlock(&tree->lock);
 618        if (prealloc)
 619                free_extent_state(prealloc);
 620
 621        return 0;
 622
 623search_again:
 624        if (start > end)
 625                goto out;
 626        spin_unlock(&tree->lock);
 627        if (mask & __GFP_WAIT)
 628                cond_resched();
 629        goto again;
 630}
 631
 632static void wait_on_state(struct extent_io_tree *tree,
 633                          struct extent_state *state)
 634                __releases(tree->lock)
 635                __acquires(tree->lock)
 636{
 637        DEFINE_WAIT(wait);
 638        prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 639        spin_unlock(&tree->lock);
 640        schedule();
 641        spin_lock(&tree->lock);
 642        finish_wait(&state->wq, &wait);
 643}
 644
 645/*
 646 * waits for one or more bits to clear on a range in the state tree.
 647 * The range [start, end] is inclusive.
 648 * The tree lock is taken by this function
 649 */
 650void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 651{
 652        struct extent_state *state;
 653        struct rb_node *node;
 654
 655        spin_lock(&tree->lock);
 656again:
 657        while (1) {
 658                /*
 659                 * this search will find all the extents that end after
 660                 * our range starts
 661                 */
 662                node = tree_search(tree, start);
 663                if (!node)
 664                        break;
 665
 666                state = rb_entry(node, struct extent_state, rb_node);
 667
 668                if (state->start > end)
 669                        goto out;
 670
 671                if (state->state & bits) {
 672                        start = state->start;
 673                        atomic_inc(&state->refs);
 674                        wait_on_state(tree, state);
 675                        free_extent_state(state);
 676                        goto again;
 677                }
 678                start = state->end + 1;
 679
 680                if (start > end)
 681                        break;
 682
 683                cond_resched_lock(&tree->lock);
 684        }
 685out:
 686        spin_unlock(&tree->lock);
 687}
 688
 689static void set_state_bits(struct extent_io_tree *tree,
 690                           struct extent_state *state,
 691                           int *bits)
 692{
 693        int bits_to_set = *bits & ~EXTENT_CTLBITS;
 694
 695        set_state_cb(tree, state, bits);
 696        if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 697                u64 range = state->end - state->start + 1;
 698                tree->dirty_bytes += range;
 699        }
 700        state->state |= bits_to_set;
 701}
 702
 703static void cache_state(struct extent_state *state,
 704                        struct extent_state **cached_ptr)
 705{
 706        if (cached_ptr && !(*cached_ptr)) {
 707                if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 708                        *cached_ptr = state;
 709                        atomic_inc(&state->refs);
 710                }
 711        }
 712}
 713
 714static void uncache_state(struct extent_state **cached_ptr)
 715{
 716        if (cached_ptr && (*cached_ptr)) {
 717                struct extent_state *state = *cached_ptr;
 718                *cached_ptr = NULL;
 719                free_extent_state(state);
 720        }
 721}
 722
 723/*
 724 * set some bits on a range in the tree.  This may require allocations or
 725 * sleeping, so the gfp mask is used to indicate what is allowed.
 726 *
 727 * If any of the exclusive bits are set, this will fail with -EEXIST if some
 728 * part of the range already has the desired bits set.  The start of the
 729 * existing range is returned in failed_start in this case.
 730 *
 731 * [start, end] is inclusive This takes the tree lock.
 732 */
 733
 734static int __must_check
 735__set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 736                 int bits, int exclusive_bits, u64 *failed_start,
 737                 struct extent_state **cached_state, gfp_t mask)
 738{
 739        struct extent_state *state;
 740        struct extent_state *prealloc = NULL;
 741        struct rb_node *node;
 742        int err = 0;
 743        u64 last_start;
 744        u64 last_end;
 745
 746        bits |= EXTENT_FIRST_DELALLOC;
 747again:
 748        if (!prealloc && (mask & __GFP_WAIT)) {
 749                prealloc = alloc_extent_state(mask);
 750                BUG_ON(!prealloc);
 751        }
 752
 753        spin_lock(&tree->lock);
 754        if (cached_state && *cached_state) {
 755                state = *cached_state;
 756                if (state->start <= start && state->end > start &&
 757                    state->tree) {
 758                        node = &state->rb_node;
 759                        goto hit_next;
 760                }
 761        }
 762        /*
 763         * this search will find all the extents that end after
 764         * our range starts.
 765         */
 766        node = tree_search(tree, start);
 767        if (!node) {
 768                prealloc = alloc_extent_state_atomic(prealloc);
 769                BUG_ON(!prealloc);
 770                err = insert_state(tree, prealloc, start, end, &bits);
 771                if (err)
 772                        extent_io_tree_panic(tree, err);
 773
 774                prealloc = NULL;
 775                goto out;
 776        }
 777        state = rb_entry(node, struct extent_state, rb_node);
 778hit_next:
 779        last_start = state->start;
 780        last_end = state->end;
 781
 782        /*
 783         * | ---- desired range ---- |
 784         * | state |
 785         *
 786         * Just lock what we found and keep going
 787         */
 788        if (state->start == start && state->end <= end) {
 789                if (state->state & exclusive_bits) {
 790                        *failed_start = state->start;
 791                        err = -EEXIST;
 792                        goto out;
 793                }
 794
 795                set_state_bits(tree, state, &bits);
 796                cache_state(state, cached_state);
 797                merge_state(tree, state);
 798                if (last_end == (u64)-1)
 799                        goto out;
 800                start = last_end + 1;
 801                state = next_state(state);
 802                if (start < end && state && state->start == start &&
 803                    !need_resched())
 804                        goto hit_next;
 805                goto search_again;
 806        }
 807
 808        /*
 809         *     | ---- desired range ---- |
 810         * | state |
 811         *   or
 812         * | ------------- state -------------- |
 813         *
 814         * We need to split the extent we found, and may flip bits on
 815         * second half.
 816         *
 817         * If the extent we found extends past our
 818         * range, we just split and search again.  It'll get split
 819         * again the next time though.
 820         *
 821         * If the extent we found is inside our range, we set the
 822         * desired bit on it.
 823         */
 824        if (state->start < start) {
 825                if (state->state & exclusive_bits) {
 826                        *failed_start = start;
 827                        err = -EEXIST;
 828                        goto out;
 829                }
 830
 831                prealloc = alloc_extent_state_atomic(prealloc);
 832                BUG_ON(!prealloc);
 833                err = split_state(tree, state, prealloc, start);
 834                if (err)
 835                        extent_io_tree_panic(tree, err);
 836
 837                prealloc = NULL;
 838                if (err)
 839                        goto out;
 840                if (state->end <= end) {
 841                        set_state_bits(tree, state, &bits);
 842                        cache_state(state, cached_state);
 843                        merge_state(tree, state);
 844                        if (last_end == (u64)-1)
 845                                goto out;
 846                        start = last_end + 1;
 847                        state = next_state(state);
 848                        if (start < end && state && state->start == start &&
 849                            !need_resched())
 850                                goto hit_next;
 851                }
 852                goto search_again;
 853        }
 854        /*
 855         * | ---- desired range ---- |
 856         *     | state | or               | state |
 857         *
 858         * There's a hole, we need to insert something in it and
 859         * ignore the extent we found.
 860         */
 861        if (state->start > start) {
 862                u64 this_end;
 863                if (end < last_start)
 864                        this_end = end;
 865                else
 866                        this_end = last_start - 1;
 867
 868                prealloc = alloc_extent_state_atomic(prealloc);
 869                BUG_ON(!prealloc);
 870
 871                /*
 872                 * Avoid to free 'prealloc' if it can be merged with
 873                 * the later extent.
 874                 */
 875                err = insert_state(tree, prealloc, start, this_end,
 876                                   &bits);
 877                if (err)
 878                        extent_io_tree_panic(tree, err);
 879
 880                cache_state(prealloc, cached_state);
 881                prealloc = NULL;
 882                start = this_end + 1;
 883                goto search_again;
 884        }
 885        /*
 886         * | ---- desired range ---- |
 887         *                        | state |
 888         * We need to split the extent, and set the bit
 889         * on the first half
 890         */
 891        if (state->start <= end && state->end > end) {
 892                if (state->state & exclusive_bits) {
 893                        *failed_start = start;
 894                        err = -EEXIST;
 895                        goto out;
 896                }
 897
 898                prealloc = alloc_extent_state_atomic(prealloc);
 899                BUG_ON(!prealloc);
 900                err = split_state(tree, state, prealloc, end + 1);
 901                if (err)
 902                        extent_io_tree_panic(tree, err);
 903
 904                set_state_bits(tree, prealloc, &bits);
 905                cache_state(prealloc, cached_state);
 906                merge_state(tree, prealloc);
 907                prealloc = NULL;
 908                goto out;
 909        }
 910
 911        goto search_again;
 912
 913out:
 914        spin_unlock(&tree->lock);
 915        if (prealloc)
 916                free_extent_state(prealloc);
 917
 918        return err;
 919
 920search_again:
 921        if (start > end)
 922                goto out;
 923        spin_unlock(&tree->lock);
 924        if (mask & __GFP_WAIT)
 925                cond_resched();
 926        goto again;
 927}
 928
 929int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits,
 930                   u64 *failed_start, struct extent_state **cached_state,
 931                   gfp_t mask)
 932{
 933        return __set_extent_bit(tree, start, end, bits, 0, failed_start,
 934                                cached_state, mask);
 935}
 936
 937
 938/**
 939 * convert_extent_bit - convert all bits in a given range from one bit to
 940 *                      another
 941 * @tree:       the io tree to search
 942 * @start:      the start offset in bytes
 943 * @end:        the end offset in bytes (inclusive)
 944 * @bits:       the bits to set in this range
 945 * @clear_bits: the bits to clear in this range
 946 * @cached_state:       state that we're going to cache
 947 * @mask:       the allocation mask
 948 *
 949 * This will go through and set bits for the given range.  If any states exist
 950 * already in this range they are set with the given bit and cleared of the
 951 * clear_bits.  This is only meant to be used by things that are mergeable, ie
 952 * converting from say DELALLOC to DIRTY.  This is not meant to be used with
 953 * boundary bits like LOCK.
 954 */
 955int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 956                       int bits, int clear_bits,
 957                       struct extent_state **cached_state, gfp_t mask)
 958{
 959        struct extent_state *state;
 960        struct extent_state *prealloc = NULL;
 961        struct rb_node *node;
 962        int err = 0;
 963        u64 last_start;
 964        u64 last_end;
 965
 966again:
 967        if (!prealloc && (mask & __GFP_WAIT)) {
 968                prealloc = alloc_extent_state(mask);
 969                if (!prealloc)
 970                        return -ENOMEM;
 971        }
 972
 973        spin_lock(&tree->lock);
 974        if (cached_state && *cached_state) {
 975                state = *cached_state;
 976                if (state->start <= start && state->end > start &&
 977                    state->tree) {
 978                        node = &state->rb_node;
 979                        goto hit_next;
 980                }
 981        }
 982
 983        /*
 984         * this search will find all the extents that end after
 985         * our range starts.
 986         */
 987        node = tree_search(tree, start);
 988        if (!node) {
 989                prealloc = alloc_extent_state_atomic(prealloc);
 990                if (!prealloc) {
 991                        err = -ENOMEM;
 992                        goto out;
 993                }
 994                err = insert_state(tree, prealloc, start, end, &bits);
 995                prealloc = NULL;
 996                if (err)
 997                        extent_io_tree_panic(tree, err);
 998                goto out;
 999        }
1000        state = rb_entry(node, struct extent_state, rb_node);
1001hit_next:
1002        last_start = state->start;
1003        last_end = state->end;
1004
1005        /*
1006         * | ---- desired range ---- |
1007         * | state |
1008         *
1009         * Just lock what we found and keep going
1010         */
1011        if (state->start == start && state->end <= end) {
1012                set_state_bits(tree, state, &bits);
1013                cache_state(state, cached_state);
1014                state = clear_state_bit(tree, state, &clear_bits, 0);
1015                if (last_end == (u64)-1)
1016                        goto out;
1017                start = last_end + 1;
1018                if (start < end && state && state->start == start &&
1019                    !need_resched())
1020                        goto hit_next;
1021                goto search_again;
1022        }
1023
1024        /*
1025         *     | ---- desired range ---- |
1026         * | state |
1027         *   or
1028         * | ------------- state -------------- |
1029         *
1030         * We need to split the extent we found, and may flip bits on
1031         * second half.
1032         *
1033         * If the extent we found extends past our
1034         * range, we just split and search again.  It'll get split
1035         * again the next time though.
1036         *
1037         * If the extent we found is inside our range, we set the
1038         * desired bit on it.
1039         */
1040        if (state->start < start) {
1041                prealloc = alloc_extent_state_atomic(prealloc);
1042                if (!prealloc) {
1043                        err = -ENOMEM;
1044                        goto out;
1045                }
1046                err = split_state(tree, state, prealloc, start);
1047                if (err)
1048                        extent_io_tree_panic(tree, err);
1049                prealloc = NULL;
1050                if (err)
1051                        goto out;
1052                if (state->end <= end) {
1053                        set_state_bits(tree, state, &bits);
1054                        cache_state(state, cached_state);
1055                        state = clear_state_bit(tree, state, &clear_bits, 0);
1056                        if (last_end == (u64)-1)
1057                                goto out;
1058                        start = last_end + 1;
1059                        if (start < end && state && state->start == start &&
1060                            !need_resched())
1061                                goto hit_next;
1062                }
1063                goto search_again;
1064        }
1065        /*
1066         * | ---- desired range ---- |
1067         *     | state | or               | state |
1068         *
1069         * There's a hole, we need to insert something in it and
1070         * ignore the extent we found.
1071         */
1072        if (state->start > start) {
1073                u64 this_end;
1074                if (end < last_start)
1075                        this_end = end;
1076                else
1077                        this_end = last_start - 1;
1078
1079                prealloc = alloc_extent_state_atomic(prealloc);
1080                if (!prealloc) {
1081                        err = -ENOMEM;
1082                        goto out;
1083                }
1084
1085                /*
1086                 * Avoid to free 'prealloc' if it can be merged with
1087                 * the later extent.
1088                 */
1089                err = insert_state(tree, prealloc, start, this_end,
1090                                   &bits);
1091                if (err)
1092                        extent_io_tree_panic(tree, err);
1093                cache_state(prealloc, cached_state);
1094                prealloc = NULL;
1095                start = this_end + 1;
1096                goto search_again;
1097        }
1098        /*
1099         * | ---- desired range ---- |
1100         *                        | state |
1101         * We need to split the extent, and set the bit
1102         * on the first half
1103         */
1104        if (state->start <= end && state->end > end) {
1105                prealloc = alloc_extent_state_atomic(prealloc);
1106                if (!prealloc) {
1107                        err = -ENOMEM;
1108                        goto out;
1109                }
1110
1111                err = split_state(tree, state, prealloc, end + 1);
1112                if (err)
1113                        extent_io_tree_panic(tree, err);
1114
1115                set_state_bits(tree, prealloc, &bits);
1116                cache_state(prealloc, cached_state);
1117                clear_state_bit(tree, prealloc, &clear_bits, 0);
1118                prealloc = NULL;
1119                goto out;
1120        }
1121
1122        goto search_again;
1123
1124out:
1125        spin_unlock(&tree->lock);
1126        if (prealloc)
1127                free_extent_state(prealloc);
1128
1129        return err;
1130
1131search_again:
1132        if (start > end)
1133                goto out;
1134        spin_unlock(&tree->lock);
1135        if (mask & __GFP_WAIT)
1136                cond_resched();
1137        goto again;
1138}
1139
1140/* wrappers around set/clear extent bit */
1141int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1142                     gfp_t mask)
1143{
1144        return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1145                              NULL, mask);
1146}
1147
1148int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1149                    int bits, gfp_t mask)
1150{
1151        return set_extent_bit(tree, start, end, bits, NULL,
1152                              NULL, mask);
1153}
1154
1155int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1156                      int bits, gfp_t mask)
1157{
1158        return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1159}
1160
1161int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1162                        struct extent_state **cached_state, gfp_t mask)
1163{
1164        return set_extent_bit(tree, start, end,
1165                              EXTENT_DELALLOC | EXTENT_UPTODATE,
1166                              NULL, cached_state, mask);
1167}
1168
1169int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1170                      struct extent_state **cached_state, gfp_t mask)
1171{
1172        return set_extent_bit(tree, start, end,
1173                              EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1174                              NULL, cached_state, mask);
1175}
1176
1177int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1178                       gfp_t mask)
1179{
1180        return clear_extent_bit(tree, start, end,
1181                                EXTENT_DIRTY | EXTENT_DELALLOC |
1182                                EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1183}
1184
1185int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1186                     gfp_t mask)
1187{
1188        return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1189                              NULL, mask);
1190}
1191
1192int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1193                        struct extent_state **cached_state, gfp_t mask)
1194{
1195        return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
1196                              cached_state, mask);
1197}
1198
1199int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1200                          struct extent_state **cached_state, gfp_t mask)
1201{
1202        return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1203                                cached_state, mask);
1204}
1205
1206/*
1207 * either insert or lock state struct between start and end use mask to tell
1208 * us if waiting is desired.
1209 */
1210int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1211                     int bits, struct extent_state **cached_state)
1212{
1213        int err;
1214        u64 failed_start;
1215        while (1) {
1216                err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1217                                       EXTENT_LOCKED, &failed_start,
1218                                       cached_state, GFP_NOFS);
1219                if (err == -EEXIST) {
1220                        wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1221                        start = failed_start;
1222                } else
1223                        break;
1224                WARN_ON(start > end);
1225        }
1226        return err;
1227}
1228
1229int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1230{
1231        return lock_extent_bits(tree, start, end, 0, NULL);
1232}
1233
1234int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1235{
1236        int err;
1237        u64 failed_start;
1238
1239        err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1240                               &failed_start, NULL, GFP_NOFS);
1241        if (err == -EEXIST) {
1242                if (failed_start > start)
1243                        clear_extent_bit(tree, start, failed_start - 1,
1244                                         EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1245                return 0;
1246        }
1247        return 1;
1248}
1249
1250int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1251                         struct extent_state **cached, gfp_t mask)
1252{
1253        return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1254                                mask);
1255}
1256
1257int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1258{
1259        return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1260                                GFP_NOFS);
1261}
1262
1263/*
1264 * helper function to set both pages and extents in the tree writeback
1265 */
1266static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1267{
1268        unsigned long index = start >> PAGE_CACHE_SHIFT;
1269        unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1270        struct page *page;
1271
1272        while (index <= end_index) {
1273                page = find_get_page(tree->mapping, index);
1274                BUG_ON(!page); /* Pages should be in the extent_io_tree */
1275                set_page_writeback(page);
1276                page_cache_release(page);
1277                index++;
1278        }
1279        return 0;
1280}
1281
1282/* find the first state struct with 'bits' set after 'start', and
1283 * return it.  tree->lock must be held.  NULL will returned if
1284 * nothing was found after 'start'
1285 */
1286struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1287                                                 u64 start, int bits)
1288{
1289        struct rb_node *node;
1290        struct extent_state *state;
1291
1292        /*
1293         * this search will find all the extents that end after
1294         * our range starts.
1295         */
1296        node = tree_search(tree, start);
1297        if (!node)
1298                goto out;
1299
1300        while (1) {
1301                state = rb_entry(node, struct extent_state, rb_node);
1302                if (state->end >= start && (state->state & bits))
1303                        return state;
1304
1305                node = rb_next(node);
1306                if (!node)
1307                        break;
1308        }
1309out:
1310        return NULL;
1311}
1312
1313/*
1314 * find the first offset in the io tree with 'bits' set. zero is
1315 * returned if we find something, and *start_ret and *end_ret are
1316 * set to reflect the state struct that was found.
1317 *
1318 * If nothing was found, 1 is returned. If found something, return 0.
1319 */
1320int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1321                          u64 *start_ret, u64 *end_ret, int bits,
1322                          struct extent_state **cached_state)
1323{
1324        struct extent_state *state;
1325        struct rb_node *n;
1326        int ret = 1;
1327
1328        spin_lock(&tree->lock);
1329        if (cached_state && *cached_state) {
1330                state = *cached_state;
1331                if (state->end == start - 1 && state->tree) {
1332                        n = rb_next(&state->rb_node);
1333                        while (n) {
1334                                state = rb_entry(n, struct extent_state,
1335                                                 rb_node);
1336                                if (state->state & bits)
1337                                        goto got_it;
1338                                n = rb_next(n);
1339                        }
1340                        free_extent_state(*cached_state);
1341                        *cached_state = NULL;
1342                        goto out;
1343                }
1344                free_extent_state(*cached_state);
1345                *cached_state = NULL;
1346        }
1347
1348        state = find_first_extent_bit_state(tree, start, bits);
1349got_it:
1350        if (state) {
1351                cache_state(state, cached_state);
1352                *start_ret = state->start;
1353                *end_ret = state->end;
1354                ret = 0;
1355        }
1356out:
1357        spin_unlock(&tree->lock);
1358        return ret;
1359}
1360
1361/*
1362 * find a contiguous range of bytes in the file marked as delalloc, not
1363 * more than 'max_bytes'.  start and end are used to return the range,
1364 *
1365 * 1 is returned if we find something, 0 if nothing was in the tree
1366 */
1367static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1368                                        u64 *start, u64 *end, u64 max_bytes,
1369                                        struct extent_state **cached_state)
1370{
1371        struct rb_node *node;
1372        struct extent_state *state;
1373        u64 cur_start = *start;
1374        u64 found = 0;
1375        u64 total_bytes = 0;
1376
1377        spin_lock(&tree->lock);
1378
1379        /*
1380         * this search will find all the extents that end after
1381         * our range starts.
1382         */
1383        node = tree_search(tree, cur_start);
1384        if (!node) {
1385                if (!found)
1386                        *end = (u64)-1;
1387                goto out;
1388        }
1389
1390        while (1) {
1391                state = rb_entry(node, struct extent_state, rb_node);
1392                if (found && (state->start != cur_start ||
1393                              (state->state & EXTENT_BOUNDARY))) {
1394                        goto out;
1395                }
1396                if (!(state->state & EXTENT_DELALLOC)) {
1397                        if (!found)
1398                                *end = state->end;
1399                        goto out;
1400                }
1401                if (!found) {
1402                        *start = state->start;
1403                        *cached_state = state;
1404                        atomic_inc(&state->refs);
1405                }
1406                found++;
1407                *end = state->end;
1408                cur_start = state->end + 1;
1409                node = rb_next(node);
1410                if (!node)
1411                        break;
1412                total_bytes += state->end - state->start + 1;
1413                if (total_bytes >= max_bytes)
1414                        break;
1415        }
1416out:
1417        spin_unlock(&tree->lock);
1418        return found;
1419}
1420
1421static noinline void __unlock_for_delalloc(struct inode *inode,
1422                                           struct page *locked_page,
1423                                           u64 start, u64 end)
1424{
1425        int ret;
1426        struct page *pages[16];
1427        unsigned long index = start >> PAGE_CACHE_SHIFT;
1428        unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1429        unsigned long nr_pages = end_index - index + 1;
1430        int i;
1431
1432        if (index == locked_page->index && end_index == index)
1433                return;
1434
1435        while (nr_pages > 0) {
1436                ret = find_get_pages_contig(inode->i_mapping, index,
1437                                     min_t(unsigned long, nr_pages,
1438                                     ARRAY_SIZE(pages)), pages);
1439                for (i = 0; i < ret; i++) {
1440                        if (pages[i] != locked_page)
1441                                unlock_page(pages[i]);
1442                        page_cache_release(pages[i]);
1443                }
1444                nr_pages -= ret;
1445                index += ret;
1446                cond_resched();
1447        }
1448}
1449
1450static noinline int lock_delalloc_pages(struct inode *inode,
1451                                        struct page *locked_page,
1452                                        u64 delalloc_start,
1453                                        u64 delalloc_end)
1454{
1455        unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1456        unsigned long start_index = index;
1457        unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1458        unsigned long pages_locked = 0;
1459        struct page *pages[16];
1460        unsigned long nrpages;
1461        int ret;
1462        int i;
1463
1464        /* the caller is responsible for locking the start index */
1465        if (index == locked_page->index && index == end_index)
1466                return 0;
1467
1468        /* skip the page at the start index */
1469        nrpages = end_index - index + 1;
1470        while (nrpages > 0) {
1471                ret = find_get_pages_contig(inode->i_mapping, index,
1472                                     min_t(unsigned long,
1473                                     nrpages, ARRAY_SIZE(pages)), pages);
1474                if (ret == 0) {
1475                        ret = -EAGAIN;
1476                        goto done;
1477                }
1478                /* now we have an array of pages, lock them all */
1479                for (i = 0; i < ret; i++) {
1480                        /*
1481                         * the caller is taking responsibility for
1482                         * locked_page
1483                         */
1484                        if (pages[i] != locked_page) {
1485                                lock_page(pages[i]);
1486                                if (!PageDirty(pages[i]) ||
1487                                    pages[i]->mapping != inode->i_mapping) {
1488                                        ret = -EAGAIN;
1489                                        unlock_page(pages[i]);
1490                                        page_cache_release(pages[i]);
1491                                        goto done;
1492                                }
1493                        }
1494                        page_cache_release(pages[i]);
1495                        pages_locked++;
1496                }
1497                nrpages -= ret;
1498                index += ret;
1499                cond_resched();
1500        }
1501        ret = 0;
1502done:
1503        if (ret && pages_locked) {
1504                __unlock_for_delalloc(inode, locked_page,
1505                              delalloc_start,
1506                              ((u64)(start_index + pages_locked - 1)) <<
1507                              PAGE_CACHE_SHIFT);
1508        }
1509        return ret;
1510}
1511
1512/*
1513 * find a contiguous range of bytes in the file marked as delalloc, not
1514 * more than 'max_bytes'.  start and end are used to return the range,
1515 *
1516 * 1 is returned if we find something, 0 if nothing was in the tree
1517 */
1518static noinline u64 find_lock_delalloc_range(struct inode *inode,
1519                                             struct extent_io_tree *tree,
1520                                             struct page *locked_page,
1521                                             u64 *start, u64 *end,
1522                                             u64 max_bytes)
1523{
1524        u64 delalloc_start;
1525        u64 delalloc_end;
1526        u64 found;
1527        struct extent_state *cached_state = NULL;
1528        int ret;
1529        int loops = 0;
1530
1531again:
1532        /* step one, find a bunch of delalloc bytes starting at start */
1533        delalloc_start = *start;
1534        delalloc_end = 0;
1535        found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1536                                    max_bytes, &cached_state);
1537        if (!found || delalloc_end <= *start) {
1538                *start = delalloc_start;
1539                *end = delalloc_end;
1540                free_extent_state(cached_state);
1541                return found;
1542        }
1543
1544        /*
1545         * start comes from the offset of locked_page.  We have to lock
1546         * pages in order, so we can't process delalloc bytes before
1547         * locked_page
1548         */
1549        if (delalloc_start < *start)
1550                delalloc_start = *start;
1551
1552        /*
1553         * make sure to limit the number of pages we try to lock down
1554         * if we're looping.
1555         */
1556        if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1557                delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1558
1559        /* step two, lock all the pages after the page that has start */
1560        ret = lock_delalloc_pages(inode, locked_page,
1561                                  delalloc_start, delalloc_end);
1562        if (ret == -EAGAIN) {
1563                /* some of the pages are gone, lets avoid looping by
1564                 * shortening the size of the delalloc range we're searching
1565                 */
1566                free_extent_state(cached_state);
1567                if (!loops) {
1568                        unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1569                        max_bytes = PAGE_CACHE_SIZE - offset;
1570                        loops = 1;
1571                        goto again;
1572                } else {
1573                        found = 0;
1574                        goto out_failed;
1575                }
1576        }
1577        BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1578
1579        /* step three, lock the state bits for the whole range */
1580        lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1581
1582        /* then test to make sure it is all still delalloc */
1583        ret = test_range_bit(tree, delalloc_start, delalloc_end,
1584                             EXTENT_DELALLOC, 1, cached_state);
1585        if (!ret) {
1586                unlock_extent_cached(tree, delalloc_start, delalloc_end,
1587                                     &cached_state, GFP_NOFS);
1588                __unlock_for_delalloc(inode, locked_page,
1589                              delalloc_start, delalloc_end);
1590                cond_resched();
1591                goto again;
1592        }
1593        free_extent_state(cached_state);
1594        *start = delalloc_start;
1595        *end = delalloc_end;
1596out_failed:
1597        return found;
1598}
1599
1600int extent_clear_unlock_delalloc(struct inode *inode,
1601                                struct extent_io_tree *tree,
1602                                u64 start, u64 end, struct page *locked_page,
1603                                unsigned long op)
1604{
1605        int ret;
1606        struct page *pages[16];
1607        unsigned long index = start >> PAGE_CACHE_SHIFT;
1608        unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1609        unsigned long nr_pages = end_index - index + 1;
1610        int i;
1611        int clear_bits = 0;
1612
1613        if (op & EXTENT_CLEAR_UNLOCK)
1614                clear_bits |= EXTENT_LOCKED;
1615        if (op & EXTENT_CLEAR_DIRTY)
1616                clear_bits |= EXTENT_DIRTY;
1617
1618        if (op & EXTENT_CLEAR_DELALLOC)
1619                clear_bits |= EXTENT_DELALLOC;
1620
1621        clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1622        if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1623                    EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1624                    EXTENT_SET_PRIVATE2)))
1625                return 0;
1626
1627        while (nr_pages > 0) {
1628                ret = find_get_pages_contig(inode->i_mapping, index,
1629                                     min_t(unsigned long,
1630                                     nr_pages, ARRAY_SIZE(pages)), pages);
1631                for (i = 0; i < ret; i++) {
1632
1633                        if (op & EXTENT_SET_PRIVATE2)
1634                                SetPagePrivate2(pages[i]);
1635
1636                        if (pages[i] == locked_page) {
1637                                page_cache_release(pages[i]);
1638                                continue;
1639                        }
1640                        if (op & EXTENT_CLEAR_DIRTY)
1641                                clear_page_dirty_for_io(pages[i]);
1642                        if (op & EXTENT_SET_WRITEBACK)
1643                                set_page_writeback(pages[i]);
1644                        if (op & EXTENT_END_WRITEBACK)
1645                                end_page_writeback(pages[i]);
1646                        if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1647                                unlock_page(pages[i]);
1648                        page_cache_release(pages[i]);
1649                }
1650                nr_pages -= ret;
1651                index += ret;
1652                cond_resched();
1653        }
1654        return 0;
1655}
1656
1657/*
1658 * count the number of bytes in the tree that have a given bit(s)
1659 * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1660 * cached.  The total number found is returned.
1661 */
1662u64 count_range_bits(struct extent_io_tree *tree,
1663                     u64 *start, u64 search_end, u64 max_bytes,
1664                     unsigned long bits, int contig)
1665{
1666        struct rb_node *node;
1667        struct extent_state *state;
1668        u64 cur_start = *start;
1669        u64 total_bytes = 0;
1670        u64 last = 0;
1671        int found = 0;
1672
1673        if (search_end <= cur_start) {
1674                WARN_ON(1);
1675                return 0;
1676        }
1677
1678        spin_lock(&tree->lock);
1679        if (cur_start == 0 && bits == EXTENT_DIRTY) {
1680                total_bytes = tree->dirty_bytes;
1681                goto out;
1682        }
1683        /*
1684         * this search will find all the extents that end after
1685         * our range starts.
1686         */
1687        node = tree_search(tree, cur_start);
1688        if (!node)
1689                goto out;
1690
1691        while (1) {
1692                state = rb_entry(node, struct extent_state, rb_node);
1693                if (state->start > search_end)
1694                        break;
1695                if (contig && found && state->start > last + 1)
1696                        break;
1697                if (state->end >= cur_start && (state->state & bits) == bits) {
1698                        total_bytes += min(search_end, state->end) + 1 -
1699                                       max(cur_start, state->start);
1700                        if (total_bytes >= max_bytes)
1701                                break;
1702                        if (!found) {
1703                                *start = max(cur_start, state->start);
1704                                found = 1;
1705                        }
1706                        last = state->end;
1707                } else if (contig && found) {
1708                        break;
1709                }
1710                node = rb_next(node);
1711                if (!node)
1712                        break;
1713        }
1714out:
1715        spin_unlock(&tree->lock);
1716        return total_bytes;
1717}
1718
1719/*
1720 * set the private field for a given byte offset in the tree.  If there isn't
1721 * an extent_state there already, this does nothing.
1722 */
1723int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1724{
1725        struct rb_node *node;
1726        struct extent_state *state;
1727        int ret = 0;
1728
1729        spin_lock(&tree->lock);
1730        /*
1731         * this search will find all the extents that end after
1732         * our range starts.
1733         */
1734        node = tree_search(tree, start);
1735        if (!node) {
1736                ret = -ENOENT;
1737                goto out;
1738        }
1739        state = rb_entry(node, struct extent_state, rb_node);
1740        if (state->start != start) {
1741                ret = -ENOENT;
1742                goto out;
1743        }
1744        state->private = private;
1745out:
1746        spin_unlock(&tree->lock);
1747        return ret;
1748}
1749
1750int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1751{
1752        struct rb_node *node;
1753        struct extent_state *state;
1754        int ret = 0;
1755
1756        spin_lock(&tree->lock);
1757        /*
1758         * this search will find all the extents that end after
1759         * our range starts.
1760         */
1761        node = tree_search(tree, start);
1762        if (!node) {
1763                ret = -ENOENT;
1764                goto out;
1765        }
1766        state = rb_entry(node, struct extent_state, rb_node);
1767        if (state->start != start) {
1768                ret = -ENOENT;
1769                goto out;
1770        }
1771        *private = state->private;
1772out:
1773        spin_unlock(&tree->lock);
1774        return ret;
1775}
1776
1777/*
1778 * searches a range in the state tree for a given mask.
1779 * If 'filled' == 1, this returns 1 only if every extent in the tree
1780 * has the bits set.  Otherwise, 1 is returned if any bit in the
1781 * range is found set.
1782 */
1783int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1784                   int bits, int filled, struct extent_state *cached)
1785{
1786        struct extent_state *state = NULL;
1787        struct rb_node *node;
1788        int bitset = 0;
1789
1790        spin_lock(&tree->lock);
1791        if (cached && cached->tree && cached->start <= start &&
1792            cached->end > start)
1793                node = &cached->rb_node;
1794        else
1795                node = tree_search(tree, start);
1796        while (node && start <= end) {
1797                state = rb_entry(node, struct extent_state, rb_node);
1798
1799                if (filled && state->start > start) {
1800                        bitset = 0;
1801                        break;
1802                }
1803
1804                if (state->start > end)
1805                        break;
1806
1807                if (state->state & bits) {
1808                        bitset = 1;
1809                        if (!filled)
1810                                break;
1811                } else if (filled) {
1812                        bitset = 0;
1813                        break;
1814                }
1815
1816                if (state->end == (u64)-1)
1817                        break;
1818
1819                start = state->end + 1;
1820                if (start > end)
1821                        break;
1822                node = rb_next(node);
1823                if (!node) {
1824                        if (filled)
1825                                bitset = 0;
1826                        break;
1827                }
1828        }
1829        spin_unlock(&tree->lock);
1830        return bitset;
1831}
1832
1833/*
1834 * helper function to set a given page up to date if all the
1835 * extents in the tree for that page are up to date
1836 */
1837static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1838{
1839        u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1840        u64 end = start + PAGE_CACHE_SIZE - 1;
1841        if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1842                SetPageUptodate(page);
1843}
1844
1845/*
1846 * helper function to unlock a page if all the extents in the tree
1847 * for that page are unlocked
1848 */
1849static void check_page_locked(struct extent_io_tree *tree, struct page *page)
1850{
1851        u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1852        u64 end = start + PAGE_CACHE_SIZE - 1;
1853        if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1854                unlock_page(page);
1855}
1856
1857/*
1858 * helper function to end page writeback if all the extents
1859 * in the tree for that page are done with writeback
1860 */
1861static void check_page_writeback(struct extent_io_tree *tree,
1862                                 struct page *page)
1863{
1864        end_page_writeback(page);
1865}
1866
1867/*
1868 * When IO fails, either with EIO or csum verification fails, we
1869 * try other mirrors that might have a good copy of the data.  This
1870 * io_failure_record is used to record state as we go through all the
1871 * mirrors.  If another mirror has good data, the page is set up to date
1872 * and things continue.  If a good mirror can't be found, the original
1873 * bio end_io callback is called to indicate things have failed.
1874 */
1875struct io_failure_record {
1876        struct page *page;
1877        u64 start;
1878        u64 len;
1879        u64 logical;
1880        unsigned long bio_flags;
1881        int this_mirror;
1882        int failed_mirror;
1883        int in_validation;
1884};
1885
1886static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1887                                int did_repair)
1888{
1889        int ret;
1890        int err = 0;
1891        struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1892
1893        set_state_private(failure_tree, rec->start, 0);
1894        ret = clear_extent_bits(failure_tree, rec->start,
1895                                rec->start + rec->len - 1,
1896                                EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1897        if (ret)
1898                err = ret;
1899
1900        if (did_repair) {
1901                ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1902                                        rec->start + rec->len - 1,
1903                                        EXTENT_DAMAGED, GFP_NOFS);
1904                if (ret && !err)
1905                        err = ret;
1906        }
1907
1908        kfree(rec);
1909        return err;
1910}
1911
1912static void repair_io_failure_callback(struct bio *bio, int err)
1913{
1914        complete(bio->bi_private);
1915}
1916
1917/*
1918 * this bypasses the standard btrfs submit functions deliberately, as
1919 * the standard behavior is to write all copies in a raid setup. here we only
1920 * want to write the one bad copy. so we do the mapping for ourselves and issue
1921 * submit_bio directly.
1922 * to avoid any synchonization issues, wait for the data after writing, which
1923 * actually prevents the read that triggered the error from finishing.
1924 * currently, there can be no more than two copies of every data bit. thus,
1925 * exactly one rewrite is required.
1926 */
1927int repair_io_failure(struct btrfs_mapping_tree *map_tree, u64 start,
1928                        u64 length, u64 logical, struct page *page,
1929                        int mirror_num)
1930{
1931        struct bio *bio;
1932        struct btrfs_device *dev;
1933        DECLARE_COMPLETION_ONSTACK(compl);
1934        u64 map_length = 0;
1935        u64 sector;
1936        struct btrfs_bio *bbio = NULL;
1937        int ret;
1938
1939        BUG_ON(!mirror_num);
1940
1941        bio = bio_alloc(GFP_NOFS, 1);
1942        if (!bio)
1943                return -EIO;
1944        bio->bi_private = &compl;
1945        bio->bi_end_io = repair_io_failure_callback;
1946        bio->bi_size = 0;
1947        map_length = length;
1948
1949        ret = btrfs_map_block(map_tree, WRITE, logical,
1950                              &map_length, &bbio, mirror_num);
1951        if (ret) {
1952                bio_put(bio);
1953                return -EIO;
1954        }
1955        BUG_ON(mirror_num != bbio->mirror_num);
1956        sector = bbio->stripes[mirror_num-1].physical >> 9;
1957        bio->bi_sector = sector;
1958        dev = bbio->stripes[mirror_num-1].dev;
1959        kfree(bbio);
1960        if (!dev || !dev->bdev || !dev->writeable) {
1961                bio_put(bio);
1962                return -EIO;
1963        }
1964        bio->bi_bdev = dev->bdev;
1965        bio_add_page(bio, page, length, start-page_offset(page));
1966        btrfsic_submit_bio(WRITE_SYNC, bio);
1967        wait_for_completion(&compl);
1968
1969        if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
1970                /* try to remap that extent elsewhere? */
1971                bio_put(bio);
1972                btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
1973                return -EIO;
1974        }
1975
1976        printk_ratelimited_in_rcu(KERN_INFO "btrfs read error corrected: ino %lu off %llu "
1977                      "(dev %s sector %llu)\n", page->mapping->host->i_ino,
1978                      start, rcu_str_deref(dev->name), sector);
1979
1980        bio_put(bio);
1981        return 0;
1982}
1983
1984int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
1985                         int mirror_num)
1986{
1987        struct btrfs_mapping_tree *map_tree = &root->fs_info->mapping_tree;
1988        u64 start = eb->start;
1989        unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
1990        int ret = 0;
1991
1992        for (i = 0; i < num_pages; i++) {
1993                struct page *p = extent_buffer_page(eb, i);
1994                ret = repair_io_failure(map_tree, start, PAGE_CACHE_SIZE,
1995                                        start, p, mirror_num);
1996                if (ret)
1997                        break;
1998                start += PAGE_CACHE_SIZE;
1999        }
2000
2001        return ret;
2002}
2003
2004/*
2005 * each time an IO finishes, we do a fast check in the IO failure tree
2006 * to see if we need to process or clean up an io_failure_record
2007 */
2008static int clean_io_failure(u64 start, struct page *page)
2009{
2010        u64 private;
2011        u64 private_failure;
2012        struct io_failure_record *failrec;
2013        struct btrfs_mapping_tree *map_tree;
2014        struct extent_state *state;
2015        int num_copies;
2016        int did_repair = 0;
2017        int ret;
2018        struct inode *inode = page->mapping->host;
2019
2020        private = 0;
2021        ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2022                                (u64)-1, 1, EXTENT_DIRTY, 0);
2023        if (!ret)
2024                return 0;
2025
2026        ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2027                                &private_failure);
2028        if (ret)
2029                return 0;
2030
2031        failrec = (struct io_failure_record *)(unsigned long) private_failure;
2032        BUG_ON(!failrec->this_mirror);
2033
2034        if (failrec->in_validation) {
2035                /* there was no real error, just free the record */
2036                pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2037                         failrec->start);
2038                did_repair = 1;
2039                goto out;
2040        }
2041
2042        spin_lock(&BTRFS_I(inode)->io_tree.lock);
2043        state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2044                                            failrec->start,
2045                                            EXTENT_LOCKED);
2046        spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2047
2048        if (state && state->start == failrec->start) {
2049                map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
2050                num_copies = btrfs_num_copies(map_tree, failrec->logical,
2051                                                failrec->len);
2052                if (num_copies > 1)  {
2053                        ret = repair_io_failure(map_tree, start, failrec->len,
2054                                                failrec->logical, page,
2055                                                failrec->failed_mirror);
2056                        did_repair = !ret;
2057                }
2058        }
2059
2060out:
2061        if (!ret)
2062                ret = free_io_failure(inode, failrec, did_repair);
2063
2064        return ret;
2065}
2066
2067/*
2068 * this is a generic handler for readpage errors (default
2069 * readpage_io_failed_hook). if other copies exist, read those and write back
2070 * good data to the failed position. does not investigate in remapping the
2071 * failed extent elsewhere, hoping the device will be smart enough to do this as
2072 * needed
2073 */
2074
2075static int bio_readpage_error(struct bio *failed_bio, struct page *page,
2076                                u64 start, u64 end, int failed_mirror,
2077                                struct extent_state *state)
2078{
2079        struct io_failure_record *failrec = NULL;
2080        u64 private;
2081        struct extent_map *em;
2082        struct inode *inode = page->mapping->host;
2083        struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2084        struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2085        struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2086        struct bio *bio;
2087        int num_copies;
2088        int ret;
2089        int read_mode;
2090        u64 logical;
2091
2092        BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2093
2094        ret = get_state_private(failure_tree, start, &private);
2095        if (ret) {
2096                failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2097                if (!failrec)
2098                        return -ENOMEM;
2099                failrec->start = start;
2100                failrec->len = end - start + 1;
2101                failrec->this_mirror = 0;
2102                failrec->bio_flags = 0;
2103                failrec->in_validation = 0;
2104
2105                read_lock(&em_tree->lock);
2106                em = lookup_extent_mapping(em_tree, start, failrec->len);
2107                if (!em) {
2108                        read_unlock(&em_tree->lock);
2109                        kfree(failrec);
2110                        return -EIO;
2111                }
2112
2113                if (em->start > start || em->start + em->len < start) {
2114                        free_extent_map(em);
2115                        em = NULL;
2116                }
2117                read_unlock(&em_tree->lock);
2118
2119                if (!em) {
2120                        kfree(failrec);
2121                        return -EIO;
2122                }
2123                logical = start - em->start;
2124                logical = em->block_start + logical;
2125                if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2126                        logical = em->block_start;
2127                        failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2128                        extent_set_compress_type(&failrec->bio_flags,
2129                                                 em->compress_type);
2130                }
2131                pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2132                         "len=%llu\n", logical, start, failrec->len);
2133                failrec->logical = logical;
2134                free_extent_map(em);
2135
2136                /* set the bits in the private failure tree */
2137                ret = set_extent_bits(failure_tree, start, end,
2138                                        EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2139                if (ret >= 0)
2140                        ret = set_state_private(failure_tree, start,
2141                                                (u64)(unsigned long)failrec);
2142                /* set the bits in the inode's tree */
2143                if (ret >= 0)
2144                        ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2145                                                GFP_NOFS);
2146                if (ret < 0) {
2147                        kfree(failrec);
2148                        return ret;
2149                }
2150        } else {
2151                failrec = (struct io_failure_record *)(unsigned long)private;
2152                pr_debug("bio_readpage_error: (found) logical=%llu, "
2153                         "start=%llu, len=%llu, validation=%d\n",
2154                         failrec->logical, failrec->start, failrec->len,
2155                         failrec->in_validation);
2156                /*
2157                 * when data can be on disk more than twice, add to failrec here
2158                 * (e.g. with a list for failed_mirror) to make
2159                 * clean_io_failure() clean all those errors at once.
2160                 */
2161        }
2162        num_copies = btrfs_num_copies(
2163                              &BTRFS_I(inode)->root->fs_info->mapping_tree,
2164                              failrec->logical, failrec->len);
2165        if (num_copies == 1) {
2166                /*
2167                 * we only have a single copy of the data, so don't bother with
2168                 * all the retry and error correction code that follows. no
2169                 * matter what the error is, it is very likely to persist.
2170                 */
2171                pr_debug("bio_readpage_error: cannot repair, num_copies == 1. "
2172                         "state=%p, num_copies=%d, next_mirror %d, "
2173                         "failed_mirror %d\n", state, num_copies,
2174                         failrec->this_mirror, failed_mirror);
2175                free_io_failure(inode, failrec, 0);
2176                return -EIO;
2177        }
2178
2179        if (!state) {
2180                spin_lock(&tree->lock);
2181                state = find_first_extent_bit_state(tree, failrec->start,
2182                                                    EXTENT_LOCKED);
2183                if (state && state->start != failrec->start)
2184                        state = NULL;
2185                spin_unlock(&tree->lock);
2186        }
2187
2188        /*
2189         * there are two premises:
2190         *      a) deliver good data to the caller
2191         *      b) correct the bad sectors on disk
2192         */
2193        if (failed_bio->bi_vcnt > 1) {
2194                /*
2195                 * to fulfill b), we need to know the exact failing sectors, as
2196                 * we don't want to rewrite any more than the failed ones. thus,
2197                 * we need separate read requests for the failed bio
2198                 *
2199                 * if the following BUG_ON triggers, our validation request got
2200                 * merged. we need separate requests for our algorithm to work.
2201                 */
2202                BUG_ON(failrec->in_validation);
2203                failrec->in_validation = 1;
2204                failrec->this_mirror = failed_mirror;
2205                read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2206        } else {
2207                /*
2208                 * we're ready to fulfill a) and b) alongside. get a good copy
2209                 * of the failed sector and if we succeed, we have setup
2210                 * everything for repair_io_failure to do the rest for us.
2211                 */
2212                if (failrec->in_validation) {
2213                        BUG_ON(failrec->this_mirror != failed_mirror);
2214                        failrec->in_validation = 0;
2215                        failrec->this_mirror = 0;
2216                }
2217                failrec->failed_mirror = failed_mirror;
2218                failrec->this_mirror++;
2219                if (failrec->this_mirror == failed_mirror)
2220                        failrec->this_mirror++;
2221                read_mode = READ_SYNC;
2222        }
2223
2224        if (!state || failrec->this_mirror > num_copies) {
2225                pr_debug("bio_readpage_error: (fail) state=%p, num_copies=%d, "
2226                         "next_mirror %d, failed_mirror %d\n", state,
2227                         num_copies, failrec->this_mirror, failed_mirror);
2228                free_io_failure(inode, failrec, 0);
2229                return -EIO;
2230        }
2231
2232        bio = bio_alloc(GFP_NOFS, 1);
2233        if (!bio) {
2234                free_io_failure(inode, failrec, 0);
2235                return -EIO;
2236        }
2237        bio->bi_private = state;
2238        bio->bi_end_io = failed_bio->bi_end_io;
2239        bio->bi_sector = failrec->logical >> 9;
2240        bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2241        bio->bi_size = 0;
2242
2243        bio_add_page(bio, page, failrec->len, start - page_offset(page));
2244
2245        pr_debug("bio_readpage_error: submitting new read[%#x] to "
2246                 "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2247                 failrec->this_mirror, num_copies, failrec->in_validation);
2248
2249        ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2250                                         failrec->this_mirror,
2251                                         failrec->bio_flags, 0);
2252        return ret;
2253}
2254
2255/* lots and lots of room for performance fixes in the end_bio funcs */
2256
2257int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2258{
2259        int uptodate = (err == 0);
2260        struct extent_io_tree *tree;
2261        int ret;
2262
2263        tree = &BTRFS_I(page->mapping->host)->io_tree;
2264
2265        if (tree->ops && tree->ops->writepage_end_io_hook) {
2266                ret = tree->ops->writepage_end_io_hook(page, start,
2267                                               end, NULL, uptodate);
2268                if (ret)
2269                        uptodate = 0;
2270        }
2271
2272        if (!uptodate) {
2273                ClearPageUptodate(page);
2274                SetPageError(page);
2275        }
2276        return 0;
2277}
2278
2279/*
2280 * after a writepage IO is done, we need to:
2281 * clear the uptodate bits on error
2282 * clear the writeback bits in the extent tree for this IO
2283 * end_page_writeback if the page has no more pending IO
2284 *
2285 * Scheduling is not allowed, so the extent state tree is expected
2286 * to have one and only one object corresponding to this IO.
2287 */
2288static void end_bio_extent_writepage(struct bio *bio, int err)
2289{
2290        struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2291        struct extent_io_tree *tree;
2292        u64 start;
2293        u64 end;
2294        int whole_page;
2295
2296        do {
2297                struct page *page = bvec->bv_page;
2298                tree = &BTRFS_I(page->mapping->host)->io_tree;
2299
2300                start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2301                         bvec->bv_offset;
2302                end = start + bvec->bv_len - 1;
2303
2304                if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2305                        whole_page = 1;
2306                else
2307                        whole_page = 0;
2308
2309                if (--bvec >= bio->bi_io_vec)
2310                        prefetchw(&bvec->bv_page->flags);
2311
2312                if (end_extent_writepage(page, err, start, end))
2313                        continue;
2314
2315                if (whole_page)
2316                        end_page_writeback(page);
2317                else
2318                        check_page_writeback(tree, page);
2319        } while (bvec >= bio->bi_io_vec);
2320
2321        bio_put(bio);
2322}
2323
2324/*
2325 * after a readpage IO is done, we need to:
2326 * clear the uptodate bits on error
2327 * set the uptodate bits if things worked
2328 * set the page up to date if all extents in the tree are uptodate
2329 * clear the lock bit in the extent tree
2330 * unlock the page if there are no other extents locked for it
2331 *
2332 * Scheduling is not allowed, so the extent state tree is expected
2333 * to have one and only one object corresponding to this IO.
2334 */
2335static void end_bio_extent_readpage(struct bio *bio, int err)
2336{
2337        int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2338        struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
2339        struct bio_vec *bvec = bio->bi_io_vec;
2340        struct extent_io_tree *tree;
2341        u64 start;
2342        u64 end;
2343        int whole_page;
2344        int mirror;
2345        int ret;
2346
2347        if (err)
2348                uptodate = 0;
2349
2350        do {
2351                struct page *page = bvec->bv_page;
2352                struct extent_state *cached = NULL;
2353                struct extent_state *state;
2354
2355                pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2356                         "mirror=%ld\n", (u64)bio->bi_sector, err,
2357                         (long int)bio->bi_bdev);
2358                tree = &BTRFS_I(page->mapping->host)->io_tree;
2359
2360                start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2361                        bvec->bv_offset;
2362                end = start + bvec->bv_len - 1;
2363
2364                if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2365                        whole_page = 1;
2366                else
2367                        whole_page = 0;
2368
2369                if (++bvec <= bvec_end)
2370                        prefetchw(&bvec->bv_page->flags);
2371
2372                spin_lock(&tree->lock);
2373                state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
2374                if (state && state->start == start) {
2375                        /*
2376                         * take a reference on the state, unlock will drop
2377                         * the ref
2378                         */
2379                        cache_state(state, &cached);
2380                }
2381                spin_unlock(&tree->lock);
2382
2383                mirror = (int)(unsigned long)bio->bi_bdev;
2384                if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
2385                        ret = tree->ops->readpage_end_io_hook(page, start, end,
2386                                                              state, mirror);
2387                        if (ret)
2388                                uptodate = 0;
2389                        else
2390                                clean_io_failure(start, page);
2391                }
2392
2393                if (!uptodate && tree->ops && tree->ops->readpage_io_failed_hook) {
2394                        ret = tree->ops->readpage_io_failed_hook(page, mirror);
2395                        if (!ret && !err &&
2396                            test_bit(BIO_UPTODATE, &bio->bi_flags))
2397                                uptodate = 1;
2398                } else if (!uptodate) {
2399                        /*
2400                         * The generic bio_readpage_error handles errors the
2401                         * following way: If possible, new read requests are
2402                         * created and submitted and will end up in
2403                         * end_bio_extent_readpage as well (if we're lucky, not
2404                         * in the !uptodate case). In that case it returns 0 and
2405                         * we just go on with the next page in our bio. If it
2406                         * can't handle the error it will return -EIO and we
2407                         * remain responsible for that page.
2408                         */
2409                        ret = bio_readpage_error(bio, page, start, end, mirror, NULL);
2410                        if (ret == 0) {
2411                                uptodate =
2412                                        test_bit(BIO_UPTODATE, &bio->bi_flags);
2413                                if (err)
2414                                        uptodate = 0;
2415                                uncache_state(&cached);
2416                                continue;
2417                        }
2418                }
2419
2420                if (uptodate && tree->track_uptodate) {
2421                        set_extent_uptodate(tree, start, end, &cached,
2422                                            GFP_ATOMIC);
2423                }
2424                unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2425
2426                if (whole_page) {
2427                        if (uptodate) {
2428                                SetPageUptodate(page);
2429                        } else {
2430                                ClearPageUptodate(page);
2431                                SetPageError(page);
2432                        }
2433                        unlock_page(page);
2434                } else {
2435                        if (uptodate) {
2436                                check_page_uptodate(tree, page);
2437                        } else {
2438                                ClearPageUptodate(page);
2439                                SetPageError(page);
2440                        }
2441                        check_page_locked(tree, page);
2442                }
2443        } while (bvec <= bvec_end);
2444
2445        bio_put(bio);
2446}
2447
2448struct bio *
2449btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2450                gfp_t gfp_flags)
2451{
2452        struct bio *bio;
2453
2454        bio = bio_alloc(gfp_flags, nr_vecs);
2455
2456        if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2457                while (!bio && (nr_vecs /= 2))
2458                        bio = bio_alloc(gfp_flags, nr_vecs);
2459        }
2460
2461        if (bio) {
2462                bio->bi_size = 0;
2463                bio->bi_bdev = bdev;
2464                bio->bi_sector = first_sector;
2465        }
2466        return bio;
2467}
2468
2469/*
2470 * Since writes are async, they will only return -ENOMEM.
2471 * Reads can return the full range of I/O error conditions.
2472 */
2473static int __must_check submit_one_bio(int rw, struct bio *bio,
2474                                       int mirror_num, unsigned long bio_flags)
2475{
2476        int ret = 0;
2477        struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2478        struct page *page = bvec->bv_page;
2479        struct extent_io_tree *tree = bio->bi_private;
2480        u64 start;
2481
2482        start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
2483
2484        bio->bi_private = NULL;
2485
2486        bio_get(bio);
2487
2488        if (tree->ops && tree->ops->submit_bio_hook)
2489                ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2490                                           mirror_num, bio_flags, start);
2491        else
2492                btrfsic_submit_bio(rw, bio);
2493
2494        if (bio_flagged(bio, BIO_EOPNOTSUPP))
2495                ret = -EOPNOTSUPP;
2496        bio_put(bio);
2497        return ret;
2498}
2499
2500static int merge_bio(struct extent_io_tree *tree, struct page *page,
2501                     unsigned long offset, size_t size, struct bio *bio,
2502                     unsigned long bio_flags)
2503{
2504        int ret = 0;
2505        if (tree->ops && tree->ops->merge_bio_hook)
2506                ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2507                                                bio_flags);
2508        BUG_ON(ret < 0);
2509        return ret;
2510
2511}
2512
2513static int submit_extent_page(int rw, struct extent_io_tree *tree,
2514                              struct page *page, sector_t sector,
2515                              size_t size, unsigned long offset,
2516                              struct block_device *bdev,
2517                              struct bio **bio_ret,
2518                              unsigned long max_pages,
2519                              bio_end_io_t end_io_func,
2520                              int mirror_num,
2521                              unsigned long prev_bio_flags,
2522                              unsigned long bio_flags)
2523{
2524        int ret = 0;
2525        struct bio *bio;
2526        int nr;
2527        int contig = 0;
2528        int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2529        int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2530        size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2531
2532        if (bio_ret && *bio_ret) {
2533                bio = *bio_ret;
2534                if (old_compressed)
2535                        contig = bio->bi_sector == sector;
2536                else
2537                        contig = bio->bi_sector + (bio->bi_size >> 9) ==
2538                                sector;
2539
2540                if (prev_bio_flags != bio_flags || !contig ||
2541                    merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2542                    bio_add_page(bio, page, page_size, offset) < page_size) {
2543                        ret = submit_one_bio(rw, bio, mirror_num,
2544                                             prev_bio_flags);
2545                        if (ret < 0)
2546                                return ret;
2547                        bio = NULL;
2548                } else {
2549                        return 0;
2550                }
2551        }
2552        if (this_compressed)
2553                nr = BIO_MAX_PAGES;
2554        else
2555                nr = bio_get_nr_vecs(bdev);
2556
2557        bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2558        if (!bio)
2559                return -ENOMEM;
2560
2561        bio_add_page(bio, page, page_size, offset);
2562        bio->bi_end_io = end_io_func;
2563        bio->bi_private = tree;
2564
2565        if (bio_ret)
2566                *bio_ret = bio;
2567        else
2568                ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2569
2570        return ret;
2571}
2572
2573void attach_extent_buffer_page(struct extent_buffer *eb, struct page *page)
2574{
2575        if (!PagePrivate(page)) {
2576                SetPagePrivate(page);
2577                page_cache_get(page);
2578                set_page_private(page, (unsigned long)eb);
2579        } else {
2580                WARN_ON(page->private != (unsigned long)eb);
2581        }
2582}
2583
2584void set_page_extent_mapped(struct page *page)
2585{
2586        if (!PagePrivate(page)) {
2587                SetPagePrivate(page);
2588                page_cache_get(page);
2589                set_page_private(page, EXTENT_PAGE_PRIVATE);
2590        }
2591}
2592
2593/*
2594 * basic readpage implementation.  Locked extent state structs are inserted
2595 * into the tree that are removed when the IO is done (by the end_io
2596 * handlers)
2597 * XXX JDM: This needs looking at to ensure proper page locking
2598 */
2599static int __extent_read_full_page(struct extent_io_tree *tree,
2600                                   struct page *page,
2601                                   get_extent_t *get_extent,
2602                                   struct bio **bio, int mirror_num,
2603                                   unsigned long *bio_flags)
2604{
2605        struct inode *inode = page->mapping->host;
2606        u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2607        u64 page_end = start + PAGE_CACHE_SIZE - 1;
2608        u64 end;
2609        u64 cur = start;
2610        u64 extent_offset;
2611        u64 last_byte = i_size_read(inode);
2612        u64 block_start;
2613        u64 cur_end;
2614        sector_t sector;
2615        struct extent_map *em;
2616        struct block_device *bdev;
2617        struct btrfs_ordered_extent *ordered;
2618        int ret;
2619        int nr = 0;
2620        size_t pg_offset = 0;
2621        size_t iosize;
2622        size_t disk_io_size;
2623        size_t blocksize = inode->i_sb->s_blocksize;
2624        unsigned long this_bio_flag = 0;
2625
2626        set_page_extent_mapped(page);
2627
2628        if (!PageUptodate(page)) {
2629                if (cleancache_get_page(page) == 0) {
2630                        BUG_ON(blocksize != PAGE_SIZE);
2631                        goto out;
2632                }
2633        }
2634
2635        end = page_end;
2636        while (1) {
2637                lock_extent(tree, start, end);
2638                ordered = btrfs_lookup_ordered_extent(inode, start);
2639                if (!ordered)
2640                        break;
2641                unlock_extent(tree, start, end);
2642                btrfs_start_ordered_extent(inode, ordered, 1);
2643                btrfs_put_ordered_extent(ordered);
2644        }
2645
2646        if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2647                char *userpage;
2648                size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2649
2650                if (zero_offset) {
2651                        iosize = PAGE_CACHE_SIZE - zero_offset;
2652                        userpage = kmap_atomic(page);
2653                        memset(userpage + zero_offset, 0, iosize);
2654                        flush_dcache_page(page);
2655                        kunmap_atomic(userpage);
2656                }
2657        }
2658        while (cur <= end) {
2659                if (cur >= last_byte) {
2660                        char *userpage;
2661                        struct extent_state *cached = NULL;
2662
2663                        iosize = PAGE_CACHE_SIZE - pg_offset;
2664                        userpage = kmap_atomic(page);
2665                        memset(userpage + pg_offset, 0, iosize);
2666                        flush_dcache_page(page);
2667                        kunmap_atomic(userpage);
2668                        set_extent_uptodate(tree, cur, cur + iosize - 1,
2669                                            &cached, GFP_NOFS);
2670                        unlock_extent_cached(tree, cur, cur + iosize - 1,
2671                                             &cached, GFP_NOFS);
2672                        break;
2673                }
2674                em = get_extent(inode, page, pg_offset, cur,
2675                                end - cur + 1, 0);
2676                if (IS_ERR_OR_NULL(em)) {
2677                        SetPageError(page);
2678                        unlock_extent(tree, cur, end);
2679                        break;
2680                }
2681                extent_offset = cur - em->start;
2682                BUG_ON(extent_map_end(em) <= cur);
2683                BUG_ON(end < cur);
2684
2685                if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2686                        this_bio_flag = EXTENT_BIO_COMPRESSED;
2687                        extent_set_compress_type(&this_bio_flag,
2688                                                 em->compress_type);
2689                }
2690
2691                iosize = min(extent_map_end(em) - cur, end - cur + 1);
2692                cur_end = min(extent_map_end(em) - 1, end);
2693                iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2694                if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2695                        disk_io_size = em->block_len;
2696                        sector = em->block_start >> 9;
2697                } else {
2698                        sector = (em->block_start + extent_offset) >> 9;
2699                        disk_io_size = iosize;
2700                }
2701                bdev = em->bdev;
2702                block_start = em->block_start;
2703                if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2704                        block_start = EXTENT_MAP_HOLE;
2705                free_extent_map(em);
2706                em = NULL;
2707
2708                /* we've found a hole, just zero and go on */
2709                if (block_start == EXTENT_MAP_HOLE) {
2710                        char *userpage;
2711                        struct extent_state *cached = NULL;
2712
2713                        userpage = kmap_atomic(page);
2714                        memset(userpage + pg_offset, 0, iosize);
2715                        flush_dcache_page(page);
2716                        kunmap_atomic(userpage);
2717
2718                        set_extent_uptodate(tree, cur, cur + iosize - 1,
2719                                            &cached, GFP_NOFS);
2720                        unlock_extent_cached(tree, cur, cur + iosize - 1,
2721                                             &cached, GFP_NOFS);
2722                        cur = cur + iosize;
2723                        pg_offset += iosize;
2724                        continue;
2725                }
2726                /* the get_extent function already copied into the page */
2727                if (test_range_bit(tree, cur, cur_end,
2728                                   EXTENT_UPTODATE, 1, NULL)) {
2729                        check_page_uptodate(tree, page);
2730                        unlock_extent(tree, cur, cur + iosize - 1);
2731                        cur = cur + iosize;
2732                        pg_offset += iosize;
2733                        continue;
2734                }
2735                /* we have an inline extent but it didn't get marked up
2736                 * to date.  Error out
2737                 */
2738                if (block_start == EXTENT_MAP_INLINE) {
2739                        SetPageError(page);
2740                        unlock_extent(tree, cur, cur + iosize - 1);
2741                        cur = cur + iosize;
2742                        pg_offset += iosize;
2743                        continue;
2744                }
2745
2746                ret = 0;
2747                if (tree->ops && tree->ops->readpage_io_hook) {
2748                        ret = tree->ops->readpage_io_hook(page, cur,
2749                                                          cur + iosize - 1);
2750                }
2751                if (!ret) {
2752                        unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2753                        pnr -= page->index;
2754                        ret = submit_extent_page(READ, tree, page,
2755                                         sector, disk_io_size, pg_offset,
2756                                         bdev, bio, pnr,
2757                                         end_bio_extent_readpage, mirror_num,
2758                                         *bio_flags,
2759                                         this_bio_flag);
2760                        if (!ret) {
2761                                nr++;
2762                                *bio_flags = this_bio_flag;
2763                        }
2764                }
2765                if (ret) {
2766                        SetPageError(page);
2767                        unlock_extent(tree, cur, cur + iosize - 1);
2768                }
2769                cur = cur + iosize;
2770                pg_offset += iosize;
2771        }
2772out:
2773        if (!nr) {
2774                if (!PageError(page))
2775                        SetPageUptodate(page);
2776                unlock_page(page);
2777        }
2778        return 0;
2779}
2780
2781int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2782                            get_extent_t *get_extent, int mirror_num)
2783{
2784        struct bio *bio = NULL;
2785        unsigned long bio_flags = 0;
2786        int ret;
2787
2788        ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
2789                                      &bio_flags);
2790        if (bio)
2791                ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
2792        return ret;
2793}
2794
2795static noinline void update_nr_written(struct page *page,
2796                                      struct writeback_control *wbc,
2797                                      unsigned long nr_written)
2798{
2799        wbc->nr_to_write -= nr_written;
2800        if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2801            wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2802                page->mapping->writeback_index = page->index + nr_written;
2803}
2804
2805/*
2806 * the writepage semantics are similar to regular writepage.  extent
2807 * records are inserted to lock ranges in the tree, and as dirty areas
2808 * are found, they are marked writeback.  Then the lock bits are removed
2809 * and the end_io handler clears the writeback ranges
2810 */
2811static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2812                              void *data)
2813{
2814        struct inode *inode = page->mapping->host;
2815        struct extent_page_data *epd = data;
2816        struct extent_io_tree *tree = epd->tree;
2817        u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2818        u64 delalloc_start;
2819        u64 page_end = start + PAGE_CACHE_SIZE - 1;
2820        u64 end;
2821        u64 cur = start;
2822        u64 extent_offset;
2823        u64 last_byte = i_size_read(inode);
2824        u64 block_start;
2825        u64 iosize;
2826        sector_t sector;
2827        struct extent_state *cached_state = NULL;
2828        struct extent_map *em;
2829        struct block_device *bdev;
2830        int ret;
2831        int nr = 0;
2832        size_t pg_offset = 0;
2833        size_t blocksize;
2834        loff_t i_size = i_size_read(inode);
2835        unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2836        u64 nr_delalloc;
2837        u64 delalloc_end;
2838        int page_started;
2839        int compressed;
2840        int write_flags;
2841        unsigned long nr_written = 0;
2842        bool fill_delalloc = true;
2843
2844        if (wbc->sync_mode == WB_SYNC_ALL)
2845                write_flags = WRITE_SYNC;
2846        else
2847                write_flags = WRITE;
2848
2849        trace___extent_writepage(page, inode, wbc);
2850
2851        WARN_ON(!PageLocked(page));
2852
2853        ClearPageError(page);
2854
2855        pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2856        if (page->index > end_index ||
2857           (page->index == end_index && !pg_offset)) {
2858                page->mapping->a_ops->invalidatepage(page, 0);
2859                unlock_page(page);
2860                return 0;
2861        }
2862
2863        if (page->index == end_index) {
2864                char *userpage;
2865
2866                userpage = kmap_atomic(page);
2867                memset(userpage + pg_offset, 0,
2868                       PAGE_CACHE_SIZE - pg_offset);
2869                kunmap_atomic(userpage);
2870                flush_dcache_page(page);
2871        }
2872        pg_offset = 0;
2873
2874        set_page_extent_mapped(page);
2875
2876        if (!tree->ops || !tree->ops->fill_delalloc)
2877                fill_delalloc = false;
2878
2879        delalloc_start = start;
2880        delalloc_end = 0;
2881        page_started = 0;
2882        if (!epd->extent_locked && fill_delalloc) {
2883                u64 delalloc_to_write = 0;
2884                /*
2885                 * make sure the wbc mapping index is at least updated
2886                 * to this page.
2887                 */
2888                update_nr_written(page, wbc, 0);
2889
2890                while (delalloc_end < page_end) {
2891                        nr_delalloc = find_lock_delalloc_range(inode, tree,
2892                                                       page,
2893                                                       &delalloc_start,
2894                                                       &delalloc_end,
2895                                                       128 * 1024 * 1024);
2896                        if (nr_delalloc == 0) {
2897                                delalloc_start = delalloc_end + 1;
2898                                continue;
2899                        }
2900                        ret = tree->ops->fill_delalloc(inode, page,
2901                                                       delalloc_start,
2902                                                       delalloc_end,
2903                                                       &page_started,
2904                                                       &nr_written);
2905                        /* File system has been set read-only */
2906                        if (ret) {
2907                                SetPageError(page);
2908                                goto done;
2909                        }
2910                        /*
2911                         * delalloc_end is already one less than the total
2912                         * length, so we don't subtract one from
2913                         * PAGE_CACHE_SIZE
2914                         */
2915                        delalloc_to_write += (delalloc_end - delalloc_start +
2916                                              PAGE_CACHE_SIZE) >>
2917                                              PAGE_CACHE_SHIFT;
2918                        delalloc_start = delalloc_end + 1;
2919                }
2920                if (wbc->nr_to_write < delalloc_to_write) {
2921                        int thresh = 8192;
2922
2923                        if (delalloc_to_write < thresh * 2)
2924                                thresh = delalloc_to_write;
2925                        wbc->nr_to_write = min_t(u64, delalloc_to_write,
2926                                                 thresh);
2927                }
2928
2929                /* did the fill delalloc function already unlock and start
2930                 * the IO?
2931                 */
2932                if (page_started) {
2933                        ret = 0;
2934                        /*
2935                         * we've unlocked the page, so we can't update
2936                         * the mapping's writeback index, just update
2937                         * nr_to_write.
2938                         */
2939                        wbc->nr_to_write -= nr_written;
2940                        goto done_unlocked;
2941                }
2942        }
2943        if (tree->ops && tree->ops->writepage_start_hook) {
2944                ret = tree->ops->writepage_start_hook(page, start,
2945                                                      page_end);
2946                if (ret) {
2947                        /* Fixup worker will requeue */
2948                        if (ret == -EBUSY)
2949                                wbc->pages_skipped++;
2950                        else
2951                                redirty_page_for_writepage(wbc, page);
2952                        update_nr_written(page, wbc, nr_written);
2953                        unlock_page(page);
2954                        ret = 0;
2955                        goto done_unlocked;
2956                }
2957        }
2958
2959        /*
2960         * we don't want to touch the inode after unlocking the page,
2961         * so we update the mapping writeback index now
2962         */
2963        update_nr_written(page, wbc, nr_written + 1);
2964
2965        end = page_end;
2966        if (last_byte <= start) {
2967                if (tree->ops && tree->ops->writepage_end_io_hook)
2968                        tree->ops->writepage_end_io_hook(page, start,
2969                                                         page_end, NULL, 1);
2970                goto done;
2971        }
2972
2973        blocksize = inode->i_sb->s_blocksize;
2974
2975        while (cur <= end) {
2976                if (cur >= last_byte) {
2977                        if (tree->ops && tree->ops->writepage_end_io_hook)
2978                                tree->ops->writepage_end_io_hook(page, cur,
2979                                                         page_end, NULL, 1);
2980                        break;
2981                }
2982                em = epd->get_extent(inode, page, pg_offset, cur,
2983                                     end - cur + 1, 1);
2984                if (IS_ERR_OR_NULL(em)) {
2985                        SetPageError(page);
2986                        break;
2987                }
2988
2989                extent_offset = cur - em->start;
2990                BUG_ON(extent_map_end(em) <= cur);
2991                BUG_ON(end < cur);
2992                iosize = min(extent_map_end(em) - cur, end - cur + 1);
2993                iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2994                sector = (em->block_start + extent_offset) >> 9;
2995                bdev = em->bdev;
2996                block_start = em->block_start;
2997                compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2998                free_extent_map(em);
2999                em = NULL;
3000
3001                /*
3002                 * compressed and inline extents are written through other
3003                 * paths in the FS
3004                 */
3005                if (compressed || block_start == EXTENT_MAP_HOLE ||
3006                    block_start == EXTENT_MAP_INLINE) {
3007                        /*
3008                         * end_io notification does not happen here for
3009                         * compressed extents
3010                         */
3011                        if (!compressed && tree->ops &&
3012                            tree->ops->writepage_end_io_hook)
3013                                tree->ops->writepage_end_io_hook(page, cur,
3014                                                         cur + iosize - 1,
3015                                                         NULL, 1);
3016                        else if (compressed) {
3017                                /* we don't want to end_page_writeback on
3018                                 * a compressed extent.  this happens
3019                                 * elsewhere
3020                                 */
3021                                nr++;
3022                        }
3023
3024                        cur += iosize;
3025                        pg_offset += iosize;
3026                        continue;
3027                }
3028                /* leave this out until we have a page_mkwrite call */
3029                if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
3030                                   EXTENT_DIRTY, 0, NULL)) {
3031                        cur = cur + iosize;
3032                        pg_offset += iosize;
3033                        continue;
3034                }
3035
3036                if (tree->ops && tree->ops->writepage_io_hook) {
3037                        ret = tree->ops->writepage_io_hook(page, cur,
3038                                                cur + iosize - 1);
3039                } else {
3040                        ret = 0;
3041                }
3042                if (ret) {
3043                        SetPageError(page);
3044                } else {
3045                        unsigned long max_nr = end_index + 1;
3046
3047                        set_range_writeback(tree, cur, cur + iosize - 1);
3048                        if (!PageWriteback(page)) {
3049                                printk(KERN_ERR "btrfs warning page %lu not "
3050                                       "writeback, cur %llu end %llu\n",
3051                                       page->index, (unsigned long long)cur,
3052                                       (unsigned long long)end);
3053                        }
3054
3055                        ret = submit_extent_page(write_flags, tree, page,
3056                                                 sector, iosize, pg_offset,
3057                                                 bdev, &epd->bio, max_nr,
3058                                                 end_bio_extent_writepage,
3059                                                 0, 0, 0);
3060                        if (ret)
3061                                SetPageError(page);
3062                }
3063                cur = cur + iosize;
3064                pg_offset += iosize;
3065                nr++;
3066        }
3067done:
3068        if (nr == 0) {
3069                /* make sure the mapping tag for page dirty gets cleared */
3070                set_page_writeback(page);
3071                end_page_writeback(page);
3072        }
3073        unlock_page(page);
3074
3075done_unlocked:
3076
3077        /* drop our reference on any cached states */
3078        free_extent_state(cached_state);
3079        return 0;
3080}
3081
3082static int eb_wait(void *word)
3083{
3084        io_schedule();
3085        return 0;
3086}
3087
3088static void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3089{
3090        wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3091                    TASK_UNINTERRUPTIBLE);
3092}
3093
3094static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3095                                     struct btrfs_fs_info *fs_info,
3096                                     struct extent_page_data *epd)
3097{
3098        unsigned long i, num_pages;
3099        int flush = 0;
3100        int ret = 0;
3101
3102        if (!btrfs_try_tree_write_lock(eb)) {
3103                flush = 1;
3104                flush_write_bio(epd);
3105                btrfs_tree_lock(eb);
3106        }
3107
3108        if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3109                btrfs_tree_unlock(eb);
3110                if (!epd->sync_io)
3111                        return 0;
3112                if (!flush) {
3113                        flush_write_bio(epd);
3114                        flush = 1;
3115                }
3116                while (1) {
3117                        wait_on_extent_buffer_writeback(eb);
3118                        btrfs_tree_lock(eb);
3119                        if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3120                                break;
3121                        btrfs_tree_unlock(eb);
3122                }
3123        }
3124
3125        /*
3126         * We need to do this to prevent races in people who check if the eb is
3127         * under IO since we can end up having no IO bits set for a short period
3128         * of time.
3129         */
3130        spin_lock(&eb->refs_lock);
3131        if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3132                set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3133                spin_unlock(&eb->refs_lock);
3134                btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3135                spin_lock(&fs_info->delalloc_lock);
3136                if (fs_info->dirty_metadata_bytes >= eb->len)
3137                        fs_info->dirty_metadata_bytes -= eb->len;
3138                else
3139                        WARN_ON(1);
3140                spin_unlock(&fs_info->delalloc_lock);
3141                ret = 1;
3142        } else {
3143                spin_unlock(&eb->refs_lock);
3144        }
3145
3146        btrfs_tree_unlock(eb);
3147
3148        if (!ret)
3149                return ret;
3150
3151        num_pages = num_extent_pages(eb->start, eb->len);
3152        for (i = 0; i < num_pages; i++) {
3153                struct page *p = extent_buffer_page(eb, i);
3154
3155                if (!trylock_page(p)) {
3156                        if (!flush) {
3157                                flush_write_bio(epd);
3158                                flush = 1;
3159                        }
3160                        lock_page(p);
3161                }
3162        }
3163
3164        return ret;
3165}
3166
3167static void end_extent_buffer_writeback(struct extent_buffer *eb)
3168{
3169        clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3170        smp_mb__after_clear_bit();
3171        wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3172}
3173
3174static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3175{
3176        int uptodate = err == 0;
3177        struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
3178        struct extent_buffer *eb;
3179        int done;
3180
3181        do {
3182                struct page *page = bvec->bv_page;
3183
3184                bvec--;
3185                eb = (struct extent_buffer *)page->private;
3186                BUG_ON(!eb);
3187                done = atomic_dec_and_test(&eb->io_pages);
3188
3189                if (!uptodate || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3190                        set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3191                        ClearPageUptodate(page);
3192                        SetPageError(page);
3193                }
3194
3195                end_page_writeback(page);
3196
3197                if (!done)
3198                        continue;
3199
3200                end_extent_buffer_writeback(eb);
3201        } while (bvec >= bio->bi_io_vec);
3202
3203        bio_put(bio);
3204
3205}
3206
3207static int write_one_eb(struct extent_buffer *eb,
3208                        struct btrfs_fs_info *fs_info,
3209                        struct writeback_control *wbc,
3210                        struct extent_page_data *epd)
3211{
3212        struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3213        u64 offset = eb->start;
3214        unsigned long i, num_pages;
3215        unsigned long bio_flags = 0;
3216        int rw = (epd->sync_io ? WRITE_SYNC : WRITE);
3217        int ret = 0;
3218
3219        clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3220        num_pages = num_extent_pages(eb->start, eb->len);
3221        atomic_set(&eb->io_pages, num_pages);
3222        if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3223                bio_flags = EXTENT_BIO_TREE_LOG;
3224
3225        for (i = 0; i < num_pages; i++) {
3226                struct page *p = extent_buffer_page(eb, i);
3227
3228                clear_page_dirty_for_io(p);
3229                set_page_writeback(p);
3230                ret = submit_extent_page(rw, eb->tree, p, offset >> 9,
3231                                         PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3232                                         -1, end_bio_extent_buffer_writepage,
3233                                         0, epd->bio_flags, bio_flags);
3234                epd->bio_flags = bio_flags;
3235                if (ret) {
3236                        set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3237                        SetPageError(p);
3238                        if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3239                                end_extent_buffer_writeback(eb);
3240                        ret = -EIO;
3241                        break;
3242                }
3243                offset += PAGE_CACHE_SIZE;
3244                update_nr_written(p, wbc, 1);
3245                unlock_page(p);
3246        }
3247
3248        if (unlikely(ret)) {
3249                for (; i < num_pages; i++) {
3250                        struct page *p = extent_buffer_page(eb, i);
3251                        unlock_page(p);
3252                }
3253        }
3254
3255        return ret;
3256}
3257
3258int btree_write_cache_pages(struct address_space *mapping,
3259                                   struct writeback_control *wbc)
3260{
3261        struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3262        struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3263        struct extent_buffer *eb, *prev_eb = NULL;
3264        struct extent_page_data epd = {
3265                .bio = NULL,
3266                .tree = tree,
3267                .extent_locked = 0,
3268                .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3269                .bio_flags = 0,
3270        };
3271        int ret = 0;
3272        int done = 0;
3273        int nr_to_write_done = 0;
3274        struct pagevec pvec;
3275        int nr_pages;
3276        pgoff_t index;
3277        pgoff_t end;            /* Inclusive */
3278        int scanned = 0;
3279        int tag;
3280
3281        pagevec_init(&pvec, 0);
3282        if (wbc->range_cyclic) {
3283                index = mapping->writeback_index; /* Start from prev offset */
3284                end = -1;
3285        } else {
3286                index = wbc->range_start >> PAGE_CACHE_SHIFT;
3287                end = wbc->range_end >> PAGE_CACHE_SHIFT;
3288                scanned = 1;
3289        }
3290        if (wbc->sync_mode == WB_SYNC_ALL)
3291                tag = PAGECACHE_TAG_TOWRITE;
3292        else
3293                tag = PAGECACHE_TAG_DIRTY;
3294retry:
3295        if (wbc->sync_mode == WB_SYNC_ALL)
3296                tag_pages_for_writeback(mapping, index, end);
3297        while (!done && !nr_to_write_done && (index <= end) &&
3298               (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3299                        min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3300                unsigned i;
3301
3302                scanned = 1;
3303                for (i = 0; i < nr_pages; i++) {
3304                        struct page *page = pvec.pages[i];
3305
3306                        if (!PagePrivate(page))
3307                                continue;
3308
3309                        if (!wbc->range_cyclic && page->index > end) {
3310                                done = 1;
3311                                break;
3312                        }
3313
3314                        spin_lock(&mapping->private_lock);
3315                        if (!PagePrivate(page)) {
3316                                spin_unlock(&mapping->private_lock);
3317                                continue;
3318                        }
3319
3320                        eb = (struct extent_buffer *)page->private;
3321
3322                        /*
3323                         * Shouldn't happen and normally this would be a BUG_ON
3324                         * but no sense in crashing the users box for something
3325                         * we can survive anyway.
3326                         */
3327                        if (!eb) {
3328                                spin_unlock(&mapping->private_lock);
3329                                WARN_ON(1);
3330                                continue;
3331                        }
3332
3333                        if (eb == prev_eb) {
3334                                spin_unlock(&mapping->private_lock);
3335                                continue;
3336                        }
3337
3338                        ret = atomic_inc_not_zero(&eb->refs);
3339                        spin_unlock(&mapping->private_lock);
3340                        if (!ret)
3341                                continue;
3342
3343                        prev_eb = eb;
3344                        ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3345                        if (!ret) {
3346                                free_extent_buffer(eb);
3347                                continue;
3348                        }
3349
3350                        ret = write_one_eb(eb, fs_info, wbc, &epd);
3351                        if (ret) {
3352                                done = 1;
3353                                free_extent_buffer(eb);
3354                                break;
3355                        }
3356                        free_extent_buffer(eb);
3357
3358                        /*
3359                         * the filesystem may choose to bump up nr_to_write.
3360                         * We have to make sure to honor the new nr_to_write
3361                         * at any time
3362                         */
3363                        nr_to_write_done = wbc->nr_to_write <= 0;
3364                }
3365                pagevec_release(&pvec);
3366                cond_resched();
3367        }
3368        if (!scanned && !done) {
3369                /*
3370                 * We hit the last page and there is more work to be done: wrap
3371                 * back to the start of the file
3372                 */
3373                scanned = 1;
3374                index = 0;
3375                goto retry;
3376        }
3377        flush_write_bio(&epd);
3378        return ret;
3379}
3380
3381/**
3382 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3383 * @mapping: address space structure to write
3384 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3385 * @writepage: function called for each page
3386 * @data: data passed to writepage function
3387 *
3388 * If a page is already under I/O, write_cache_pages() skips it, even
3389 * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3390 * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3391 * and msync() need to guarantee that all the data which was dirty at the time
3392 * the call was made get new I/O started against them.  If wbc->sync_mode is
3393 * WB_SYNC_ALL then we were called for data integrity and we must wait for
3394 * existing IO to complete.
3395 */
3396static int extent_write_cache_pages(struct extent_io_tree *tree,
3397                             struct address_space *mapping,
3398                             struct writeback_control *wbc,
3399                             writepage_t writepage, void *data,
3400                             void (*flush_fn)(void *))
3401{
3402        struct inode *inode = mapping->host;
3403        int ret = 0;
3404        int done = 0;
3405        int nr_to_write_done = 0;
3406        struct pagevec pvec;
3407        int nr_pages;
3408        pgoff_t index;
3409        pgoff_t end;            /* Inclusive */
3410        int scanned = 0;
3411        int tag;
3412
3413        /*
3414         * We have to hold onto the inode so that ordered extents can do their
3415         * work when the IO finishes.  The alternative to this is failing to add
3416         * an ordered extent if the igrab() fails there and that is a huge pain
3417         * to deal with, so instead just hold onto the inode throughout the
3418         * writepages operation.  If it fails here we are freeing up the inode
3419         * anyway and we'd rather not waste our time writing out stuff that is
3420         * going to be truncated anyway.
3421         */
3422        if (!igrab(inode))
3423                return 0;
3424
3425        pagevec_init(&pvec, 0);
3426        if (wbc->range_cyclic) {
3427                index = mapping->writeback_index; /* Start from prev offset */
3428                end = -1;
3429        } else {
3430                index = wbc->range_start >> PAGE_CACHE_SHIFT;
3431                end = wbc->range_end >> PAGE_CACHE_SHIFT;
3432                scanned = 1;
3433        }
3434        if (wbc->sync_mode == WB_SYNC_ALL)
3435                tag = PAGECACHE_TAG_TOWRITE;
3436        else
3437                tag = PAGECACHE_TAG_DIRTY;
3438retry:
3439        if (wbc->sync_mode == WB_SYNC_ALL)
3440                tag_pages_for_writeback(mapping, index, end);
3441        while (!done && !nr_to_write_done && (index <= end) &&
3442               (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3443                        min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3444                unsigned i;
3445
3446                scanned = 1;
3447                for (i = 0; i < nr_pages; i++) {
3448                        struct page *page = pvec.pages[i];
3449
3450                        /*
3451                         * At this point we hold neither mapping->tree_lock nor
3452                         * lock on the page itself: the page may be truncated or
3453                         * invalidated (changing page->mapping to NULL), or even
3454                         * swizzled back from swapper_space to tmpfs file
3455                         * mapping
3456                         */
3457                        if (tree->ops &&
3458                            tree->ops->write_cache_pages_lock_hook) {
3459                                tree->ops->write_cache_pages_lock_hook(page,
3460                                                               data, flush_fn);
3461                        } else {
3462                                if (!trylock_page(page)) {
3463                                        flush_fn(data);
3464                                        lock_page(page);
3465                                }
3466                        }
3467
3468                        if (unlikely(page->mapping != mapping)) {
3469                                unlock_page(page);
3470                                continue;
3471                        }
3472
3473                        if (!wbc->range_cyclic && page->index > end) {
3474                                done = 1;
3475                                unlock_page(page);
3476                                continue;
3477                        }
3478
3479                        if (wbc->sync_mode != WB_SYNC_NONE) {
3480                                if (PageWriteback(page))
3481                                        flush_fn(data);
3482                                wait_on_page_writeback(page);
3483                        }
3484
3485                        if (PageWriteback(page) ||
3486                            !clear_page_dirty_for_io(page)) {
3487                                unlock_page(page);
3488                                continue;
3489                        }
3490
3491                        ret = (*writepage)(page, wbc, data);
3492
3493                        if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3494                                unlock_page(page);
3495                                ret = 0;
3496                        }
3497                        if (ret)
3498                                done = 1;
3499
3500                        /*
3501                         * the filesystem may choose to bump up nr_to_write.
3502                         * We have to make sure to honor the new nr_to_write
3503                         * at any time
3504                         */
3505                        nr_to_write_done = wbc->nr_to_write <= 0;
3506                }
3507                pagevec_release(&pvec);
3508                cond_resched();
3509        }
3510        if (!scanned && !done) {
3511                /*
3512                 * We hit the last page and there is more work to be done: wrap
3513                 * back to the start of the file
3514                 */
3515                scanned = 1;
3516                index = 0;
3517                goto retry;
3518        }
3519        btrfs_add_delayed_iput(inode);
3520        return ret;
3521}
3522
3523static void flush_epd_write_bio(struct extent_page_data *epd)
3524{
3525        if (epd->bio) {
3526                int rw = WRITE;
3527                int ret;
3528
3529                if (epd->sync_io)
3530                        rw = WRITE_SYNC;
3531
3532                ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
3533                BUG_ON(ret < 0); /* -ENOMEM */
3534                epd->bio = NULL;
3535        }
3536}
3537
3538static noinline void flush_write_bio(void *data)
3539{
3540        struct extent_page_data *epd = data;
3541        flush_epd_write_bio(epd);
3542}
3543
3544int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3545                          get_extent_t *get_extent,
3546                          struct writeback_control *wbc)
3547{
3548        int ret;
3549        struct extent_page_data epd = {
3550                .bio = NULL,
3551                .tree = tree,
3552                .get_extent = get_extent,
3553                .extent_locked = 0,
3554                .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3555                .bio_flags = 0,
3556        };
3557
3558        ret = __extent_writepage(page, wbc, &epd);
3559
3560        flush_epd_write_bio(&epd);
3561        return ret;
3562}
3563
3564int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3565                              u64 start, u64 end, get_extent_t *get_extent,
3566                              int mode)
3567{
3568        int ret = 0;
3569        struct address_space *mapping = inode->i_mapping;
3570        struct page *page;
3571        unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3572                PAGE_CACHE_SHIFT;
3573
3574        struct extent_page_data epd = {
3575                .bio = NULL,
3576                .tree = tree,
3577                .get_extent = get_extent,
3578                .extent_locked = 1,
3579                .sync_io = mode == WB_SYNC_ALL,
3580                .bio_flags = 0,
3581        };
3582        struct writeback_control wbc_writepages = {
3583                .sync_mode      = mode,
3584                .nr_to_write    = nr_pages * 2,
3585                .range_start    = start,
3586                .range_end      = end + 1,
3587        };
3588
3589        while (start <= end) {
3590                page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3591                if (clear_page_dirty_for_io(page))
3592                        ret = __extent_writepage(page, &wbc_writepages, &epd);
3593                else {
3594                        if (tree->ops && tree->ops->writepage_end_io_hook)
3595                                tree->ops->writepage_end_io_hook(page, start,
3596                                                 start + PAGE_CACHE_SIZE - 1,
3597                                                 NULL, 1);
3598                        unlock_page(page);
3599                }
3600                page_cache_release(page);
3601                start += PAGE_CACHE_SIZE;
3602        }
3603
3604        flush_epd_write_bio(&epd);
3605        return ret;
3606}
3607
3608int extent_writepages(struct extent_io_tree *tree,
3609                      struct address_space *mapping,
3610                      get_extent_t *get_extent,
3611                      struct writeback_control *wbc)
3612{
3613        int ret = 0;
3614        struct extent_page_data epd = {
3615                .bio = NULL,
3616                .tree = tree,
3617                .get_extent = get_extent,
3618                .extent_locked = 0,
3619                .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3620                .bio_flags = 0,
3621        };
3622
3623        ret = extent_write_cache_pages(tree, mapping, wbc,
3624                                       __extent_writepage, &epd,
3625                                       flush_write_bio);
3626        flush_epd_write_bio(&epd);
3627        return ret;
3628}
3629
3630int extent_readpages(struct extent_io_tree *tree,
3631                     struct address_space *mapping,
3632                     struct list_head *pages, unsigned nr_pages,
3633                     get_extent_t get_extent)
3634{
3635        struct bio *bio = NULL;
3636        unsigned page_idx;
3637        unsigned long bio_flags = 0;
3638        struct page *pagepool[16];
3639        struct page *page;
3640        int i = 0;
3641        int nr = 0;
3642
3643        for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3644                page = list_entry(pages->prev, struct page, lru);
3645
3646                prefetchw(&page->flags);
3647                list_del(&page->lru);
3648                if (add_to_page_cache_lru(page, mapping,
3649                                        page->index, GFP_NOFS)) {
3650                        page_cache_release(page);
3651                        continue;
3652                }
3653
3654                pagepool[nr++] = page;
3655                if (nr < ARRAY_SIZE(pagepool))
3656                        continue;
3657                for (i = 0; i < nr; i++) {
3658                        __extent_read_full_page(tree, pagepool[i], get_extent,
3659                                        &bio, 0, &bio_flags);
3660                        page_cache_release(pagepool[i]);
3661                }
3662                nr = 0;
3663        }
3664        for (i = 0; i < nr; i++) {
3665                __extent_read_full_page(tree, pagepool[i], get_extent,
3666                                        &bio, 0, &bio_flags);
3667                page_cache_release(pagepool[i]);
3668        }
3669
3670        BUG_ON(!list_empty(pages));
3671        if (bio)
3672                return submit_one_bio(READ, bio, 0, bio_flags);
3673        return 0;
3674}
3675
3676/*
3677 * basic invalidatepage code, this waits on any locked or writeback
3678 * ranges corresponding to the page, and then deletes any extent state
3679 * records from the tree
3680 */
3681int extent_invalidatepage(struct extent_io_tree *tree,
3682                          struct page *page, unsigned long offset)
3683{
3684        struct extent_state *cached_state = NULL;
3685        u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
3686        u64 end = start + PAGE_CACHE_SIZE - 1;
3687        size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3688
3689        start += (offset + blocksize - 1) & ~(blocksize - 1);
3690        if (start > end)
3691                return 0;
3692
3693        lock_extent_bits(tree, start, end, 0, &cached_state);
3694        wait_on_page_writeback(page);
3695        clear_extent_bit(tree, start, end,
3696                         EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3697                         EXTENT_DO_ACCOUNTING,
3698                         1, 1, &cached_state, GFP_NOFS);
3699        return 0;
3700}
3701
3702/*
3703 * a helper for releasepage, this tests for areas of the page that
3704 * are locked or under IO and drops the related state bits if it is safe
3705 * to drop the page.
3706 */
3707int try_release_extent_state(struct extent_map_tree *map,
3708                             struct extent_io_tree *tree, struct page *page,
3709                             gfp_t mask)
3710{
3711        u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3712        u64 end = start + PAGE_CACHE_SIZE - 1;
3713        int ret = 1;
3714
3715        if (test_range_bit(tree, start, end,
3716                           EXTENT_IOBITS, 0, NULL))
3717                ret = 0;
3718        else {
3719                if ((mask & GFP_NOFS) == GFP_NOFS)
3720                        mask = GFP_NOFS;
3721                /*
3722                 * at this point we can safely clear everything except the
3723                 * locked bit and the nodatasum bit
3724                 */
3725                ret = clear_extent_bit(tree, start, end,
3726                                 ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3727                                 0, 0, NULL, mask);
3728
3729                /* if clear_extent_bit failed for enomem reasons,
3730                 * we can't allow the release to continue.
3731                 */
3732                if (ret < 0)
3733                        ret = 0;
3734                else
3735                        ret = 1;
3736        }
3737        return ret;
3738}
3739
3740/*
3741 * a helper for releasepage.  As long as there are no locked extents
3742 * in the range corresponding to the page, both state records and extent
3743 * map records are removed
3744 */
3745int try_release_extent_mapping(struct extent_map_tree *map,
3746                               struct extent_io_tree *tree, struct page *page,
3747                               gfp_t mask)
3748{
3749        struct extent_map *em;
3750        u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3751        u64 end = start + PAGE_CACHE_SIZE - 1;
3752
3753        if ((mask & __GFP_WAIT) &&
3754            page->mapping->host->i_size > 16 * 1024 * 1024) {
3755                u64 len;
3756                while (start <= end) {
3757                        len = end - start + 1;
3758                        write_lock(&map->lock);
3759                        em = lookup_extent_mapping(map, start, len);
3760                        if (!em) {
3761                                write_unlock(&map->lock);
3762                                break;
3763                        }
3764                        if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
3765                            em->start != start) {
3766                                write_unlock(&map->lock);
3767                                free_extent_map(em);
3768                                break;
3769                        }
3770                        if (!test_range_bit(tree, em->start,
3771                                            extent_map_end(em) - 1,
3772                                            EXTENT_LOCKED | EXTENT_WRITEBACK,
3773                                            0, NULL)) {
3774                                remove_extent_mapping(map, em);
3775                                /* once for the rb tree */
3776                                free_extent_map(em);
3777                        }
3778                        start = extent_map_end(em);
3779                        write_unlock(&map->lock);
3780
3781                        /* once for us */
3782                        free_extent_map(em);
3783                }
3784        }
3785        return try_release_extent_state(map, tree, page, mask);
3786}
3787
3788/*
3789 * helper function for fiemap, which doesn't want to see any holes.
3790 * This maps until we find something past 'last'
3791 */
3792static struct extent_map *get_extent_skip_holes(struct inode *inode,
3793                                                u64 offset,
3794                                                u64 last,
3795                                                get_extent_t *get_extent)
3796{
3797        u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
3798        struct extent_map *em;
3799        u64 len;
3800
3801        if (offset >= last)
3802                return NULL;
3803
3804        while(1) {
3805                len = last - offset;
3806                if (len == 0)
3807                        break;
3808                len = (len + sectorsize - 1) & ~(sectorsize - 1);
3809                em = get_extent(inode, NULL, 0, offset, len, 0);
3810                if (IS_ERR_OR_NULL(em))
3811                        return em;
3812
3813                /* if this isn't a hole return it */
3814                if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
3815                    em->block_start != EXTENT_MAP_HOLE) {
3816                        return em;
3817                }
3818
3819                /* this is a hole, advance to the next extent */
3820                offset = extent_map_end(em);
3821                free_extent_map(em);
3822                if (offset >= last)
3823                        break;
3824        }
3825        return NULL;
3826}
3827
3828int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
3829                __u64 start, __u64 len, get_extent_t *get_extent)
3830{
3831        int ret = 0;
3832        u64 off = start;
3833        u64 max = start + len;
3834        u32 flags = 0;
3835        u32 found_type;
3836        u64 last;
3837        u64 last_for_get_extent = 0;
3838        u64 disko = 0;
3839        u64 isize = i_size_read(inode);
3840        struct btrfs_key found_key;
3841        struct extent_map *em = NULL;
3842        struct extent_state *cached_state = NULL;
3843        struct btrfs_path *path;
3844        struct btrfs_file_extent_item *item;
3845        int end = 0;
3846        u64 em_start = 0;
3847        u64 em_len = 0;
3848        u64 em_end = 0;
3849        unsigned long emflags;
3850
3851        if (len == 0)
3852                return -EINVAL;
3853
3854        path = btrfs_alloc_path();
3855        if (!path)
3856                return -ENOMEM;
3857        path->leave_spinning = 1;
3858
3859        start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
3860        len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
3861
3862        /*
3863         * lookup the last file extent.  We're not using i_size here
3864         * because there might be preallocation past i_size
3865         */
3866        ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
3867                                       path, btrfs_ino(inode), -1, 0);
3868        if (ret < 0) {
3869                btrfs_free_path(path);
3870                return ret;
3871        }
3872        WARN_ON(!ret);
3873        path->slots[0]--;
3874        item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3875                              struct btrfs_file_extent_item);
3876        btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
3877        found_type = btrfs_key_type(&found_key);
3878
3879        /* No extents, but there might be delalloc bits */
3880        if (found_key.objectid != btrfs_ino(inode) ||
3881            found_type != BTRFS_EXTENT_DATA_KEY) {
3882                /* have to trust i_size as the end */
3883                last = (u64)-1;
3884                last_for_get_extent = isize;
3885        } else {
3886                /*
3887                 * remember the start of the last extent.  There are a
3888                 * bunch of different factors that go into the length of the
3889                 * extent, so its much less complex to remember where it started
3890                 */
3891                last = found_key.offset;
3892                last_for_get_extent = last + 1;
3893        }
3894        btrfs_free_path(path);
3895
3896        /*
3897         * we might have some extents allocated but more delalloc past those
3898         * extents.  so, we trust isize unless the start of the last extent is
3899         * beyond isize
3900         */
3901        if (last < isize) {
3902                last = (u64)-1;
3903                last_for_get_extent = isize;
3904        }
3905
3906        lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
3907                         &cached_state);
3908
3909        em = get_extent_skip_holes(inode, start, last_for_get_extent,
3910                                   get_extent);
3911        if (!em)
3912                goto out;
3913        if (IS_ERR(em)) {
3914                ret = PTR_ERR(em);
3915                goto out;
3916        }
3917
3918        while (!end) {
3919                u64 offset_in_extent;
3920
3921                /* break if the extent we found is outside the range */
3922                if (em->start >= max || extent_map_end(em) < off)
3923                        break;
3924
3925                /*
3926                 * get_extent may return an extent that starts before our
3927                 * requested range.  We have to make sure the ranges
3928                 * we return to fiemap always move forward and don't
3929                 * overlap, so adjust the offsets here
3930                 */
3931                em_start = max(em->start, off);
3932
3933                /*
3934                 * record the offset from the start of the extent
3935                 * for adjusting the disk offset below
3936                 */
3937                offset_in_extent = em_start - em->start;
3938                em_end = extent_map_end(em);
3939                em_len = em_end - em_start;
3940                emflags = em->flags;
3941                disko = 0;
3942                flags = 0;
3943
3944                /*
3945                 * bump off for our next call to get_extent
3946                 */
3947                off = extent_map_end(em);
3948                if (off >= max)
3949                        end = 1;
3950
3951                if (em->block_start == EXTENT_MAP_LAST_BYTE) {
3952                        end = 1;
3953                        flags |= FIEMAP_EXTENT_LAST;
3954                } else if (em->block_start == EXTENT_MAP_INLINE) {
3955                        flags |= (FIEMAP_EXTENT_DATA_INLINE |
3956                                  FIEMAP_EXTENT_NOT_ALIGNED);
3957                } else if (em->block_start == EXTENT_MAP_DELALLOC) {
3958                        flags |= (FIEMAP_EXTENT_DELALLOC |
3959                                  FIEMAP_EXTENT_UNKNOWN);
3960                } else {
3961                        disko = em->block_start + offset_in_extent;
3962                }
3963                if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3964                        flags |= FIEMAP_EXTENT_ENCODED;
3965
3966                free_extent_map(em);
3967                em = NULL;
3968                if ((em_start >= last) || em_len == (u64)-1 ||
3969                   (last == (u64)-1 && isize <= em_end)) {
3970                        flags |= FIEMAP_EXTENT_LAST;
3971                        end = 1;
3972                }
3973
3974                /* now scan forward to see if this is really the last extent. */
3975                em = get_extent_skip_holes(inode, off, last_for_get_extent,
3976                                           get_extent);
3977                if (IS_ERR(em)) {
3978                        ret = PTR_ERR(em);
3979                        goto out;
3980                }
3981                if (!em) {
3982                        flags |= FIEMAP_EXTENT_LAST;
3983                        end = 1;
3984                }
3985                ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3986                                              em_len, flags);
3987                if (ret)
3988                        goto out_free;
3989        }
3990out_free:
3991        free_extent_map(em);
3992out:
3993        unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3994                             &cached_state, GFP_NOFS);
3995        return ret;
3996}
3997
3998static void __free_extent_buffer(struct extent_buffer *eb)
3999{
4000#if LEAK_DEBUG
4001        unsigned long flags;
4002        spin_lock_irqsave(&leak_lock, flags);
4003        list_del(&eb->leak_list);
4004        spin_unlock_irqrestore(&leak_lock, flags);
4005#endif
4006        if (eb->pages && eb->pages != eb->inline_pages)
4007                kfree(eb->pages);
4008        kmem_cache_free(extent_buffer_cache, eb);
4009}
4010
4011static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
4012                                                   u64 start,
4013                                                   unsigned long len,
4014                                                   gfp_t mask)
4015{
4016        struct extent_buffer *eb = NULL;
4017#if LEAK_DEBUG
4018        unsigned long flags;
4019#endif
4020
4021        eb = kmem_cache_zalloc(extent_buffer_cache, mask);
4022        if (eb == NULL)
4023                return NULL;
4024        eb->start = start;
4025        eb->len = len;
4026        eb->tree = tree;
4027        eb->bflags = 0;
4028        rwlock_init(&eb->lock);
4029        atomic_set(&eb->write_locks, 0);
4030        atomic_set(&eb->read_locks, 0);
4031        atomic_set(&eb->blocking_readers, 0);
4032        atomic_set(&eb->blocking_writers, 0);
4033        atomic_set(&eb->spinning_readers, 0);
4034        atomic_set(&eb->spinning_writers, 0);
4035        eb->lock_nested = 0;
4036        init_waitqueue_head(&eb->write_lock_wq);
4037        init_waitqueue_head(&eb->read_lock_wq);
4038
4039#if LEAK_DEBUG
4040        spin_lock_irqsave(&leak_lock, flags);
4041        list_add(&eb->leak_list, &buffers);
4042        spin_unlock_irqrestore(&leak_lock, flags);
4043#endif
4044        spin_lock_init(&eb->refs_lock);
4045        atomic_set(&eb->refs, 1);
4046        atomic_set(&eb->io_pages, 0);
4047
4048        if (len > MAX_INLINE_EXTENT_BUFFER_SIZE) {
4049                struct page **pages;
4050                int num_pages = (len + PAGE_CACHE_SIZE - 1) >>
4051                        PAGE_CACHE_SHIFT;
4052                pages = kzalloc(num_pages, mask);
4053                if (!pages) {
4054                        __free_extent_buffer(eb);
4055                        return NULL;
4056                }
4057                eb->pages = pages;
4058        } else {
4059                eb->pages = eb->inline_pages;
4060        }
4061
4062        return eb;
4063}
4064
4065struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4066{
4067        unsigned long i;
4068        struct page *p;
4069        struct extent_buffer *new;
4070        unsigned long num_pages = num_extent_pages(src->start, src->len);
4071
4072        new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_ATOMIC);
4073        if (new == NULL)
4074                return NULL;
4075
4076        for (i = 0; i < num_pages; i++) {
4077                p = alloc_page(GFP_ATOMIC);
4078                BUG_ON(!p);
4079                attach_extent_buffer_page(new, p);
4080                WARN_ON(PageDirty(p));
4081                SetPageUptodate(p);
4082                new->pages[i] = p;
4083        }
4084
4085        copy_extent_buffer(new, src, 0, 0, src->len);
4086        set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4087        set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4088
4089        return new;
4090}
4091
4092struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4093{
4094        struct extent_buffer *eb;
4095        unsigned long num_pages = num_extent_pages(0, len);
4096        unsigned long i;
4097
4098        eb = __alloc_extent_buffer(NULL, start, len, GFP_ATOMIC);
4099        if (!eb)
4100                return NULL;
4101
4102        for (i = 0; i < num_pages; i++) {
4103                eb->pages[i] = alloc_page(GFP_ATOMIC);
4104                if (!eb->pages[i])
4105                        goto err;
4106        }
4107        set_extent_buffer_uptodate(eb);
4108        btrfs_set_header_nritems(eb, 0);
4109        set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4110
4111        return eb;
4112err:
4113        for (; i > 0; i--)
4114                __free_page(eb->pages[i - 1]);
4115        __free_extent_buffer(eb);
4116        return NULL;
4117}
4118
4119static int extent_buffer_under_io(struct extent_buffer *eb)
4120{
4121        return (atomic_read(&eb->io_pages) ||
4122                test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4123                test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4124}
4125
4126/*
4127 * Helper for releasing extent buffer page.
4128 */
4129static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4130                                                unsigned long start_idx)
4131{
4132        unsigned long index;
4133        unsigned long num_pages;
4134        struct page *page;
4135        int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4136
4137        BUG_ON(extent_buffer_under_io(eb));
4138
4139        num_pages = num_extent_pages(eb->start, eb->len);
4140        index = start_idx + num_pages;
4141        if (start_idx >= index)
4142                return;
4143
4144        do {
4145                index--;
4146                page = extent_buffer_page(eb, index);
4147                if (page && mapped) {
4148                        spin_lock(&page->mapping->private_lock);
4149                        /*
4150                         * We do this since we'll remove the pages after we've
4151                         * removed the eb from the radix tree, so we could race
4152                         * and have this page now attached to the new eb.  So
4153                         * only clear page_private if it's still connected to
4154                         * this eb.
4155                         */
4156                        if (PagePrivate(page) &&
4157                            page->private == (unsigned long)eb) {
4158                                BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4159                                BUG_ON(PageDirty(page));
4160                                BUG_ON(PageWriteback(page));
4161                                /*
4162                                 * We need to make sure we haven't be attached
4163                                 * to a new eb.
4164                                 */
4165                                ClearPagePrivate(page);
4166                                set_page_private(page, 0);
4167                                /* One for the page private */
4168                                page_cache_release(page);
4169                        }
4170                        spin_unlock(&page->mapping->private_lock);
4171
4172                }
4173                if (page) {
4174                        /* One for when we alloced the page */
4175                        page_cache_release(page);
4176                }
4177        } while (index != start_idx);
4178}
4179
4180/*
4181 * Helper for releasing the extent buffer.
4182 */
4183static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4184{
4185        btrfs_release_extent_buffer_page(eb, 0);
4186        __free_extent_buffer(eb);
4187}
4188
4189static void check_buffer_tree_ref(struct extent_buffer *eb)
4190{
4191        /* the ref bit is tricky.  We have to make sure it is set
4192         * if we have the buffer dirty.   Otherwise the
4193         * code to free a buffer can end up dropping a dirty
4194         * page
4195         *
4196         * Once the ref bit is set, it won't go away while the
4197         * buffer is dirty or in writeback, and it also won't
4198         * go away while we have the reference count on the
4199         * eb bumped.
4200         *
4201         * We can't just set the ref bit without bumping the
4202         * ref on the eb because free_extent_buffer might
4203         * see the ref bit and try to clear it.  If this happens
4204         * free_extent_buffer might end up dropping our original
4205         * ref by mistake and freeing the page before we are able
4206         * to add one more ref.
4207         *
4208         * So bump the ref count first, then set the bit.  If someone
4209         * beat us to it, drop the ref we added.
4210         */
4211        spin_lock(&eb->refs_lock);
4212        if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4213                atomic_inc(&eb->refs);
4214        spin_unlock(&eb->refs_lock);
4215}
4216
4217static void mark_extent_buffer_accessed(struct extent_buffer *eb)
4218{
4219        unsigned long num_pages, i;
4220
4221        check_buffer_tree_ref(eb);
4222
4223        num_pages = num_extent_pages(eb->start, eb->len);
4224        for (i = 0; i < num_pages; i++) {
4225                struct page *p = extent_buffer_page(eb, i);
4226                mark_page_accessed(p);
4227        }
4228}
4229
4230struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
4231                                          u64 start, unsigned long len)
4232{
4233        unsigned long num_pages = num_extent_pages(start, len);
4234        unsigned long i;
4235        unsigned long index = start >> PAGE_CACHE_SHIFT;
4236        struct extent_buffer *eb;
4237        struct extent_buffer *exists = NULL;
4238        struct page *p;
4239        struct address_space *mapping = tree->mapping;
4240        int uptodate = 1;
4241        int ret;
4242
4243        rcu_read_lock();
4244        eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4245        if (eb && atomic_inc_not_zero(&eb->refs)) {
4246                rcu_read_unlock();
4247                mark_extent_buffer_accessed(eb);
4248                return eb;
4249        }
4250        rcu_read_unlock();
4251
4252        eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
4253        if (!eb)
4254                return NULL;
4255
4256        for (i = 0; i < num_pages; i++, index++) {
4257                p = find_or_create_page(mapping, index, GFP_NOFS);
4258                if (!p)
4259                        goto free_eb;
4260
4261                spin_lock(&mapping->private_lock);
4262                if (PagePrivate(p)) {
4263                        /*
4264                         * We could have already allocated an eb for this page
4265                         * and attached one so lets see if we can get a ref on
4266                         * the existing eb, and if we can we know it's good and
4267                         * we can just return that one, else we know we can just
4268                         * overwrite page->private.
4269                         */
4270                        exists = (struct extent_buffer *)p->private;
4271                        if (atomic_inc_not_zero(&exists->refs)) {
4272                                spin_unlock(&mapping->private_lock);
4273                                unlock_page(p);
4274                                page_cache_release(p);
4275                                mark_extent_buffer_accessed(exists);
4276                                goto free_eb;
4277                        }
4278
4279                        /*
4280                         * Do this so attach doesn't complain and we need to
4281                         * drop the ref the old guy had.
4282                         */
4283                        ClearPagePrivate(p);
4284                        WARN_ON(PageDirty(p));
4285                        page_cache_release(p);
4286                }
4287                attach_extent_buffer_page(eb, p);
4288                spin_unlock(&mapping->private_lock);
4289                WARN_ON(PageDirty(p));
4290                mark_page_accessed(p);
4291                eb->pages[i] = p;
4292                if (!PageUptodate(p))
4293                        uptodate = 0;
4294
4295                /*
4296                 * see below about how we avoid a nasty race with release page
4297                 * and why we unlock later
4298                 */
4299        }
4300        if (uptodate)
4301                set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4302again:
4303        ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4304        if (ret)
4305                goto free_eb;
4306
4307        spin_lock(&tree->buffer_lock);
4308        ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
4309        if (ret == -EEXIST) {
4310                exists = radix_tree_lookup(&tree->buffer,
4311                                                start >> PAGE_CACHE_SHIFT);
4312                if (!atomic_inc_not_zero(&exists->refs)) {
4313                        spin_unlock(&tree->buffer_lock);
4314                        radix_tree_preload_end();
4315                        exists = NULL;
4316                        goto again;
4317                }
4318                spin_unlock(&tree->buffer_lock);
4319                radix_tree_preload_end();
4320                mark_extent_buffer_accessed(exists);
4321                goto free_eb;
4322        }
4323        /* add one reference for the tree */
4324        check_buffer_tree_ref(eb);
4325        spin_unlock(&tree->buffer_lock);
4326        radix_tree_preload_end();
4327
4328        /*
4329         * there is a race where release page may have
4330         * tried to find this extent buffer in the radix
4331         * but failed.  It will tell the VM it is safe to
4332         * reclaim the, and it will clear the page private bit.
4333         * We must make sure to set the page private bit properly
4334         * after the extent buffer is in the radix tree so
4335         * it doesn't get lost
4336         */
4337        SetPageChecked(eb->pages[0]);
4338        for (i = 1; i < num_pages; i++) {
4339                p = extent_buffer_page(eb, i);
4340                ClearPageChecked(p);
4341                unlock_page(p);
4342        }
4343        unlock_page(eb->pages[0]);
4344        return eb;
4345
4346free_eb:
4347        for (i = 0; i < num_pages; i++) {
4348                if (eb->pages[i])
4349                        unlock_page(eb->pages[i]);
4350        }
4351
4352        WARN_ON(!atomic_dec_and_test(&eb->refs));
4353        btrfs_release_extent_buffer(eb);
4354        return exists;
4355}
4356
4357struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
4358                                         u64 start, unsigned long len)
4359{
4360        struct extent_buffer *eb;
4361
4362        rcu_read_lock();
4363        eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4364        if (eb && atomic_inc_not_zero(&eb->refs)) {
4365                rcu_read_unlock();
4366                mark_extent_buffer_accessed(eb);
4367                return eb;
4368        }
4369        rcu_read_unlock();
4370
4371        return NULL;
4372}
4373
4374static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4375{
4376        struct extent_buffer *eb =
4377                        container_of(head, struct extent_buffer, rcu_head);
4378
4379        __free_extent_buffer(eb);
4380}
4381
4382/* Expects to have eb->eb_lock already held */
4383static int release_extent_buffer(struct extent_buffer *eb, gfp_t mask)
4384{
4385        WARN_ON(atomic_read(&eb->refs) == 0);
4386        if (atomic_dec_and_test(&eb->refs)) {
4387                if (test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags)) {
4388                        spin_unlock(&eb->refs_lock);
4389                } else {
4390                        struct extent_io_tree *tree = eb->tree;
4391
4392                        spin_unlock(&eb->refs_lock);
4393
4394                        spin_lock(&tree->buffer_lock);
4395                        radix_tree_delete(&tree->buffer,
4396                                          eb->start >> PAGE_CACHE_SHIFT);
4397                        spin_unlock(&tree->buffer_lock);
4398                }
4399
4400                /* Should be safe to release our pages at this point */
4401                btrfs_release_extent_buffer_page(eb, 0);
4402                call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4403                return 1;
4404        }
4405        spin_unlock(&eb->refs_lock);
4406
4407        return 0;
4408}
4409
4410void free_extent_buffer(struct extent_buffer *eb)
4411{
4412        if (!eb)
4413                return;
4414
4415        spin_lock(&eb->refs_lock);
4416        if (atomic_read(&eb->refs) == 2 &&
4417            test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4418                atomic_dec(&eb->refs);
4419
4420        if (atomic_read(&eb->refs) == 2 &&
4421            test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4422            !extent_buffer_under_io(eb) &&
4423            test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4424                atomic_dec(&eb->refs);
4425
4426        /*
4427         * I know this is terrible, but it's temporary until we stop tracking
4428         * the uptodate bits and such for the extent buffers.
4429         */
4430        release_extent_buffer(eb, GFP_ATOMIC);
4431}
4432
4433void free_extent_buffer_stale(struct extent_buffer *eb)
4434{
4435        if (!eb)
4436                return;
4437
4438        spin_lock(&eb->refs_lock);
4439        set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4440
4441        if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4442            test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4443                atomic_dec(&eb->refs);
4444        release_extent_buffer(eb, GFP_NOFS);
4445}
4446
4447void clear_extent_buffer_dirty(struct extent_buffer *eb)
4448{
4449        unsigned long i;
4450        unsigned long num_pages;
4451        struct page *page;
4452
4453        num_pages = num_extent_pages(eb->start, eb->len);
4454
4455        for (i = 0; i < num_pages; i++) {
4456                page = extent_buffer_page(eb, i);
4457                if (!PageDirty(page))
4458                        continue;
4459
4460                lock_page(page);
4461                WARN_ON(!PagePrivate(page));
4462
4463                clear_page_dirty_for_io(page);
4464                spin_lock_irq(&page->mapping->tree_lock);
4465                if (!PageDirty(page)) {
4466                        radix_tree_tag_clear(&page->mapping->page_tree,
4467                                                page_index(page),
4468                                                PAGECACHE_TAG_DIRTY);
4469                }
4470                spin_unlock_irq(&page->mapping->tree_lock);
4471                ClearPageError(page);
4472                unlock_page(page);
4473        }
4474        WARN_ON(atomic_read(&eb->refs) == 0);
4475}
4476
4477int set_extent_buffer_dirty(struct extent_buffer *eb)
4478{
4479        unsigned long i;
4480        unsigned long num_pages;
4481        int was_dirty = 0;
4482
4483        check_buffer_tree_ref(eb);
4484
4485        was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4486
4487        num_pages = num_extent_pages(eb->start, eb->len);
4488        WARN_ON(atomic_read(&eb->refs) == 0);
4489        WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4490
4491        for (i = 0; i < num_pages; i++)
4492                set_page_dirty(extent_buffer_page(eb, i));
4493        return was_dirty;
4494}
4495
4496static int range_straddles_pages(u64 start, u64 len)
4497{
4498        if (len < PAGE_CACHE_SIZE)
4499                return 1;
4500        if (start & (PAGE_CACHE_SIZE - 1))
4501                return 1;
4502        if ((start + len) & (PAGE_CACHE_SIZE - 1))
4503                return 1;
4504        return 0;
4505}
4506
4507int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4508{
4509        unsigned long i;
4510        struct page *page;
4511        unsigned long num_pages;
4512
4513        clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4514        num_pages = num_extent_pages(eb->start, eb->len);
4515        for (i = 0; i < num_pages; i++) {
4516                page = extent_buffer_page(eb, i);
4517                if (page)
4518                        ClearPageUptodate(page);
4519        }
4520        return 0;
4521}
4522
4523int set_extent_buffer_uptodate(struct extent_buffer *eb)
4524{
4525        unsigned long i;
4526        struct page *page;
4527        unsigned long num_pages;
4528
4529        set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4530        num_pages = num_extent_pages(eb->start, eb->len);
4531        for (i = 0; i < num_pages; i++) {
4532                page = extent_buffer_page(eb, i);
4533                SetPageUptodate(page);
4534        }
4535        return 0;
4536}
4537
4538int extent_range_uptodate(struct extent_io_tree *tree,
4539                          u64 start, u64 end)
4540{
4541        struct page *page;
4542        int ret;
4543        int pg_uptodate = 1;
4544        int uptodate;
4545        unsigned long index;
4546
4547        if (range_straddles_pages(start, end - start + 1)) {
4548                ret = test_range_bit(tree, start, end,
4549                                     EXTENT_UPTODATE, 1, NULL);
4550                if (ret)
4551                        return 1;
4552        }
4553        while (start <= end) {
4554                index = start >> PAGE_CACHE_SHIFT;
4555                page = find_get_page(tree->mapping, index);
4556                if (!page)
4557                        return 1;
4558                uptodate = PageUptodate(page);
4559                page_cache_release(page);
4560                if (!uptodate) {
4561                        pg_uptodate = 0;
4562                        break;
4563                }
4564                start += PAGE_CACHE_SIZE;
4565        }
4566        return pg_uptodate;
4567}
4568
4569int extent_buffer_uptodate(struct extent_buffer *eb)
4570{
4571        return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4572}
4573
4574int read_extent_buffer_pages(struct extent_io_tree *tree,
4575                             struct extent_buffer *eb, u64 start, int wait,
4576                             get_extent_t *get_extent, int mirror_num)
4577{
4578        unsigned long i;
4579        unsigned long start_i;
4580        struct page *page;
4581        int err;
4582        int ret = 0;
4583        int locked_pages = 0;
4584        int all_uptodate = 1;
4585        unsigned long num_pages;
4586        unsigned long num_reads = 0;
4587        struct bio *bio = NULL;
4588        unsigned long bio_flags = 0;
4589
4590        if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4591                return 0;
4592
4593        if (start) {
4594                WARN_ON(start < eb->start);
4595                start_i = (start >> PAGE_CACHE_SHIFT) -
4596                        (eb->start >> PAGE_CACHE_SHIFT);
4597        } else {
4598                start_i = 0;
4599        }
4600
4601        num_pages = num_extent_pages(eb->start, eb->len);
4602        for (i = start_i; i < num_pages; i++) {
4603                page = extent_buffer_page(eb, i);
4604                if (wait == WAIT_NONE) {
4605                        if (!trylock_page(page))
4606                                goto unlock_exit;
4607                } else {
4608                        lock_page(page);
4609                }
4610                locked_pages++;
4611                if (!PageUptodate(page)) {
4612                        num_reads++;
4613                        all_uptodate = 0;
4614                }
4615        }
4616        if (all_uptodate) {
4617                if (start_i == 0)
4618                        set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4619                goto unlock_exit;
4620        }
4621
4622        clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4623        eb->read_mirror = 0;
4624        atomic_set(&eb->io_pages, num_reads);
4625        for (i = start_i; i < num_pages; i++) {
4626                page = extent_buffer_page(eb, i);
4627                if (!PageUptodate(page)) {
4628                        ClearPageError(page);
4629                        err = __extent_read_full_page(tree, page,
4630                                                      get_extent, &bio,
4631                                                      mirror_num, &bio_flags);
4632                        if (err)
4633                                ret = err;
4634                } else {
4635                        unlock_page(page);
4636                }
4637        }
4638
4639        if (bio) {
4640                err = submit_one_bio(READ, bio, mirror_num, bio_flags);
4641                if (err)
4642                        return err;
4643        }
4644
4645        if (ret || wait != WAIT_COMPLETE)
4646                return ret;
4647
4648        for (i = start_i; i < num_pages; i++) {
4649                page = extent_buffer_page(eb, i);
4650                wait_on_page_locked(page);
4651                if (!PageUptodate(page))
4652                        ret = -EIO;
4653        }
4654
4655        return ret;
4656
4657unlock_exit:
4658        i = start_i;
4659        while (locked_pages > 0) {
4660                page = extent_buffer_page(eb, i);
4661                i++;
4662                unlock_page(page);
4663                locked_pages--;
4664        }
4665        return ret;
4666}
4667
4668void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4669                        unsigned long start,
4670                        unsigned long len)
4671{
4672        size_t cur;
4673        size_t offset;
4674        struct page *page;
4675        char *kaddr;
4676        char *dst = (char *)dstv;
4677        size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4678        unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4679
4680        WARN_ON(start > eb->len);
4681        WARN_ON(start + len > eb->start + eb->len);
4682
4683        offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4684
4685        while (len > 0) {
4686                page = extent_buffer_page(eb, i);
4687
4688                cur = min(len, (PAGE_CACHE_SIZE - offset));
4689                kaddr = page_address(page);
4690                memcpy(dst, kaddr + offset, cur);
4691
4692                dst += cur;
4693                len -= cur;
4694                offset = 0;
4695                i++;
4696        }
4697}
4698
4699int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4700                               unsigned long min_len, char **map,
4701                               unsigned long *map_start,
4702                               unsigned long *map_len)
4703{
4704        size_t offset = start & (PAGE_CACHE_SIZE - 1);
4705        char *kaddr;
4706        struct page *p;
4707        size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4708        unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4709        unsigned long end_i = (start_offset + start + min_len - 1) >>
4710                PAGE_CACHE_SHIFT;
4711
4712        if (i != end_i)
4713                return -EINVAL;
4714
4715        if (i == 0) {
4716                offset = start_offset;
4717                *map_start = 0;
4718        } else {
4719                offset = 0;
4720                *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4721        }
4722
4723        if (start + min_len > eb->len) {
4724                printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4725                       "wanted %lu %lu\n", (unsigned long long)eb->start,
4726                       eb->len, start, min_len);
4727                WARN_ON(1);
4728                return -EINVAL;
4729        }
4730
4731        p = extent_buffer_page(eb, i);
4732        kaddr = page_address(p);
4733        *map = kaddr + offset;
4734        *map_len = PAGE_CACHE_SIZE - offset;
4735        return 0;
4736}
4737
4738int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4739                          unsigned long start,
4740                          unsigned long len)
4741{
4742        size_t cur;
4743        size_t offset;
4744        struct page *page;
4745        char *kaddr;
4746        char *ptr = (char *)ptrv;
4747        size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4748        unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4749        int ret = 0;
4750
4751        WARN_ON(start > eb->len);
4752        WARN_ON(start + len > eb->start + eb->len);
4753
4754        offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4755
4756        while (len > 0) {
4757                page = extent_buffer_page(eb, i);
4758
4759                cur = min(len, (PAGE_CACHE_SIZE - offset));
4760
4761                kaddr = page_address(page);
4762                ret = memcmp(ptr, kaddr + offset, cur);
4763                if (ret)
4764                        break;
4765
4766                ptr += cur;
4767                len -= cur;
4768                offset = 0;
4769                i++;
4770        }
4771        return ret;
4772}
4773
4774void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4775                         unsigned long start, unsigned long len)
4776{
4777        size_t cur;
4778        size_t offset;
4779        struct page *page;
4780        char *kaddr;
4781        char *src = (char *)srcv;
4782        size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4783        unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4784
4785        WARN_ON(start > eb->len);
4786        WARN_ON(start + len > eb->start + eb->len);
4787
4788        offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4789
4790        while (len > 0) {
4791                page = extent_buffer_page(eb, i);
4792                WARN_ON(!PageUptodate(page));
4793
4794                cur = min(len, PAGE_CACHE_SIZE - offset);
4795                kaddr = page_address(page);
4796                memcpy(kaddr + offset, src, cur);
4797
4798                src += cur;
4799                len -= cur;
4800                offset = 0;
4801                i++;
4802        }
4803}
4804
4805void memset_extent_buffer(struct extent_buffer *eb, char c,
4806                          unsigned long start, unsigned long len)
4807{
4808        size_t cur;
4809        size_t offset;
4810        struct page *page;
4811        char *kaddr;
4812        size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4813        unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4814
4815        WARN_ON(start > eb->len);
4816        WARN_ON(start + len > eb->start + eb->len);
4817
4818        offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4819
4820        while (len > 0) {
4821                page = extent_buffer_page(eb, i);
4822                WARN_ON(!PageUptodate(page));
4823
4824                cur = min(len, PAGE_CACHE_SIZE - offset);
4825                kaddr = page_address(page);
4826                memset(kaddr + offset, c, cur);
4827
4828                len -= cur;
4829                offset = 0;
4830                i++;
4831        }
4832}
4833
4834void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
4835                        unsigned long dst_offset, unsigned long src_offset,
4836                        unsigned long len)
4837{
4838        u64 dst_len = dst->len;
4839        size_t cur;
4840        size_t offset;
4841        struct page *page;
4842        char *kaddr;
4843        size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4844        unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4845
4846        WARN_ON(src->len != dst_len);
4847
4848        offset = (start_offset + dst_offset) &
4849                ((unsigned long)PAGE_CACHE_SIZE - 1);
4850
4851        while (len > 0) {
4852                page = extent_buffer_page(dst, i);
4853                WARN_ON(!PageUptodate(page));
4854
4855                cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
4856
4857                kaddr = page_address(page);
4858                read_extent_buffer(src, kaddr + offset, src_offset, cur);
4859
4860                src_offset += cur;
4861                len -= cur;
4862                offset = 0;
4863                i++;
4864        }
4865}
4866
4867static void move_pages(struct page *dst_page, struct page *src_page,
4868                       unsigned long dst_off, unsigned long src_off,
4869                       unsigned long len)
4870{
4871        char *dst_kaddr = page_address(dst_page);
4872        if (dst_page == src_page) {
4873                memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
4874        } else {
4875                char *src_kaddr = page_address(src_page);
4876                char *p = dst_kaddr + dst_off + len;
4877                char *s = src_kaddr + src_off + len;
4878
4879                while (len--)
4880                        *--p = *--s;
4881        }
4882}
4883
4884static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4885{
4886        unsigned long distance = (src > dst) ? src - dst : dst - src;
4887        return distance < len;
4888}
4889
4890static void copy_pages(struct page *dst_page, struct page *src_page,
4891                       unsigned long dst_off, unsigned long src_off,
4892                       unsigned long len)
4893{
4894        char *dst_kaddr = page_address(dst_page);
4895        char *src_kaddr;
4896        int must_memmove = 0;
4897
4898        if (dst_page != src_page) {
4899                src_kaddr = page_address(src_page);
4900        } else {
4901                src_kaddr = dst_kaddr;
4902                if (areas_overlap(src_off, dst_off, len))
4903                        must_memmove = 1;
4904        }
4905
4906        if (must_memmove)
4907                memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
4908        else
4909                memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
4910}
4911
4912void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4913                           unsigned long src_offset, unsigned long len)
4914{
4915        size_t cur;
4916        size_t dst_off_in_page;
4917        size_t src_off_in_page;
4918        size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4919        unsigned long dst_i;
4920        unsigned long src_i;
4921
4922        if (src_offset + len > dst->len) {
4923                printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4924                       "len %lu dst len %lu\n", src_offset, len, dst->len);
4925                BUG_ON(1);
4926        }
4927        if (dst_offset + len > dst->len) {
4928                printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4929                       "len %lu dst len %lu\n", dst_offset, len, dst->len);
4930                BUG_ON(1);
4931        }
4932
4933        while (len > 0) {
4934                dst_off_in_page = (start_offset + dst_offset) &
4935                        ((unsigned long)PAGE_CACHE_SIZE - 1);
4936                src_off_in_page = (start_offset + src_offset) &
4937                        ((unsigned long)PAGE_CACHE_SIZE - 1);
4938
4939                dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4940                src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
4941
4942                cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
4943                                               src_off_in_page));
4944                cur = min_t(unsigned long, cur,
4945                        (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
4946
4947                copy_pages(extent_buffer_page(dst, dst_i),
4948                           extent_buffer_page(dst, src_i),
4949                           dst_off_in_page, src_off_in_page, cur);
4950
4951                src_offset += cur;
4952                dst_offset += cur;
4953                len -= cur;
4954        }
4955}
4956
4957void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4958                           unsigned long src_offset, unsigned long len)
4959{
4960        size_t cur;
4961        size_t dst_off_in_page;
4962        size_t src_off_in_page;
4963        unsigned long dst_end = dst_offset + len - 1;
4964        unsigned long src_end = src_offset + len - 1;
4965        size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4966        unsigned long dst_i;
4967        unsigned long src_i;
4968
4969        if (src_offset + len > dst->len) {
4970                printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4971                       "len %lu len %lu\n", src_offset, len, dst->len);
4972                BUG_ON(1);
4973        }
4974        if (dst_offset + len > dst->len) {
4975                printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4976                       "len %lu len %lu\n", dst_offset, len, dst->len);
4977                BUG_ON(1);
4978        }
4979        if (dst_offset < src_offset) {
4980                memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4981                return;
4982        }
4983        while (len > 0) {
4984                dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
4985                src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
4986
4987                dst_off_in_page = (start_offset + dst_end) &
4988                        ((unsigned long)PAGE_CACHE_SIZE - 1);
4989                src_off_in_page = (start_offset + src_end) &
4990                        ((unsigned long)PAGE_CACHE_SIZE - 1);
4991
4992                cur = min_t(unsigned long, len, src_off_in_page + 1);
4993                cur = min(cur, dst_off_in_page + 1);
4994                move_pages(extent_buffer_page(dst, dst_i),
4995                           extent_buffer_page(dst, src_i),
4996                           dst_off_in_page - cur + 1,
4997                           src_off_in_page - cur + 1, cur);
4998
4999                dst_end -= cur;
5000                src_end -= cur;
5001                len -= cur;
5002        }
5003}
5004
5005int try_release_extent_buffer(struct page *page, gfp_t mask)
5006{
5007        struct extent_buffer *eb;
5008
5009        /*
5010         * We need to make sure noboody is attaching this page to an eb right
5011         * now.
5012         */
5013        spin_lock(&page->mapping->private_lock);
5014        if (!PagePrivate(page)) {
5015                spin_unlock(&page->mapping->private_lock);
5016                return 1;
5017        }
5018
5019        eb = (struct extent_buffer *)page->private;
5020        BUG_ON(!eb);
5021
5022        /*
5023         * This is a little awful but should be ok, we need to make sure that
5024         * the eb doesn't disappear out from under us while we're looking at
5025         * this page.
5026         */
5027        spin_lock(&eb->refs_lock);
5028        if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5029                spin_unlock(&eb->refs_lock);
5030                spin_unlock(&page->mapping->private_lock);
5031                return 0;
5032        }
5033        spin_unlock(&page->mapping->private_lock);
5034
5035        if ((mask & GFP_NOFS) == GFP_NOFS)
5036                mask = GFP_NOFS;
5037
5038        /*
5039         * If tree ref isn't set then we know the ref on this eb is a real ref,
5040         * so just return, this page will likely be freed soon anyway.
5041         */
5042        if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5043                spin_unlock(&eb->refs_lock);
5044                return 0;
5045        }
5046
5047        return release_extent_buffer(eb, mask);
5048}
5049
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.