fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include <linux/percpu_counter.h>
  28 #include "hash.h"
  29 #include "tree-log.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "volumes.h"
  33 #include "raid56.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36 #include "math.h"
  37 #include "sysfs.h"
  38 #include "qgroup.h"
  39
  40 #undef SCRAMBLE_DELAYED_REFS
  41
  42 /*
  43  * control flags for do_chunk_alloc's force field
  44  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  45  * if we really need one.
  46  *
  47  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  48  * if we have very few chunks already allocated.  This is
  49  * used as part of the clustering code to help make sure
  50  * we have a good pool of storage to cluster in, without
  51  * filling the FS with empty chunks
  52  *
  53  * CHUNK_ALLOC_FORCE means it must try to allocate one
  54  *
  55  */
  56 enum {
  57         CHUNK_ALLOC_NO_FORCE = 0,
  58         CHUNK_ALLOC_LIMITED = 1,
  59         CHUNK_ALLOC_FORCE = 2,
  60 };
  61
  62 /*
  63  * Control how reservations are dealt with.
  64  *
  65  * RESERVE_FREE - freeing a reservation.
  66  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  67  *   ENOSPC accounting
  68  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  69  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  70  */
  71 enum {
  72         RESERVE_FREE = 0,
  73         RESERVE_ALLOC = 1,
  74         RESERVE_ALLOC_NO_ACCOUNT = 2,
  75 };
  76
  77 static int update_block_group(struct btrfs_trans_handle *trans,
  78                               struct btrfs_root *root, u64 bytenr,
  79                               u64 num_bytes, int alloc);
  80 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  81                                 struct btrfs_root *root,
  82                                 u64 bytenr, u64 num_bytes, u64 parent,
  83                                 u64 root_objectid, u64 owner_objectid,
  84                                 u64 owner_offset, int refs_to_drop,
  85                                 struct btrfs_delayed_extent_op *extra_op,
  86                                 int no_quota);
  87 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  88                                     struct extent_buffer *leaf,
  89                                     struct btrfs_extent_item *ei);
  90 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  91                                       struct btrfs_root *root,
  92                                       u64 parent, u64 root_objectid,
  93                                       u64 flags, u64 owner, u64 offset,
  94                                       struct btrfs_key *ins, int ref_mod);
  95 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  96                                      struct btrfs_root *root,
  97                                      u64 parent, u64 root_objectid,
  98                                      u64 flags, struct btrfs_disk_key *key,
  99                                      int level, struct btrfs_key *ins,
 100                                      int no_quota);
 101 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
 102                           struct btrfs_root *extent_root, u64 flags,
 103                           int force);
 104 static int find_next_key(struct btrfs_path *path, int level,
 105                          struct btrfs_key *key);
 106 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 107                             int dump_block_groups);
 108 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 109                                        u64 num_bytes, int reserve,
 110                                        int delalloc);
 111 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 112                                u64 num_bytes);
 113 int btrfs_pin_extent(struct btrfs_root *root,
 114                      u64 bytenr, u64 num_bytes, int reserved);
 115
 116 static noinline int
 117 block_group_cache_done(struct btrfs_block_group_cache *cache)
 118 {
 119         smp_mb();
 120         return cache->cached == BTRFS_CACHE_FINISHED ||
 121                 cache->cached == BTRFS_CACHE_ERROR;
 122 }
 123
 124 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 125 {
 126         return (cache->flags & bits) == bits;
 127 }
 128
 129 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 130 {
 131         atomic_inc(&cache->count);
 132 }
 133
 134 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 135 {
 136         if (atomic_dec_and_test(&cache->count)) {
 137                 WARN_ON(cache->pinned > 0);
 138                 WARN_ON(cache->reserved > 0);
 139                 kfree(cache->free_space_ctl);
 140                 kfree(cache);
 141         }
 142 }
 143
 144 /*
 145  * this adds the block group to the fs_info rb tree for the block group
 146  * cache
 147  */
 148 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 149                                 struct btrfs_block_group_cache *block_group)
 150 {
 151         struct rb_node **p;
 152         struct rb_node *parent = NULL;
 153         struct btrfs_block_group_cache *cache;
 154
 155         spin_lock(&info->block_group_cache_lock);
 156         p = &info->block_group_cache_tree.rb_node;
 157
 158         while (*p) {
 159                 parent = *p;
 160                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 161                                  cache_node);
 162                 if (block_group->key.objectid < cache->key.objectid) {
 163                         p = &(*p)->rb_left;
 164                 } else if (block_group->key.objectid > cache->key.objectid) {
 165                         p = &(*p)->rb_right;
 166                 } else {
 167                         spin_unlock(&info->block_group_cache_lock);
 168                         return -EEXIST;
 169                 }
 170         }
 171
 172         rb_link_node(&block_group->cache_node, parent, p);
 173         rb_insert_color(&block_group->cache_node,
 174                         &info->block_group_cache_tree);
 175
 176         if (info->first_logical_byte > block_group->key.objectid)
 177                 info->first_logical_byte = block_group->key.objectid;
 178
 179         spin_unlock(&info->block_group_cache_lock);
 180
 181         return 0;
 182 }
 183
 184 /*
 185  * This will return the block group at or after bytenr if contains is 0, else
 186  * it will return the block group that contains the bytenr
 187  */
 188 static struct btrfs_block_group_cache *
 189 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 190                               int contains)
 191 {
 192         struct btrfs_block_group_cache *cache, *ret = NULL;
 193         struct rb_node *n;
 194         u64 end, start;
 195
 196         spin_lock(&info->block_group_cache_lock);
 197         n = info->block_group_cache_tree.rb_node;
 198
 199         while (n) {
 200                 cache = rb_entry(n, struct btrfs_block_group_cache,
 201                                  cache_node);
 202                 end = cache->key.objectid + cache->key.offset - 1;
 203                 start = cache->key.objectid;
 204
 205                 if (bytenr < start) {
 206                         if (!contains && (!ret || start < ret->key.objectid))
 207                                 ret = cache;
 208                         n = n->rb_left;
 209                 } else if (bytenr > start) {
 210                         if (contains && bytenr <= end) {
 211                                 ret = cache;
 212                                 break;
 213                         }
 214                         n = n->rb_right;
 215                 } else {
 216                         ret = cache;
 217                         break;
 218                 }
 219         }
 220         if (ret) {
 221                 btrfs_get_block_group(ret);
 222                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 223                         info->first_logical_byte = ret->key.objectid;
 224         }
 225         spin_unlock(&info->block_group_cache_lock);
 226
 227         return ret;
 228 }
 229
 230 static int add_excluded_extent(struct btrfs_root *root,
 231                                u64 start, u64 num_bytes)
 232 {
 233         u64 end = start + num_bytes - 1;
 234         set_extent_bits(&root->fs_info->freed_extents[0],
 235                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 236         set_extent_bits(&root->fs_info->freed_extents[1],
 237                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 238         return 0;
 239 }
 240
 241 static void free_excluded_extents(struct btrfs_root *root,
 242                                   struct btrfs_block_group_cache *cache)
 243 {
 244         u64 start, end;
 245
 246         start = cache->key.objectid;
 247         end = start + cache->key.offset - 1;
 248
 249         clear_extent_bits(&root->fs_info->freed_extents[0],
 250                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 251         clear_extent_bits(&root->fs_info->freed_extents[1],
 252                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 253 }
 254
 255 static int exclude_super_stripes(struct btrfs_root *root,
 256                                  struct btrfs_block_group_cache *cache)
 257 {
 258         u64 bytenr;
 259         u64 *logical;
 260         int stripe_len;
 261         int i, nr, ret;
 262
 263         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 264                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 265                 cache->bytes_super += stripe_len;
 266                 ret = add_excluded_extent(root, cache->key.objectid,
 267                                           stripe_len);
 268                 if (ret)
 269                         return ret;
 270         }
 271
 272         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 273                 bytenr = btrfs_sb_offset(i);
 274                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 275                                        cache->key.objectid, bytenr,
 276                                        0, &logical, &nr, &stripe_len);
 277                 if (ret)
 278                         return ret;
 279
 280                 while (nr--) {
 281                         u64 start, len;
 282
 283                         if (logical[nr] > cache->key.objectid +
 284                             cache->key.offset)
 285                                 continue;
 286
 287                         if (logical[nr] + stripe_len <= cache->key.objectid)
 288                                 continue;
 289
 290                         start = logical[nr];
 291                         if (start < cache->key.objectid) {
 292                                 start = cache->key.objectid;
 293                                 len = (logical[nr] + stripe_len) - start;
 294                         } else {
 295                                 len = min_t(u64, stripe_len,
 296                                             cache->key.objectid +
 297                                             cache->key.offset - start);
 298                         }
 299
 300                         cache->bytes_super += len;
 301                         ret = add_excluded_extent(root, start, len);
 302                         if (ret) {
 303                                 kfree(logical);
 304                                 return ret;
 305                         }
 306                 }
 307
 308                 kfree(logical);
 309         }
 310         return 0;
 311 }
 312
 313 static struct btrfs_caching_control *
 314 get_caching_control(struct btrfs_block_group_cache *cache)
 315 {
 316         struct btrfs_caching_control *ctl;
 317
 318         spin_lock(&cache->lock);
 319         if (!cache->caching_ctl) {
 320                 spin_unlock(&cache->lock);
 321                 return NULL;
 322         }
 323
 324         ctl = cache->caching_ctl;
 325         atomic_inc(&ctl->count);
 326         spin_unlock(&cache->lock);
 327         return ctl;
 328 }
 329
 330 static void put_caching_control(struct btrfs_caching_control *ctl)
 331 {
 332         if (atomic_dec_and_test(&ctl->count))
 333                 kfree(ctl);
 334 }
 335
 336 /*
 337  * this is only called by cache_block_group, since we could have freed extents
 338  * we need to check the pinned_extents for any extents that can't be used yet
 339  * since their free space will be released as soon as the transaction commits.
 340  */
 341 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 342                               struct btrfs_fs_info *info, u64 start, u64 end)
 343 {
 344         u64 extent_start, extent_end, size, total_added = 0;
 345         int ret;
 346
 347         while (start < end) {
 348                 ret = find_first_extent_bit(info->pinned_extents, start,
 349                                             &extent_start, &extent_end,
 350                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 351                                             NULL);
 352                 if (ret)
 353                         break;
 354
 355                 if (extent_start <= start) {
 356                         start = extent_end + 1;
 357                 } else if (extent_start > start && extent_start < end) {
 358                         size = extent_start - start;
 359                         total_added += size;
 360                         ret = btrfs_add_free_space(block_group, start,
 361                                                    size);
 362                         BUG_ON(ret); /* -ENOMEM or logic error */
 363                         start = extent_end + 1;
 364                 } else {
 365                         break;
 366                 }
 367         }
 368
 369         if (start < end) {
 370                 size = end - start;
 371                 total_added += size;
 372                 ret = btrfs_add_free_space(block_group, start, size);
 373                 BUG_ON(ret); /* -ENOMEM or logic error */
 374         }
 375
 376         return total_added;
 377 }
 378
 379 static noinline void caching_thread(struct btrfs_work *work)
 380 {
 381         struct btrfs_block_group_cache *block_group;
 382         struct btrfs_fs_info *fs_info;
 383         struct btrfs_caching_control *caching_ctl;
 384         struct btrfs_root *extent_root;
 385         struct btrfs_path *path;
 386         struct extent_buffer *leaf;
 387         struct btrfs_key key;
 388         u64 total_found = 0;
 389         u64 last = 0;
 390         u32 nritems;
 391         int ret = -ENOMEM;
 392
 393         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 394         block_group = caching_ctl->block_group;
 395         fs_info = block_group->fs_info;
 396         extent_root = fs_info->extent_root;
 397
 398         path = btrfs_alloc_path();
 399         if (!path)
 400                 goto out;
 401
 402         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 403
 404         /*
 405          * We don't want to deadlock with somebody trying to allocate a new
 406          * extent for the extent root while also trying to search the extent
 407          * root to add free space.  So we skip locking and search the commit
 408          * root, since its read-only
 409          */
 410         path->skip_locking = 1;
 411         path->search_commit_root = 1;
 412         path->reada = 1;
 413
 414         key.objectid = last;
 415         key.offset = 0;
 416         key.type = BTRFS_EXTENT_ITEM_KEY;
 417 again:
 418         mutex_lock(&caching_ctl->mutex);
 419         /* need to make sure the commit_root doesn't disappear */
 420         down_read(&fs_info->commit_root_sem);
 421
 422 next:
 423         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 424         if (ret < 0)
 425                 goto err;
 426
 427         leaf = path->nodes[0];
 428         nritems = btrfs_header_nritems(leaf);
 429
 430         while (1) {
 431                 if (btrfs_fs_closing(fs_info) > 1) {
 432                         last = (u64)-1;
 433                         break;
 434                 }
 435
 436                 if (path->slots[0] < nritems) {
 437                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 438                 } else {
 439                         ret = find_next_key(path, 0, &key);
 440                         if (ret)
 441                                 break;
 442
 443                         if (need_resched() ||
 444                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 445                                 caching_ctl->progress = last;
 446                                 btrfs_release_path(path);
 447                                 up_read(&fs_info->commit_root_sem);
 448                                 mutex_unlock(&caching_ctl->mutex);
 449                                 cond_resched();
 450                                 goto again;
 451                         }
 452
 453                         ret = btrfs_next_leaf(extent_root, path);
 454                         if (ret < 0)
 455                                 goto err;
 456                         if (ret)
 457                                 break;
 458                         leaf = path->nodes[0];
 459                         nritems = btrfs_header_nritems(leaf);
 460                         continue;
 461                 }
 462
 463                 if (key.objectid < last) {
 464                         key.objectid = last;
 465                         key.offset = 0;
 466                         key.type = BTRFS_EXTENT_ITEM_KEY;
 467
 468                         caching_ctl->progress = last;
 469                         btrfs_release_path(path);
 470                         goto next;
 471                 }
 472
 473                 if (key.objectid < block_group->key.objectid) {
 474                         path->slots[0]++;
 475                         continue;
 476                 }
 477
 478                 if (key.objectid >= block_group->key.objectid +
 479                     block_group->key.offset)
 480                         break;
 481
 482                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 483                     key.type == BTRFS_METADATA_ITEM_KEY) {
 484                         total_found += add_new_free_space(block_group,
 485                                                           fs_info, last,
 486                                                           key.objectid);
 487                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 488                                 last = key.objectid +
 489                                         fs_info->tree_root->nodesize;
 490                         else
 491                                 last = key.objectid + key.offset;
 492
 493                         if (total_found > (1024 * 1024 * 2)) {
 494                                 total_found = 0;
 495                                 wake_up(&caching_ctl->wait);
 496                         }
 497                 }
 498                 path->slots[0]++;
 499         }
 500         ret = 0;
 501
 502         total_found += add_new_free_space(block_group, fs_info, last,
 503                                           block_group->key.objectid +
 504                                           block_group->key.offset);
 505         caching_ctl->progress = (u64)-1;
 506
 507         spin_lock(&block_group->lock);
 508         block_group->caching_ctl = NULL;
 509         block_group->cached = BTRFS_CACHE_FINISHED;
 510         spin_unlock(&block_group->lock);
 511
 512 err:
 513         btrfs_free_path(path);
 514         up_read(&fs_info->commit_root_sem);
 515
 516         free_excluded_extents(extent_root, block_group);
 517
 518         mutex_unlock(&caching_ctl->mutex);
 519 out:
 520         if (ret) {
 521                 spin_lock(&block_group->lock);
 522                 block_group->caching_ctl = NULL;
 523                 block_group->cached = BTRFS_CACHE_ERROR;
 524                 spin_unlock(&block_group->lock);
 525         }
 526         wake_up(&caching_ctl->wait);
 527
 528         put_caching_control(caching_ctl);
 529         btrfs_put_block_group(block_group);
 530 }
 531
 532 static int cache_block_group(struct btrfs_block_group_cache *cache,
 533                              int load_cache_only)
 534 {
 535         DEFINE_WAIT(wait);
 536         struct btrfs_fs_info *fs_info = cache->fs_info;
 537         struct btrfs_caching_control *caching_ctl;
 538         int ret = 0;
 539
 540         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 541         if (!caching_ctl)
 542                 return -ENOMEM;
 543
 544         INIT_LIST_HEAD(&caching_ctl->list);
 545         mutex_init(&caching_ctl->mutex);
 546         init_waitqueue_head(&caching_ctl->wait);
 547         caching_ctl->block_group = cache;
 548         caching_ctl->progress = cache->key.objectid;
 549         atomic_set(&caching_ctl->count, 1);
 550         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 551                         caching_thread, NULL, NULL);
 552
 553         spin_lock(&cache->lock);
 554         /*
 555          * This should be a rare occasion, but this could happen I think in the
 556          * case where one thread starts to load the space cache info, and then
 557          * some other thread starts a transaction commit which tries to do an
 558          * allocation while the other thread is still loading the space cache
 559          * info.  The previous loop should have kept us from choosing this block
 560          * group, but if we've moved to the state where we will wait on caching
 561          * block groups we need to first check if we're doing a fast load here,
 562          * so we can wait for it to finish, otherwise we could end up allocating
 563          * from a block group who's cache gets evicted for one reason or
 564          * another.
 565          */
 566         while (cache->cached == BTRFS_CACHE_FAST) {
 567                 struct btrfs_caching_control *ctl;
 568
 569                 ctl = cache->caching_ctl;
 570                 atomic_inc(&ctl->count);
 571                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 572                 spin_unlock(&cache->lock);
 573
 574                 schedule();
 575
 576                 finish_wait(&ctl->wait, &wait);
 577                 put_caching_control(ctl);
 578                 spin_lock(&cache->lock);
 579         }
 580
 581         if (cache->cached != BTRFS_CACHE_NO) {
 582                 spin_unlock(&cache->lock);
 583                 kfree(caching_ctl);
 584                 return 0;
 585         }
 586         WARN_ON(cache->caching_ctl);
 587         cache->caching_ctl = caching_ctl;
 588         cache->cached = BTRFS_CACHE_FAST;
 589         spin_unlock(&cache->lock);
 590
 591         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 592                 mutex_lock(&caching_ctl->mutex);
 593                 ret = load_free_space_cache(fs_info, cache);
 594
 595                 spin_lock(&cache->lock);
 596                 if (ret == 1) {
 597                         cache->caching_ctl = NULL;
 598                         cache->cached = BTRFS_CACHE_FINISHED;
 599                         cache->last_byte_to_unpin = (u64)-1;
 600                         caching_ctl->progress = (u64)-1;
 601                 } else {
 602                         if (load_cache_only) {
 603                                 cache->caching_ctl = NULL;
 604                                 cache->cached = BTRFS_CACHE_NO;
 605                         } else {
 606                                 cache->cached = BTRFS_CACHE_STARTED;
 607                                 cache->has_caching_ctl = 1;
 608                         }
 609                 }
 610                 spin_unlock(&cache->lock);
 611                 mutex_unlock(&caching_ctl->mutex);
 612
 613                 wake_up(&caching_ctl->wait);
 614                 if (ret == 1) {
 615                         put_caching_control(caching_ctl);
 616                         free_excluded_extents(fs_info->extent_root, cache);
 617                         return 0;
 618                 }
 619         } else {
 620                 /*
 621                  * We are not going to do the fast caching, set cached to the
 622                  * appropriate value and wakeup any waiters.
 623                  */
 624                 spin_lock(&cache->lock);
 625                 if (load_cache_only) {
 626                         cache->caching_ctl = NULL;
 627                         cache->cached = BTRFS_CACHE_NO;
 628                 } else {
 629                         cache->cached = BTRFS_CACHE_STARTED;
 630                         cache->has_caching_ctl = 1;
 631                 }
 632                 spin_unlock(&cache->lock);
 633                 wake_up(&caching_ctl->wait);
 634         }
 635
 636         if (load_cache_only) {
 637                 put_caching_control(caching_ctl);
 638                 return 0;
 639         }
 640
 641         down_write(&fs_info->commit_root_sem);
 642         atomic_inc(&caching_ctl->count);
 643         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 644         up_write(&fs_info->commit_root_sem);
 645
 646         btrfs_get_block_group(cache);
 647
 648         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 649
 650         return ret;
 651 }
 652
 653 /*
 654  * return the block group that starts at or after bytenr
 655  */
 656 static struct btrfs_block_group_cache *
 657 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 658 {
 659         struct btrfs_block_group_cache *cache;
 660
 661         cache = block_group_cache_tree_search(info, bytenr, 0);
 662
 663         return cache;
 664 }
 665
 666 /*
 667  * return the block group that contains the given bytenr
 668  */
 669 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 670                                                  struct btrfs_fs_info *info,
 671                                                  u64 bytenr)
 672 {
 673         struct btrfs_block_group_cache *cache;
 674
 675         cache = block_group_cache_tree_search(info, bytenr, 1);
 676
 677         return cache;
 678 }
 679
 680 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 681                                                   u64 flags)
 682 {
 683         struct list_head *head = &info->space_info;
 684         struct btrfs_space_info *found;
 685
 686         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 687
 688         rcu_read_lock();
 689         list_for_each_entry_rcu(found, head, list) {
 690                 if (found->flags & flags) {
 691                         rcu_read_unlock();
 692                         return found;
 693                 }
 694         }
 695         rcu_read_unlock();
 696         return NULL;
 697 }
 698
 699 /*
 700  * after adding space to the filesystem, we need to clear the full flags
 701  * on all the space infos.
 702  */
 703 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 704 {
 705         struct list_head *head = &info->space_info;
 706         struct btrfs_space_info *found;
 707
 708         rcu_read_lock();
 709         list_for_each_entry_rcu(found, head, list)
 710                 found->full = 0;
 711         rcu_read_unlock();
 712 }
 713
 714 /* simple helper to search for an existing data extent at a given offset */
 715 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 716 {
 717         int ret;
 718         struct btrfs_key key;
 719         struct btrfs_path *path;
 720
 721         path = btrfs_alloc_path();
 722         if (!path)
 723                 return -ENOMEM;
 724
 725         key.objectid = start;
 726         key.offset = len;
 727         key.type = BTRFS_EXTENT_ITEM_KEY;
 728         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 729                                 0, 0);
 730         btrfs_free_path(path);
 731         return ret;
 732 }
 733
 734 /*
 735  * helper function to lookup reference count and flags of a tree block.
 736  *
 737  * the head node for delayed ref is used to store the sum of all the
 738  * reference count modifications queued up in the rbtree. the head
 739  * node may also store the extent flags to set. This way you can check
 740  * to see what the reference count and extent flags would be if all of
 741  * the delayed refs are not processed.
 742  */
 743 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 744                              struct btrfs_root *root, u64 bytenr,
 745                              u64 offset, int metadata, u64 *refs, u64 *flags)
 746 {
 747         struct btrfs_delayed_ref_head *head;
 748         struct btrfs_delayed_ref_root *delayed_refs;
 749         struct btrfs_path *path;
 750         struct btrfs_extent_item *ei;
 751         struct extent_buffer *leaf;
 752         struct btrfs_key key;
 753         u32 item_size;
 754         u64 num_refs;
 755         u64 extent_flags;
 756         int ret;
 757
 758         /*
 759          * If we don't have skinny metadata, don't bother doing anything
 760          * different
 761          */
 762         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 763                 offset = root->nodesize;
 764                 metadata = 0;
 765         }
 766
 767         path = btrfs_alloc_path();
 768         if (!path)
 769                 return -ENOMEM;
 770
 771         if (!trans) {
 772                 path->skip_locking = 1;
 773                 path->search_commit_root = 1;
 774         }
 775
 776 search_again:
 777         key.objectid = bytenr;
 778         key.offset = offset;
 779         if (metadata)
 780                 key.type = BTRFS_METADATA_ITEM_KEY;
 781         else
 782                 key.type = BTRFS_EXTENT_ITEM_KEY;
 783
 784         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 785                                 &key, path, 0, 0);
 786         if (ret < 0)
 787                 goto out_free;
 788
 789         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 790                 if (path->slots[0]) {
 791                         path->slots[0]--;
 792                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 793                                               path->slots[0]);
 794                         if (key.objectid == bytenr &&
 795                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 796                             key.offset == root->nodesize)
 797                                 ret = 0;
 798                 }
 799         }
 800
 801         if (ret == 0) {
 802                 leaf = path->nodes[0];
 803                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 804                 if (item_size >= sizeof(*ei)) {
 805                         ei = btrfs_item_ptr(leaf, path->slots[0],
 806                                             struct btrfs_extent_item);
 807                         num_refs = btrfs_extent_refs(leaf, ei);
 808                         extent_flags = btrfs_extent_flags(leaf, ei);
 809                 } else {
 810 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 811                         struct btrfs_extent_item_v0 *ei0;
 812                         BUG_ON(item_size != sizeof(*ei0));
 813                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 814                                              struct btrfs_extent_item_v0);
 815                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 816                         /* FIXME: this isn't correct for data */
 817                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 818 #else
 819                         BUG();
 820 #endif
 821                 }
 822                 BUG_ON(num_refs == 0);
 823         } else {
 824                 num_refs = 0;
 825                 extent_flags = 0;
 826                 ret = 0;
 827         }
 828
 829         if (!trans)
 830                 goto out;
 831
 832         delayed_refs = &trans->transaction->delayed_refs;
 833         spin_lock(&delayed_refs->lock);
 834         head = btrfs_find_delayed_ref_head(trans, bytenr);
 835         if (head) {
 836                 if (!mutex_trylock(&head->mutex)) {
 837                         atomic_inc(&head->node.refs);
 838                         spin_unlock(&delayed_refs->lock);
 839
 840                         btrfs_release_path(path);
 841
 842                         /*
 843                          * Mutex was contended, block until it's released and try
 844                          * again
 845                          */
 846                         mutex_lock(&head->mutex);
 847                         mutex_unlock(&head->mutex);
 848                         btrfs_put_delayed_ref(&head->node);
 849                         goto search_again;
 850                 }
 851                 spin_lock(&head->lock);
 852                 if (head->extent_op && head->extent_op->update_flags)
 853                         extent_flags |= head->extent_op->flags_to_set;
 854                 else
 855                         BUG_ON(num_refs == 0);
 856
 857                 num_refs += head->node.ref_mod;
 858                 spin_unlock(&head->lock);
 859                 mutex_unlock(&head->mutex);
 860         }
 861         spin_unlock(&delayed_refs->lock);
 862 out:
 863         WARN_ON(num_refs == 0);
 864         if (refs)
 865                 *refs = num_refs;
 866         if (flags)
 867                 *flags = extent_flags;
 868 out_free:
 869         btrfs_free_path(path);
 870         return ret;
 871 }
 872
 873 /*
 874  * Back reference rules.  Back refs have three main goals:
 875  *
 876  * 1) differentiate between all holders of references to an extent so that
 877  *    when a reference is dropped we can make sure it was a valid reference
 878  *    before freeing the extent.
 879  *
 880  * 2) Provide enough information to quickly find the holders of an extent
 881  *    if we notice a given block is corrupted or bad.
 882  *
 883  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 884  *    maintenance.  This is actually the same as #2, but with a slightly
 885  *    different use case.
 886  *
 887  * There are two kinds of back refs. The implicit back refs is optimized
 888  * for pointers in non-shared tree blocks. For a given pointer in a block,
 889  * back refs of this kind provide information about the block's owner tree
 890  * and the pointer's key. These information allow us to find the block by
 891  * b-tree searching. The full back refs is for pointers in tree blocks not
 892  * referenced by their owner trees. The location of tree block is recorded
 893  * in the back refs. Actually the full back refs is generic, and can be
 894  * used in all cases the implicit back refs is used. The major shortcoming
 895  * of the full back refs is its overhead. Every time a tree block gets
 896  * COWed, we have to update back refs entry for all pointers in it.
 897  *
 898  * For a newly allocated tree block, we use implicit back refs for
 899  * pointers in it. This means most tree related operations only involve
 900  * implicit back refs. For a tree block created in old transaction, the
 901  * only way to drop a reference to it is COW it. So we can detect the
 902  * event that tree block loses its owner tree's reference and do the
 903  * back refs conversion.
 904  *
 905  * When a tree block is COW'd through a tree, there are four cases:
 906  *
 907  * The reference count of the block is one and the tree is the block's
 908  * owner tree. Nothing to do in this case.
 909  *
 910  * The reference count of the block is one and the tree is not the
 911  * block's owner tree. In this case, full back refs is used for pointers
 912  * in the block. Remove these full back refs, add implicit back refs for
 913  * every pointers in the new block.
 914  *
 915  * The reference count of the block is greater than one and the tree is
 916  * the block's owner tree. In this case, implicit back refs is used for
 917  * pointers in the block. Add full back refs for every pointers in the
 918  * block, increase lower level extents' reference counts. The original
 919  * implicit back refs are entailed to the new block.
 920  *
 921  * The reference count of the block is greater than one and the tree is
 922  * not the block's owner tree. Add implicit back refs for every pointer in
 923  * the new block, increase lower level extents' reference count.
 924  *
 925  * Back Reference Key composing:
 926  *
 927  * The key objectid corresponds to the first byte in the extent,
 928  * The key type is used to differentiate between types of back refs.
 929  * There are different meanings of the key offset for different types
 930  * of back refs.
 931  *
 932  * File extents can be referenced by:
 933  *
 934  * - multiple snapshots, subvolumes, or different generations in one subvol
 935  * - different files inside a single subvolume
 936  * - different offsets inside a file (bookend extents in file.c)
 937  *
 938  * The extent ref structure for the implicit back refs has fields for:
 939  *
 940  * - Objectid of the subvolume root
 941  * - objectid of the file holding the reference
 942  * - original offset in the file
 943  * - how many bookend extents
 944  *
 945  * The key offset for the implicit back refs is hash of the first
 946  * three fields.
 947  *
 948  * The extent ref structure for the full back refs has field for:
 949  *
 950  * - number of pointers in the tree leaf
 951  *
 952  * The key offset for the implicit back refs is the first byte of
 953  * the tree leaf
 954  *
 955  * When a file extent is allocated, The implicit back refs is used.
 956  * the fields are filled in:
 957  *
 958  *     (root_key.objectid, inode objectid, offset in file, 1)
 959  *
 960  * When a file extent is removed file truncation, we find the
 961  * corresponding implicit back refs and check the following fields:
 962  *
 963  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 964  *
 965  * Btree extents can be referenced by:
 966  *
 967  * - Different subvolumes
 968  *
 969  * Both the implicit back refs and the full back refs for tree blocks
 970  * only consist of key. The key offset for the implicit back refs is
 971  * objectid of block's owner tree. The key offset for the full back refs
 972  * is the first byte of parent block.
 973  *
 974  * When implicit back refs is used, information about the lowest key and
 975  * level of the tree block are required. These information are stored in
 976  * tree block info structure.
 977  */
 978
 979 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 980 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 981                                   struct btrfs_root *root,
 982                                   struct btrfs_path *path,
 983                                   u64 owner, u32 extra_size)
 984 {
 985         struct btrfs_extent_item *item;
 986         struct btrfs_extent_item_v0 *ei0;
 987         struct btrfs_extent_ref_v0 *ref0;
 988         struct btrfs_tree_block_info *bi;
 989         struct extent_buffer *leaf;
 990         struct btrfs_key key;
 991         struct btrfs_key found_key;
 992         u32 new_size = sizeof(*item);
 993         u64 refs;
 994         int ret;
 995
 996         leaf = path->nodes[0];
 997         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 998
 999         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1000         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1001                              struct btrfs_extent_item_v0);
1002         refs = btrfs_extent_refs_v0(leaf, ei0);
1003
1004         if (owner == (u64)-1) {
1005                 while (1) {
1006                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1007                                 ret = btrfs_next_leaf(root, path);
1008                                 if (ret < 0)
1009                                         return ret;
1010                                 BUG_ON(ret > 0); /* Corruption */
1011                                 leaf = path->nodes[0];
1012                         }
1013                         btrfs_item_key_to_cpu(leaf, &found_key,
1014                                               path->slots[0]);
1015                         BUG_ON(key.objectid != found_key.objectid);
1016                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1017                                 path->slots[0]++;
1018                                 continue;
1019                         }
1020                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1021                                               struct btrfs_extent_ref_v0);
1022                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1023                         break;
1024                 }
1025         }
1026         btrfs_release_path(path);
1027
1028         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1029                 new_size += sizeof(*bi);
1030
1031         new_size -= sizeof(*ei0);
1032         ret = btrfs_search_slot(trans, root, &key, path,
1033                                 new_size + extra_size, 1);
1034         if (ret < 0)
1035                 return ret;
1036         BUG_ON(ret); /* Corruption */
1037
1038         btrfs_extend_item(root, path, new_size);
1039
1040         leaf = path->nodes[0];
1041         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1042         btrfs_set_extent_refs(leaf, item, refs);
1043         /* FIXME: get real generation */
1044         btrfs_set_extent_generation(leaf, item, 0);
1045         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1046                 btrfs_set_extent_flags(leaf, item,
1047                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1048                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1049                 bi = (struct btrfs_tree_block_info *)(item + 1);
1050                 /* FIXME: get first key of the block */
1051                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1052                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1053         } else {
1054                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1055         }
1056         btrfs_mark_buffer_dirty(leaf);
1057         return 0;
1058 }
1059 #endif
1060
1061 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1062 {
1063         u32 high_crc = ~(u32)0;
1064         u32 low_crc = ~(u32)0;
1065         __le64 lenum;
1066
1067         lenum = cpu_to_le64(root_objectid);
1068         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1069         lenum = cpu_to_le64(owner);
1070         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1071         lenum = cpu_to_le64(offset);
1072         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1073
1074         return ((u64)high_crc << 31) ^ (u64)low_crc;
1075 }
1076
1077 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1078                                      struct btrfs_extent_data_ref *ref)
1079 {
1080         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1081                                     btrfs_extent_data_ref_objectid(leaf, ref),
1082                                     btrfs_extent_data_ref_offset(leaf, ref));
1083 }
1084
1085 static int match_extent_data_ref(struct extent_buffer *leaf,
1086                                  struct btrfs_extent_data_ref *ref,
1087                                  u64 root_objectid, u64 owner, u64 offset)
1088 {
1089         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1090             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1091             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1092                 return 0;
1093         return 1;
1094 }
1095
1096 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1097                                            struct btrfs_root *root,
1098                                            struct btrfs_path *path,
1099                                            u64 bytenr, u64 parent,
1100                                            u64 root_objectid,
1101                                            u64 owner, u64 offset)
1102 {
1103         struct btrfs_key key;
1104         struct btrfs_extent_data_ref *ref;
1105         struct extent_buffer *leaf;
1106         u32 nritems;
1107         int ret;
1108         int recow;
1109         int err = -ENOENT;
1110
1111         key.objectid = bytenr;
1112         if (parent) {
1113                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1114                 key.offset = parent;
1115         } else {
1116                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1117                 key.offset = hash_extent_data_ref(root_objectid,
1118                                                   owner, offset);
1119         }
1120 again:
1121         recow = 0;
1122         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1123         if (ret < 0) {
1124                 err = ret;
1125                 goto fail;
1126         }
1127
1128         if (parent) {
1129                 if (!ret)
1130                         return 0;
1131 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1132                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1133                 btrfs_release_path(path);
1134                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1135                 if (ret < 0) {
1136                         err = ret;
1137                         goto fail;
1138                 }
1139                 if (!ret)
1140                         return 0;
1141 #endif
1142                 goto fail;
1143         }
1144
1145         leaf = path->nodes[0];
1146         nritems = btrfs_header_nritems(leaf);
1147         while (1) {
1148                 if (path->slots[0] >= nritems) {
1149                         ret = btrfs_next_leaf(root, path);
1150                         if (ret < 0)
1151                                 err = ret;
1152                         if (ret)
1153                                 goto fail;
1154
1155                         leaf = path->nodes[0];
1156                         nritems = btrfs_header_nritems(leaf);
1157                         recow = 1;
1158                 }
1159
1160                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1161                 if (key.objectid != bytenr ||
1162                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1163                         goto fail;
1164
1165                 ref = btrfs_item_ptr(leaf, path->slots[0],
1166                                      struct btrfs_extent_data_ref);
1167
1168                 if (match_extent_data_ref(leaf, ref, root_objectid,
1169                                           owner, offset)) {
1170                         if (recow) {
1171                                 btrfs_release_path(path);
1172                                 goto again;
1173                         }
1174                         err = 0;
1175                         break;
1176                 }
1177                 path->slots[0]++;
1178         }
1179 fail:
1180         return err;
1181 }
1182
1183 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1184                                            struct btrfs_root *root,
1185                                            struct btrfs_path *path,
1186                                            u64 bytenr, u64 parent,
1187                                            u64 root_objectid, u64 owner,
1188                                            u64 offset, int refs_to_add)
1189 {
1190         struct btrfs_key key;
1191         struct extent_buffer *leaf;
1192         u32 size;
1193         u32 num_refs;
1194         int ret;
1195
1196         key.objectid = bytenr;
1197         if (parent) {
1198                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1199                 key.offset = parent;
1200                 size = sizeof(struct btrfs_shared_data_ref);
1201         } else {
1202                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1203                 key.offset = hash_extent_data_ref(root_objectid,
1204                                                   owner, offset);
1205                 size = sizeof(struct btrfs_extent_data_ref);
1206         }
1207
1208         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1209         if (ret && ret != -EEXIST)
1210                 goto fail;
1211
1212         leaf = path->nodes[0];
1213         if (parent) {
1214                 struct btrfs_shared_data_ref *ref;
1215                 ref = btrfs_item_ptr(leaf, path->slots[0],
1216                                      struct btrfs_shared_data_ref);
1217                 if (ret == 0) {
1218                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1219                 } else {
1220                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1221                         num_refs += refs_to_add;
1222                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1223                 }
1224         } else {
1225                 struct btrfs_extent_data_ref *ref;
1226                 while (ret == -EEXIST) {
1227                         ref = btrfs_item_ptr(leaf, path->slots[0],
1228                                              struct btrfs_extent_data_ref);
1229                         if (match_extent_data_ref(leaf, ref, root_objectid,
1230                                                   owner, offset))
1231                                 break;
1232                         btrfs_release_path(path);
1233                         key.offset++;
1234                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1235                                                       size);
1236                         if (ret && ret != -EEXIST)
1237                                 goto fail;
1238
1239                         leaf = path->nodes[0];
1240                 }
1241                 ref = btrfs_item_ptr(leaf, path->slots[0],
1242                                      struct btrfs_extent_data_ref);
1243                 if (ret == 0) {
1244                         btrfs_set_extent_data_ref_root(leaf, ref,
1245                                                        root_objectid);
1246                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1247                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1248                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1249                 } else {
1250                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1251                         num_refs += refs_to_add;
1252                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1253                 }
1254         }
1255         btrfs_mark_buffer_dirty(leaf);
1256         ret = 0;
1257 fail:
1258         btrfs_release_path(path);
1259         return ret;
1260 }
1261
1262 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1263                                            struct btrfs_root *root,
1264                                            struct btrfs_path *path,
1265                                            int refs_to_drop, int *last_ref)
1266 {
1267         struct btrfs_key key;
1268         struct btrfs_extent_data_ref *ref1 = NULL;
1269         struct btrfs_shared_data_ref *ref2 = NULL;
1270         struct extent_buffer *leaf;
1271         u32 num_refs = 0;
1272         int ret = 0;
1273
1274         leaf = path->nodes[0];
1275         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1276
1277         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1278                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1279                                       struct btrfs_extent_data_ref);
1280                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1281         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1282                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1283                                       struct btrfs_shared_data_ref);
1284                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1285 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1286         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1287                 struct btrfs_extent_ref_v0 *ref0;
1288                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1289                                       struct btrfs_extent_ref_v0);
1290                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1291 #endif
1292         } else {
1293                 BUG();
1294         }
1295
1296         BUG_ON(num_refs < refs_to_drop);
1297         num_refs -= refs_to_drop;
1298
1299         if (num_refs == 0) {
1300                 ret = btrfs_del_item(trans, root, path);
1301                 *last_ref = 1;
1302         } else {
1303                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1304                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1305                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1306                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1307 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1308                 else {
1309                         struct btrfs_extent_ref_v0 *ref0;
1310                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1311                                         struct btrfs_extent_ref_v0);
1312                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1313                 }
1314 #endif
1315                 btrfs_mark_buffer_dirty(leaf);
1316         }
1317         return ret;
1318 }
1319
1320 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1321                                           struct btrfs_path *path,
1322                                           struct btrfs_extent_inline_ref *iref)
1323 {
1324         struct btrfs_key key;
1325         struct extent_buffer *leaf;
1326         struct btrfs_extent_data_ref *ref1;
1327         struct btrfs_shared_data_ref *ref2;
1328         u32 num_refs = 0;
1329
1330         leaf = path->nodes[0];
1331         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1332         if (iref) {
1333                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1334                     BTRFS_EXTENT_DATA_REF_KEY) {
1335                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1336                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1337                 } else {
1338                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1339                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1340                 }
1341         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1342                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1343                                       struct btrfs_extent_data_ref);
1344                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1345         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1346                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1347                                       struct btrfs_shared_data_ref);
1348                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1349 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1350         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1351                 struct btrfs_extent_ref_v0 *ref0;
1352                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1353                                       struct btrfs_extent_ref_v0);
1354                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1355 #endif
1356         } else {
1357                 WARN_ON(1);
1358         }
1359         return num_refs;
1360 }
1361
1362 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1363                                           struct btrfs_root *root,
1364                                           struct btrfs_path *path,
1365                                           u64 bytenr, u64 parent,
1366                                           u64 root_objectid)
1367 {
1368         struct btrfs_key key;
1369         int ret;
1370
1371         key.objectid = bytenr;
1372         if (parent) {
1373                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1374                 key.offset = parent;
1375         } else {
1376                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1377                 key.offset = root_objectid;
1378         }
1379
1380         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1381         if (ret > 0)
1382                 ret = -ENOENT;
1383 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1384         if (ret == -ENOENT && parent) {
1385                 btrfs_release_path(path);
1386                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1387                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1388                 if (ret > 0)
1389                         ret = -ENOENT;
1390         }
1391 #endif
1392         return ret;
1393 }
1394
1395 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1396                                           struct btrfs_root *root,
1397                                           struct btrfs_path *path,
1398                                           u64 bytenr, u64 parent,
1399                                           u64 root_objectid)
1400 {
1401         struct btrfs_key key;
1402         int ret;
1403
1404         key.objectid = bytenr;
1405         if (parent) {
1406                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1407                 key.offset = parent;
1408         } else {
1409                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1410                 key.offset = root_objectid;
1411         }
1412
1413         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1414         btrfs_release_path(path);
1415         return ret;
1416 }
1417
1418 static inline int extent_ref_type(u64 parent, u64 owner)
1419 {
1420         int type;
1421         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1422                 if (parent > 0)
1423                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1424                 else
1425                         type = BTRFS_TREE_BLOCK_REF_KEY;
1426         } else {
1427                 if (parent > 0)
1428                         type = BTRFS_SHARED_DATA_REF_KEY;
1429                 else
1430                         type = BTRFS_EXTENT_DATA_REF_KEY;
1431         }
1432         return type;
1433 }
1434
1435 static int find_next_key(struct btrfs_path *path, int level,
1436                          struct btrfs_key *key)
1437
1438 {
1439         for (; level < BTRFS_MAX_LEVEL; level++) {
1440                 if (!path->nodes[level])
1441                         break;
1442                 if (path->slots[level] + 1 >=
1443                     btrfs_header_nritems(path->nodes[level]))
1444                         continue;
1445                 if (level == 0)
1446                         btrfs_item_key_to_cpu(path->nodes[level], key,
1447                                               path->slots[level] + 1);
1448                 else
1449                         btrfs_node_key_to_cpu(path->nodes[level], key,
1450                                               path->slots[level] + 1);
1451                 return 0;
1452         }
1453         return 1;
1454 }
1455
1456 /*
1457  * look for inline back ref. if back ref is found, *ref_ret is set
1458  * to the address of inline back ref, and 0 is returned.
1459  *
1460  * if back ref isn't found, *ref_ret is set to the address where it
1461  * should be inserted, and -ENOENT is returned.
1462  *
1463  * if insert is true and there are too many inline back refs, the path
1464  * points to the extent item, and -EAGAIN is returned.
1465  *
1466  * NOTE: inline back refs are ordered in the same way that back ref
1467  *       items in the tree are ordered.
1468  */
1469 static noinline_for_stack
1470 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1471                                  struct btrfs_root *root,
1472                                  struct btrfs_path *path,
1473                                  struct btrfs_extent_inline_ref **ref_ret,
1474                                  u64 bytenr, u64 num_bytes,
1475                                  u64 parent, u64 root_objectid,
1476                                  u64 owner, u64 offset, int insert)
1477 {
1478         struct btrfs_key key;
1479         struct extent_buffer *leaf;
1480         struct btrfs_extent_item *ei;
1481         struct btrfs_extent_inline_ref *iref;
1482         u64 flags;
1483         u64 item_size;
1484         unsigned long ptr;
1485         unsigned long end;
1486         int extra_size;
1487         int type;
1488         int want;
1489         int ret;
1490         int err = 0;
1491         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1492                                                  SKINNY_METADATA);
1493
1494         key.objectid = bytenr;
1495         key.type = BTRFS_EXTENT_ITEM_KEY;
1496         key.offset = num_bytes;
1497
1498         want = extent_ref_type(parent, owner);
1499         if (insert) {
1500                 extra_size = btrfs_extent_inline_ref_size(want);
1501                 path->keep_locks = 1;
1502         } else
1503                 extra_size = -1;
1504
1505         /*
1506          * Owner is our parent level, so we can just add one to get the level
1507          * for the block we are interested in.
1508          */
1509         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1510                 key.type = BTRFS_METADATA_ITEM_KEY;
1511                 key.offset = owner;
1512         }
1513
1514 again:
1515         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1516         if (ret < 0) {
1517                 err = ret;
1518                 goto out;
1519         }
1520
1521         /*
1522          * We may be a newly converted file system which still has the old fat
1523          * extent entries for metadata, so try and see if we have one of those.
1524          */
1525         if (ret > 0 && skinny_metadata) {
1526                 skinny_metadata = false;
1527                 if (path->slots[0]) {
1528                         path->slots[0]--;
1529                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1530                                               path->slots[0]);
1531                         if (key.objectid == bytenr &&
1532                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1533                             key.offset == num_bytes)
1534                                 ret = 0;
1535                 }
1536                 if (ret) {
1537                         key.objectid = bytenr;
1538                         key.type = BTRFS_EXTENT_ITEM_KEY;
1539                         key.offset = num_bytes;
1540                         btrfs_release_path(path);
1541                         goto again;
1542                 }
1543         }
1544
1545         if (ret && !insert) {
1546                 err = -ENOENT;
1547                 goto out;
1548         } else if (WARN_ON(ret)) {
1549                 err = -EIO;
1550                 goto out;
1551         }
1552
1553         leaf = path->nodes[0];
1554         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1555 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1556         if (item_size < sizeof(*ei)) {
1557                 if (!insert) {
1558                         err = -ENOENT;
1559                         goto out;
1560                 }
1561                 ret = convert_extent_item_v0(trans, root, path, owner,
1562                                              extra_size);
1563                 if (ret < 0) {
1564                         err = ret;
1565                         goto out;
1566                 }
1567                 leaf = path->nodes[0];
1568                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1569         }
1570 #endif
1571         BUG_ON(item_size < sizeof(*ei));
1572
1573         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1574         flags = btrfs_extent_flags(leaf, ei);
1575
1576         ptr = (unsigned long)(ei + 1);
1577         end = (unsigned long)ei + item_size;
1578
1579         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1580                 ptr += sizeof(struct btrfs_tree_block_info);
1581                 BUG_ON(ptr > end);
1582         }
1583
1584         err = -ENOENT;
1585         while (1) {
1586                 if (ptr >= end) {
1587                         WARN_ON(ptr > end);
1588                         break;
1589                 }
1590                 iref = (struct btrfs_extent_inline_ref *)ptr;
1591                 type = btrfs_extent_inline_ref_type(leaf, iref);
1592                 if (want < type)
1593                         break;
1594                 if (want > type) {
1595                         ptr += btrfs_extent_inline_ref_size(type);
1596                         continue;
1597                 }
1598
1599                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1600                         struct btrfs_extent_data_ref *dref;
1601                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1602                         if (match_extent_data_ref(leaf, dref, root_objectid,
1603                                                   owner, offset)) {
1604                                 err = 0;
1605                                 break;
1606                         }
1607                         if (hash_extent_data_ref_item(leaf, dref) <
1608                             hash_extent_data_ref(root_objectid, owner, offset))
1609                                 break;
1610                 } else {
1611                         u64 ref_offset;
1612                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1613                         if (parent > 0) {
1614                                 if (parent == ref_offset) {
1615                                         err = 0;
1616                                         break;
1617                                 }
1618                                 if (ref_offset < parent)
1619                                         break;
1620                         } else {
1621                                 if (root_objectid == ref_offset) {
1622                                         err = 0;
1623                                         break;
1624                                 }
1625                                 if (ref_offset < root_objectid)
1626                                         break;
1627                         }
1628                 }
1629                 ptr += btrfs_extent_inline_ref_size(type);
1630         }
1631         if (err == -ENOENT && insert) {
1632                 if (item_size + extra_size >=
1633                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1634                         err = -EAGAIN;
1635                         goto out;
1636                 }
1637                 /*
1638                  * To add new inline back ref, we have to make sure
1639                  * there is no corresponding back ref item.
1640                  * For simplicity, we just do not add new inline back
1641                  * ref if there is any kind of item for this block
1642                  */
1643                 if (find_next_key(path, 0, &key) == 0 &&
1644                     key.objectid == bytenr &&
1645                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1646                         err = -EAGAIN;
1647                         goto out;
1648                 }
1649         }
1650         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1651 out:
1652         if (insert) {
1653                 path->keep_locks = 0;
1654                 btrfs_unlock_up_safe(path, 1);
1655         }
1656         return err;
1657 }
1658
1659 /*
1660  * helper to add new inline back ref
1661  */
1662 static noinline_for_stack
1663 void setup_inline_extent_backref(struct btrfs_root *root,
1664                                  struct btrfs_path *path,
1665                                  struct btrfs_extent_inline_ref *iref,
1666                                  u64 parent, u64 root_objectid,
1667                                  u64 owner, u64 offset, int refs_to_add,
1668                                  struct btrfs_delayed_extent_op *extent_op)
1669 {
1670         struct extent_buffer *leaf;
1671         struct btrfs_extent_item *ei;
1672         unsigned long ptr;
1673         unsigned long end;
1674         unsigned long item_offset;
1675         u64 refs;
1676         int size;
1677         int type;
1678
1679         leaf = path->nodes[0];
1680         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1681         item_offset = (unsigned long)iref - (unsigned long)ei;
1682
1683         type = extent_ref_type(parent, owner);
1684         size = btrfs_extent_inline_ref_size(type);
1685
1686         btrfs_extend_item(root, path, size);
1687
1688         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1689         refs = btrfs_extent_refs(leaf, ei);
1690         refs += refs_to_add;
1691         btrfs_set_extent_refs(leaf, ei, refs);
1692         if (extent_op)
1693                 __run_delayed_extent_op(extent_op, leaf, ei);
1694
1695         ptr = (unsigned long)ei + item_offset;
1696         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1697         if (ptr < end - size)
1698                 memmove_extent_buffer(leaf, ptr + size, ptr,
1699                                       end - size - ptr);
1700
1701         iref = (struct btrfs_extent_inline_ref *)ptr;
1702         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1703         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1704                 struct btrfs_extent_data_ref *dref;
1705                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1706                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1707                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1708                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1709                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1710         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1711                 struct btrfs_shared_data_ref *sref;
1712                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1713                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1714                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1715         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1716                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1717         } else {
1718                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1719         }
1720         btrfs_mark_buffer_dirty(leaf);
1721 }
1722
1723 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1724                                  struct btrfs_root *root,
1725                                  struct btrfs_path *path,
1726                                  struct btrfs_extent_inline_ref **ref_ret,
1727                                  u64 bytenr, u64 num_bytes, u64 parent,
1728                                  u64 root_objectid, u64 owner, u64 offset)
1729 {
1730         int ret;
1731
1732         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1733                                            bytenr, num_bytes, parent,
1734                                            root_objectid, owner, offset, 0);
1735         if (ret != -ENOENT)
1736                 return ret;
1737
1738         btrfs_release_path(path);
1739         *ref_ret = NULL;
1740
1741         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1742                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1743                                             root_objectid);
1744         } else {
1745                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1746                                              root_objectid, owner, offset);
1747         }
1748         return ret;
1749 }
1750
1751 /*
1752  * helper to update/remove inline back ref
1753  */
1754 static noinline_for_stack
1755 void update_inline_extent_backref(struct btrfs_root *root,
1756                                   struct btrfs_path *path,
1757                                   struct btrfs_extent_inline_ref *iref,
1758                                   int refs_to_mod,
1759                                   struct btrfs_delayed_extent_op *extent_op,
1760                                   int *last_ref)
1761 {
1762         struct extent_buffer *leaf;
1763         struct btrfs_extent_item *ei;
1764         struct btrfs_extent_data_ref *dref = NULL;
1765         struct btrfs_shared_data_ref *sref = NULL;
1766         unsigned long ptr;
1767         unsigned long end;
1768         u32 item_size;
1769         int size;
1770         int type;
1771         u64 refs;
1772
1773         leaf = path->nodes[0];
1774         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1775         refs = btrfs_extent_refs(leaf, ei);
1776         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1777         refs += refs_to_mod;
1778         btrfs_set_extent_refs(leaf, ei, refs);
1779         if (extent_op)
1780                 __run_delayed_extent_op(extent_op, leaf, ei);
1781
1782         type = btrfs_extent_inline_ref_type(leaf, iref);
1783
1784         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1785                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1786                 refs = btrfs_extent_data_ref_count(leaf, dref);
1787         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1788                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1789                 refs = btrfs_shared_data_ref_count(leaf, sref);
1790         } else {
1791                 refs = 1;
1792                 BUG_ON(refs_to_mod != -1);
1793         }
1794
1795         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1796         refs += refs_to_mod;
1797
1798         if (refs > 0) {
1799                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1800                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1801                 else
1802                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1803         } else {
1804                 *last_ref = 1;
1805                 size =  btrfs_extent_inline_ref_size(type);
1806                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1807                 ptr = (unsigned long)iref;
1808                 end = (unsigned long)ei + item_size;
1809                 if (ptr + size < end)
1810                         memmove_extent_buffer(leaf, ptr, ptr + size,
1811                                               end - ptr - size);
1812                 item_size -= size;
1813                 btrfs_truncate_item(root, path, item_size, 1);
1814         }
1815         btrfs_mark_buffer_dirty(leaf);
1816 }
1817
1818 static noinline_for_stack
1819 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1820                                  struct btrfs_root *root,
1821                                  struct btrfs_path *path,
1822                                  u64 bytenr, u64 num_bytes, u64 parent,
1823                                  u64 root_objectid, u64 owner,
1824                                  u64 offset, int refs_to_add,
1825                                  struct btrfs_delayed_extent_op *extent_op)
1826 {
1827         struct btrfs_extent_inline_ref *iref;
1828         int ret;
1829
1830         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1831                                            bytenr, num_bytes, parent,
1832                                            root_objectid, owner, offset, 1);
1833         if (ret == 0) {
1834                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1835                 update_inline_extent_backref(root, path, iref,
1836                                              refs_to_add, extent_op, NULL);
1837         } else if (ret == -ENOENT) {
1838                 setup_inline_extent_backref(root, path, iref, parent,
1839                                             root_objectid, owner, offset,
1840                                             refs_to_add, extent_op);
1841                 ret = 0;
1842         }
1843         return ret;
1844 }
1845
1846 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1847                                  struct btrfs_root *root,
1848                                  struct btrfs_path *path,
1849                                  u64 bytenr, u64 parent, u64 root_objectid,
1850                                  u64 owner, u64 offset, int refs_to_add)
1851 {
1852         int ret;
1853         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1854                 BUG_ON(refs_to_add != 1);
1855                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1856                                             parent, root_objectid);
1857         } else {
1858                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1859                                              parent, root_objectid,
1860                                              owner, offset, refs_to_add);
1861         }
1862         return ret;
1863 }
1864
1865 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1866                                  struct btrfs_root *root,
1867                                  struct btrfs_path *path,
1868                                  struct btrfs_extent_inline_ref *iref,
1869                                  int refs_to_drop, int is_data, int *last_ref)
1870 {
1871         int ret = 0;
1872
1873         BUG_ON(!is_data && refs_to_drop != 1);
1874         if (iref) {
1875                 update_inline_extent_backref(root, path, iref,
1876                                              -refs_to_drop, NULL, last_ref);
1877         } else if (is_data) {
1878                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1879                                              last_ref);
1880         } else {
1881                 *last_ref = 1;
1882                 ret = btrfs_del_item(trans, root, path);
1883         }
1884         return ret;
1885 }
1886
1887 static int btrfs_issue_discard(struct block_device *bdev,
1888                                 u64 start, u64 len)
1889 {
1890         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1891 }
1892
1893 int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1894                          u64 num_bytes, u64 *actual_bytes)
1895 {
1896         int ret;
1897         u64 discarded_bytes = 0;
1898         struct btrfs_bio *bbio = NULL;
1899
1900
1901         /* Tell the block device(s) that the sectors can be discarded */
1902         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1903                               bytenr, &num_bytes, &bbio, 0);
1904         /* Error condition is -ENOMEM */
1905         if (!ret) {
1906                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1907                 int i;
1908
1909
1910                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1911                         if (!stripe->dev->can_discard)
1912                                 continue;
1913
1914                         ret = btrfs_issue_discard(stripe->dev->bdev,
1915                                                   stripe->physical,
1916                                                   stripe->length);
1917                         if (!ret)
1918                                 discarded_bytes += stripe->length;
1919                         else if (ret != -EOPNOTSUPP)
1920                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1921
1922                         /*
1923                          * Just in case we get back EOPNOTSUPP for some reason,
1924                          * just ignore the return value so we don't screw up
1925                          * people calling discard_extent.
1926                          */
1927                         ret = 0;
1928                 }
1929                 btrfs_put_bbio(bbio);
1930         }
1931
1932         if (actual_bytes)
1933                 *actual_bytes = discarded_bytes;
1934
1935
1936         if (ret == -EOPNOTSUPP)
1937                 ret = 0;
1938         return ret;
1939 }
1940
1941 /* Can return -ENOMEM */
1942 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1943                          struct btrfs_root *root,
1944                          u64 bytenr, u64 num_bytes, u64 parent,
1945                          u64 root_objectid, u64 owner, u64 offset,
1946                          int no_quota)
1947 {
1948         int ret;
1949         struct btrfs_fs_info *fs_info = root->fs_info;
1950
1951         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1952                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1953
1954         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1955                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1956                                         num_bytes,
1957                                         parent, root_objectid, (int)owner,
1958                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1959         } else {
1960                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1961                                         num_bytes,
1962                                         parent, root_objectid, owner, offset,
1963                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1964         }
1965         return ret;
1966 }
1967
1968 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1969                                   struct btrfs_root *root,
1970                                   u64 bytenr, u64 num_bytes,
1971                                   u64 parent, u64 root_objectid,
1972                                   u64 owner, u64 offset, int refs_to_add,
1973                                   int no_quota,
1974                                   struct btrfs_delayed_extent_op *extent_op)
1975 {
1976         struct btrfs_fs_info *fs_info = root->fs_info;
1977         struct btrfs_path *path;
1978         struct extent_buffer *leaf;
1979         struct btrfs_extent_item *item;
1980         struct btrfs_key key;
1981         u64 refs;
1982         int ret;
1983         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_ADD_EXCL;
1984
1985         path = btrfs_alloc_path();
1986         if (!path)
1987                 return -ENOMEM;
1988
1989         if (!is_fstree(root_objectid) || !root->fs_info->quota_enabled)
1990                 no_quota = 1;
1991
1992         path->reada = 1;
1993         path->leave_spinning = 1;
1994         /* this will setup the path even if it fails to insert the back ref */
1995         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
1996                                            bytenr, num_bytes, parent,
1997                                            root_objectid, owner, offset,
1998                                            refs_to_add, extent_op);
1999         if ((ret < 0 && ret != -EAGAIN) || (!ret && no_quota))
2000                 goto out;
2001         /*
2002          * Ok we were able to insert an inline extent and it appears to be a new
2003          * reference, deal with the qgroup accounting.
2004          */
2005         if (!ret && !no_quota) {
2006                 ASSERT(root->fs_info->quota_enabled);
2007                 leaf = path->nodes[0];
2008                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2009                 item = btrfs_item_ptr(leaf, path->slots[0],
2010                                       struct btrfs_extent_item);
2011                 if (btrfs_extent_refs(leaf, item) > (u64)refs_to_add)
2012                         type = BTRFS_QGROUP_OPER_ADD_SHARED;
2013                 btrfs_release_path(path);
2014
2015                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2016                                               bytenr, num_bytes, type, 0);
2017                 goto out;
2018         }
2019
2020         /*
2021          * Ok we had -EAGAIN which means we didn't have space to insert and
2022          * inline extent ref, so just update the reference count and add a
2023          * normal backref.
2024          */
2025         leaf = path->nodes[0];
2026         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2027         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2028         refs = btrfs_extent_refs(leaf, item);
2029         if (refs)
2030                 type = BTRFS_QGROUP_OPER_ADD_SHARED;
2031         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2032         if (extent_op)
2033                 __run_delayed_extent_op(extent_op, leaf, item);
2034
2035         btrfs_mark_buffer_dirty(leaf);
2036         btrfs_release_path(path);
2037
2038         if (!no_quota) {
2039                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2040                                               bytenr, num_bytes, type, 0);
2041                 if (ret)
2042                         goto out;
2043         }
2044
2045         path->reada = 1;
2046         path->leave_spinning = 1;
2047         /* now insert the actual backref */
2048         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2049                                     path, bytenr, parent, root_objectid,
2050                                     owner, offset, refs_to_add);
2051         if (ret)
2052                 btrfs_abort_transaction(trans, root, ret);
2053 out:
2054         btrfs_free_path(path);
2055         return ret;
2056 }
2057
2058 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2059                                 struct btrfs_root *root,
2060                                 struct btrfs_delayed_ref_node *node,
2061                                 struct btrfs_delayed_extent_op *extent_op,
2062                                 int insert_reserved)
2063 {
2064         int ret = 0;
2065         struct btrfs_delayed_data_ref *ref;
2066         struct btrfs_key ins;
2067         u64 parent = 0;
2068         u64 ref_root = 0;
2069         u64 flags = 0;
2070
2071         ins.objectid = node->bytenr;
2072         ins.offset = node->num_bytes;
2073         ins.type = BTRFS_EXTENT_ITEM_KEY;
2074
2075         ref = btrfs_delayed_node_to_data_ref(node);
2076         trace_run_delayed_data_ref(node, ref, node->action);
2077
2078         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2079                 parent = ref->parent;
2080         ref_root = ref->root;
2081
2082         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2083                 if (extent_op)
2084                         flags |= extent_op->flags_to_set;
2085                 ret = alloc_reserved_file_extent(trans, root,
2086                                                  parent, ref_root, flags,
2087                                                  ref->objectid, ref->offset,
2088                                                  &ins, node->ref_mod);
2089         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2090                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2091                                              node->num_bytes, parent,
2092                                              ref_root, ref->objectid,
2093                                              ref->offset, node->ref_mod,
2094                                              node->no_quota, extent_op);
2095         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2096                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2097                                           node->num_bytes, parent,
2098                                           ref_root, ref->objectid,
2099                                           ref->offset, node->ref_mod,
2100                                           extent_op, node->no_quota);
2101         } else {
2102                 BUG();
2103         }
2104         return ret;
2105 }
2106
2107 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2108                                     struct extent_buffer *leaf,
2109                                     struct btrfs_extent_item *ei)
2110 {
2111         u64 flags = btrfs_extent_flags(leaf, ei);
2112         if (extent_op->update_flags) {
2113                 flags |= extent_op->flags_to_set;
2114                 btrfs_set_extent_flags(leaf, ei, flags);
2115         }
2116
2117         if (extent_op->update_key) {
2118                 struct btrfs_tree_block_info *bi;
2119                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2120                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2121                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2122         }
2123 }
2124
2125 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2126                                  struct btrfs_root *root,
2127                                  struct btrfs_delayed_ref_node *node,
2128                                  struct btrfs_delayed_extent_op *extent_op)
2129 {
2130         struct btrfs_key key;
2131         struct btrfs_path *path;
2132         struct btrfs_extent_item *ei;
2133         struct extent_buffer *leaf;
2134         u32 item_size;
2135         int ret;
2136         int err = 0;
2137         int metadata = !extent_op->is_data;
2138
2139         if (trans->aborted)
2140                 return 0;
2141
2142         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2143                 metadata = 0;
2144
2145         path = btrfs_alloc_path();
2146         if (!path)
2147                 return -ENOMEM;
2148
2149         key.objectid = node->bytenr;
2150
2151         if (metadata) {
2152                 key.type = BTRFS_METADATA_ITEM_KEY;
2153                 key.offset = extent_op->level;
2154         } else {
2155                 key.type = BTRFS_EXTENT_ITEM_KEY;
2156                 key.offset = node->num_bytes;
2157         }
2158
2159 again:
2160         path->reada = 1;
2161         path->leave_spinning = 1;
2162         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2163                                 path, 0, 1);
2164         if (ret < 0) {
2165                 err = ret;
2166                 goto out;
2167         }
2168         if (ret > 0) {
2169                 if (metadata) {
2170                         if (path->slots[0] > 0) {
2171                                 path->slots[0]--;
2172                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2173                                                       path->slots[0]);
2174                                 if (key.objectid == node->bytenr &&
2175                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2176                                     key.offset == node->num_bytes)
2177                                         ret = 0;
2178                         }
2179                         if (ret > 0) {
2180                                 btrfs_release_path(path);
2181                                 metadata = 0;
2182
2183                                 key.objectid = node->bytenr;
2184                                 key.offset = node->num_bytes;
2185                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2186                                 goto again;
2187                         }
2188                 } else {
2189                         err = -EIO;
2190                         goto out;
2191                 }
2192         }
2193
2194         leaf = path->nodes[0];
2195         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2196 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2197         if (item_size < sizeof(*ei)) {
2198                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2199                                              path, (u64)-1, 0);
2200                 if (ret < 0) {
2201                         err = ret;
2202                         goto out;
2203                 }
2204                 leaf = path->nodes[0];
2205                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2206         }
2207 #endif
2208         BUG_ON(item_size < sizeof(*ei));
2209         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2210         __run_delayed_extent_op(extent_op, leaf, ei);
2211
2212         btrfs_mark_buffer_dirty(leaf);
2213 out:
2214         btrfs_free_path(path);
2215         return err;
2216 }
2217
2218 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2219                                 struct btrfs_root *root,
2220                                 struct btrfs_delayed_ref_node *node,
2221                                 struct btrfs_delayed_extent_op *extent_op,
2222                                 int insert_reserved)
2223 {
2224         int ret = 0;
2225         struct btrfs_delayed_tree_ref *ref;
2226         struct btrfs_key ins;
2227         u64 parent = 0;
2228         u64 ref_root = 0;
2229         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2230                                                  SKINNY_METADATA);
2231
2232         ref = btrfs_delayed_node_to_tree_ref(node);
2233         trace_run_delayed_tree_ref(node, ref, node->action);
2234
2235         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2236                 parent = ref->parent;
2237         ref_root = ref->root;
2238
2239         ins.objectid = node->bytenr;
2240         if (skinny_metadata) {
2241                 ins.offset = ref->level;
2242                 ins.type = BTRFS_METADATA_ITEM_KEY;
2243         } else {
2244                 ins.offset = node->num_bytes;
2245                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2246         }
2247
2248         BUG_ON(node->ref_mod != 1);
2249         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2250                 BUG_ON(!extent_op || !extent_op->update_flags);
2251                 ret = alloc_reserved_tree_block(trans, root,
2252                                                 parent, ref_root,
2253                                                 extent_op->flags_to_set,
2254                                                 &extent_op->key,
2255                                                 ref->level, &ins,
2256                                                 node->no_quota);
2257         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2258                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2259                                              node->num_bytes, parent, ref_root,
2260                                              ref->level, 0, 1, node->no_quota,
2261                                              extent_op);
2262         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2263                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2264                                           node->num_bytes, parent, ref_root,
2265                                           ref->level, 0, 1, extent_op,
2266                                           node->no_quota);
2267         } else {
2268                 BUG();
2269         }
2270         return ret;
2271 }
2272
2273 /* helper function to actually process a single delayed ref entry */
2274 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2275                                struct btrfs_root *root,
2276                                struct btrfs_delayed_ref_node *node,
2277                                struct btrfs_delayed_extent_op *extent_op,
2278                                int insert_reserved)
2279 {
2280         int ret = 0;
2281
2282         if (trans->aborted) {
2283                 if (insert_reserved)
2284                         btrfs_pin_extent(root, node->bytenr,
2285                                          node->num_bytes, 1);
2286                 return 0;
2287         }
2288
2289         if (btrfs_delayed_ref_is_head(node)) {
2290                 struct btrfs_delayed_ref_head *head;
2291                 /*
2292                  * we've hit the end of the chain and we were supposed
2293                  * to insert this extent into the tree.  But, it got
2294                  * deleted before we ever needed to insert it, so all
2295                  * we have to do is clean up the accounting
2296                  */
2297                 BUG_ON(extent_op);
2298                 head = btrfs_delayed_node_to_head(node);
2299                 trace_run_delayed_ref_head(node, head, node->action);
2300
2301                 if (insert_reserved) {
2302                         btrfs_pin_extent(root, node->bytenr,
2303                                          node->num_bytes, 1);
2304                         if (head->is_data) {
2305                                 ret = btrfs_del_csums(trans, root,
2306                                                       node->bytenr,
2307                                                       node->num_bytes);
2308                         }
2309                 }
2310                 return ret;
2311         }
2312
2313         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2314             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2315                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2316                                            insert_reserved);
2317         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2318                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2319                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2320                                            insert_reserved);
2321         else
2322                 BUG();
2323         return ret;
2324 }
2325
2326 static noinline struct btrfs_delayed_ref_node *
2327 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2328 {
2329         struct rb_node *node;
2330         struct btrfs_delayed_ref_node *ref, *last = NULL;;
2331
2332         /*
2333          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2334          * this prevents ref count from going down to zero when
2335          * there still are pending delayed ref.
2336          */
2337         node = rb_first(&head->ref_root);
2338         while (node) {
2339                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2340                                 rb_node);
2341                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2342                         return ref;
2343                 else if (last == NULL)
2344                         last = ref;
2345                 node = rb_next(node);
2346         }
2347         return last;
2348 }
2349
2350 /*
2351  * Returns 0 on success or if called with an already aborted transaction.
2352  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2353  */
2354 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2355                                              struct btrfs_root *root,
2356                                              unsigned long nr)
2357 {
2358         struct btrfs_delayed_ref_root *delayed_refs;
2359         struct btrfs_delayed_ref_node *ref;
2360         struct btrfs_delayed_ref_head *locked_ref = NULL;
2361         struct btrfs_delayed_extent_op *extent_op;
2362         struct btrfs_fs_info *fs_info = root->fs_info;
2363         ktime_t start = ktime_get();
2364         int ret;
2365         unsigned long count = 0;
2366         unsigned long actual_count = 0;
2367         int must_insert_reserved = 0;
2368
2369         delayed_refs = &trans->transaction->delayed_refs;
2370         while (1) {
2371                 if (!locked_ref) {
2372                         if (count >= nr)
2373                                 break;
2374
2375                         spin_lock(&delayed_refs->lock);
2376                         locked_ref = btrfs_select_ref_head(trans);
2377                         if (!locked_ref) {
2378                                 spin_unlock(&delayed_refs->lock);
2379                                 break;
2380                         }
2381
2382                         /* grab the lock that says we are going to process
2383                          * all the refs for this head */
2384                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2385                         spin_unlock(&delayed_refs->lock);
2386                         /*
2387                          * we may have dropped the spin lock to get the head
2388                          * mutex lock, and that might have given someone else
2389                          * time to free the head.  If that's true, it has been
2390                          * removed from our list and we can move on.
2391                          */
2392                         if (ret == -EAGAIN) {
2393                                 locked_ref = NULL;
2394                                 count++;
2395                                 continue;
2396                         }
2397                 }
2398
2399                 /*
2400                  * We need to try and merge add/drops of the same ref since we
2401                  * can run into issues with relocate dropping the implicit ref
2402                  * and then it being added back again before the drop can
2403                  * finish.  If we merged anything we need to re-loop so we can
2404                  * get a good ref.
2405                  */
2406                 spin_lock(&locked_ref->lock);
2407                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2408                                          locked_ref);
2409
2410                 /*
2411                  * locked_ref is the head node, so we have to go one
2412                  * node back for any delayed ref updates
2413                  */
2414                 ref = select_delayed_ref(locked_ref);
2415
2416                 if (ref && ref->seq &&
2417                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2418                         spin_unlock(&locked_ref->lock);
2419                         btrfs_delayed_ref_unlock(locked_ref);
2420                         spin_lock(&delayed_refs->lock);
2421                         locked_ref->processing = 0;
2422                         delayed_refs->num_heads_ready++;
2423                         spin_unlock(&delayed_refs->lock);
2424                         locked_ref = NULL;
2425                         cond_resched();
2426                         count++;
2427                         continue;
2428                 }
2429
2430                 /*
2431                  * record the must insert reserved flag before we
2432                  * drop the spin lock.
2433                  */
2434                 must_insert_reserved = locked_ref->must_insert_reserved;
2435                 locked_ref->must_insert_reserved = 0;
2436
2437                 extent_op = locked_ref->extent_op;
2438                 locked_ref->extent_op = NULL;
2439
2440                 if (!ref) {
2441
2442
2443                         /* All delayed refs have been processed, Go ahead
2444                          * and send the head node to run_one_delayed_ref,
2445                          * so that any accounting fixes can happen
2446                          */
2447                         ref = &locked_ref->node;
2448
2449                         if (extent_op && must_insert_reserved) {
2450                                 btrfs_free_delayed_extent_op(extent_op);
2451                                 extent_op = NULL;
2452                         }
2453
2454                         if (extent_op) {
2455                                 spin_unlock(&locked_ref->lock);
2456                                 ret = run_delayed_extent_op(trans, root,
2457                                                             ref, extent_op);
2458                                 btrfs_free_delayed_extent_op(extent_op);
2459
2460                                 if (ret) {
2461                                         /*
2462                                          * Need to reset must_insert_reserved if
2463                                          * there was an error so the abort stuff
2464                                          * can cleanup the reserved space
2465                                          * properly.
2466                                          */
2467                                         if (must_insert_reserved)
2468                                                 locked_ref->must_insert_reserved = 1;
2469                                         locked_ref->processing = 0;
2470                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2471                                         btrfs_delayed_ref_unlock(locked_ref);
2472                                         return ret;
2473                                 }
2474                                 continue;
2475                         }
2476
2477                         /*
2478                          * Need to drop our head ref lock and re-aqcuire the
2479                          * delayed ref lock and then re-check to make sure
2480                          * nobody got added.
2481                          */
2482                         spin_unlock(&locked_ref->lock);
2483                         spin_lock(&delayed_refs->lock);
2484                         spin_lock(&locked_ref->lock);
2485                         if (rb_first(&locked_ref->ref_root) ||
2486                             locked_ref->extent_op) {
2487                                 spin_unlock(&locked_ref->lock);
2488                                 spin_unlock(&delayed_refs->lock);
2489                                 continue;
2490                         }
2491                         ref->in_tree = 0;
2492                         delayed_refs->num_heads--;
2493                         rb_erase(&locked_ref->href_node,
2494                                  &delayed_refs->href_root);
2495                         spin_unlock(&delayed_refs->lock);
2496                 } else {
2497                         actual_count++;
2498                         ref->in_tree = 0;
2499                         rb_erase(&ref->rb_node, &locked_ref->ref_root);
2500                 }
2501                 atomic_dec(&delayed_refs->num_entries);
2502
2503                 if (!btrfs_delayed_ref_is_head(ref)) {
2504                         /*
2505                          * when we play the delayed ref, also correct the
2506                          * ref_mod on head
2507                          */
2508                         switch (ref->action) {
2509                         case BTRFS_ADD_DELAYED_REF:
2510                         case BTRFS_ADD_DELAYED_EXTENT:
2511                                 locked_ref->node.ref_mod -= ref->ref_mod;
2512                                 break;
2513                         case BTRFS_DROP_DELAYED_REF:
2514                                 locked_ref->node.ref_mod += ref->ref_mod;
2515                                 break;
2516                         default:
2517                                 WARN_ON(1);
2518                         }
2519                 }
2520                 spin_unlock(&locked_ref->lock);
2521
2522                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2523                                           must_insert_reserved);
2524
2525                 btrfs_free_delayed_extent_op(extent_op);
2526                 if (ret) {
2527                         locked_ref->processing = 0;
2528                         btrfs_delayed_ref_unlock(locked_ref);
2529                         btrfs_put_delayed_ref(ref);
2530                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2531                         return ret;
2532                 }
2533
2534                 /*
2535                  * If this node is a head, that means all the refs in this head
2536                  * have been dealt with, and we will pick the next head to deal
2537                  * with, so we must unlock the head and drop it from the cluster
2538                  * list before we release it.
2539                  */
2540                 if (btrfs_delayed_ref_is_head(ref)) {
2541                         if (locked_ref->is_data &&
2542                             locked_ref->total_ref_mod < 0) {
2543                                 spin_lock(&delayed_refs->lock);
2544                                 delayed_refs->pending_csums -= ref->num_bytes;
2545                                 spin_unlock(&delayed_refs->lock);
2546                         }
2547                         btrfs_delayed_ref_unlock(locked_ref);
2548                         locked_ref = NULL;
2549                 }
2550                 btrfs_put_delayed_ref(ref);
2551                 count++;
2552                 cond_resched();
2553         }
2554
2555         /*
2556          * We don't want to include ref heads since we can have empty ref heads
2557          * and those will drastically skew our runtime down since we just do
2558          * accounting, no actual extent tree updates.
2559          */
2560         if (actual_count > 0) {
2561                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2562                 u64 avg;
2563
2564                 /*
2565                  * We weigh the current average higher than our current runtime
2566                  * to avoid large swings in the average.
2567                  */
2568                 spin_lock(&delayed_refs->lock);
2569                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2570                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2571                 spin_unlock(&delayed_refs->lock);
2572         }
2573         return 0;
2574 }
2575
2576 #ifdef SCRAMBLE_DELAYED_REFS
2577 /*
2578  * Normally delayed refs get processed in ascending bytenr order. This
2579  * correlates in most cases to the order added. To expose dependencies on this
2580  * order, we start to process the tree in the middle instead of the beginning
2581  */
2582 static u64 find_middle(struct rb_root *root)
2583 {
2584         struct rb_node *n = root->rb_node;
2585         struct btrfs_delayed_ref_node *entry;
2586         int alt = 1;
2587         u64 middle;
2588         u64 first = 0, last = 0;
2589
2590         n = rb_first(root);
2591         if (n) {
2592                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2593                 first = entry->bytenr;
2594         }
2595         n = rb_last(root);
2596         if (n) {
2597                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2598                 last = entry->bytenr;
2599         }
2600         n = root->rb_node;
2601
2602         while (n) {
2603                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2604                 WARN_ON(!entry->in_tree);
2605
2606                 middle = entry->bytenr;
2607
2608                 if (alt)
2609                         n = n->rb_left;
2610                 else
2611                         n = n->rb_right;
2612
2613                 alt = 1 - alt;
2614         }
2615         return middle;
2616 }
2617 #endif
2618
2619 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2620 {
2621         u64 num_bytes;
2622
2623         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2624                              sizeof(struct btrfs_extent_inline_ref));
2625         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2626                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2627
2628         /*
2629          * We don't ever fill up leaves all the way so multiply by 2 just to be
2630          * closer to what we're really going to want to ouse.
2631          */
2632         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2633 }
2634
2635 /*
2636  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2637  * would require to store the csums for that many bytes.
2638  */
2639 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2640 {
2641         u64 csum_size;
2642         u64 num_csums_per_leaf;
2643         u64 num_csums;
2644
2645         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
2646         num_csums_per_leaf = div64_u64(csum_size,
2647                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2648         num_csums = div64_u64(csum_bytes, root->sectorsize);
2649         num_csums += num_csums_per_leaf - 1;
2650         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2651         return num_csums;
2652 }
2653
2654 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2655                                        struct btrfs_root *root)
2656 {
2657         struct btrfs_block_rsv *global_rsv;
2658         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2659         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2660         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2661         u64 num_bytes, num_dirty_bgs_bytes;
2662         int ret = 0;
2663
2664         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2665         num_heads = heads_to_leaves(root, num_heads);
2666         if (num_heads > 1)
2667                 num_bytes += (num_heads - 1) * root->nodesize;
2668         num_bytes <<= 1;
2669         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2670         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2671                                                              num_dirty_bgs);
2672         global_rsv = &root->fs_info->global_block_rsv;
2673
2674         /*
2675          * If we can't allocate any more chunks lets make sure we have _lots_ of
2676          * wiggle room since running delayed refs can create more delayed refs.
2677          */
2678         if (global_rsv->space_info->full) {
2679                 num_dirty_bgs_bytes <<= 1;
2680                 num_bytes <<= 1;
2681         }
2682
2683         spin_lock(&global_rsv->lock);
2684         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2685                 ret = 1;
2686         spin_unlock(&global_rsv->lock);
2687         return ret;
2688 }
2689
2690 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2691                                        struct btrfs_root *root)
2692 {
2693         struct btrfs_fs_info *fs_info = root->fs_info;
2694         u64 num_entries =
2695                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2696         u64 avg_runtime;
2697         u64 val;
2698
2699         smp_mb();
2700         avg_runtime = fs_info->avg_delayed_ref_runtime;
2701         val = num_entries * avg_runtime;
2702         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2703                 return 1;
2704         if (val >= NSEC_PER_SEC / 2)
2705                 return 2;
2706
2707         return btrfs_check_space_for_delayed_refs(trans, root);
2708 }
2709
2710 struct async_delayed_refs {
2711         struct btrfs_root *root;
2712         int count;
2713         int error;
2714         int sync;
2715         struct completion wait;
2716         struct btrfs_work work;
2717 };
2718
2719 static void delayed_ref_async_start(struct btrfs_work *work)
2720 {
2721         struct async_delayed_refs *async;
2722         struct btrfs_trans_handle *trans;
2723         int ret;
2724
2725         async = container_of(work, struct async_delayed_refs, work);
2726
2727         trans = btrfs_join_transaction(async->root);
2728         if (IS_ERR(trans)) {
2729                 async->error = PTR_ERR(trans);
2730                 goto done;
2731         }
2732
2733         /*
2734          * trans->sync means that when we call end_transaciton, we won't
2735          * wait on delayed refs
2736          */
2737         trans->sync = true;
2738         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2739         if (ret)
2740                 async->error = ret;
2741
2742         ret = btrfs_end_transaction(trans, async->root);
2743         if (ret && !async->error)
2744                 async->error = ret;
2745 done:
2746         if (async->sync)
2747                 complete(&async->wait);
2748         else
2749                 kfree(async);
2750 }
2751
2752 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2753                                  unsigned long count, int wait)
2754 {
2755         struct async_delayed_refs *async;
2756         int ret;
2757
2758         async = kmalloc(sizeof(*async), GFP_NOFS);
2759         if (!async)
2760                 return -ENOMEM;
2761
2762         async->root = root->fs_info->tree_root;
2763         async->count = count;
2764         async->error = 0;
2765         if (wait)
2766                 async->sync = 1;
2767         else
2768                 async->sync = 0;
2769         init_completion(&async->wait);
2770
2771         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2772                         delayed_ref_async_start, NULL, NULL);
2773
2774         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2775
2776         if (wait) {
2777                 wait_for_completion(&async->wait);
2778                 ret = async->error;
2779                 kfree(async);
2780                 return ret;
2781         }
2782         return 0;
2783 }
2784
2785 /*
2786  * this starts processing the delayed reference count updates and
2787  * extent insertions we have queued up so far.  count can be
2788  * 0, which means to process everything in the tree at the start
2789  * of the run (but not newly added entries), or it can be some target
2790  * number you'd like to process.
2791  *
2792  * Returns 0 on success or if called with an aborted transaction
2793  * Returns <0 on error and aborts the transaction
2794  */
2795 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2796                            struct btrfs_root *root, unsigned long count)
2797 {
2798         struct rb_node *node;
2799         struct btrfs_delayed_ref_root *delayed_refs;
2800         struct btrfs_delayed_ref_head *head;
2801         int ret;
2802         int run_all = count == (unsigned long)-1;
2803
2804         /* We'll clean this up in btrfs_cleanup_transaction */
2805         if (trans->aborted)
2806                 return 0;
2807
2808         if (root == root->fs_info->extent_root)
2809                 root = root->fs_info->tree_root;
2810
2811         delayed_refs = &trans->transaction->delayed_refs;
2812         if (count == 0)
2813                 count = atomic_read(&delayed_refs->num_entries) * 2;
2814
2815 again:
2816 #ifdef SCRAMBLE_DELAYED_REFS
2817         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2818 #endif
2819         ret = __btrfs_run_delayed_refs(trans, root, count);
2820         if (ret < 0) {
2821                 btrfs_abort_transaction(trans, root, ret);
2822                 return ret;
2823         }
2824
2825         if (run_all) {
2826                 if (!list_empty(&trans->new_bgs))
2827                         btrfs_create_pending_block_groups(trans, root);
2828
2829                 spin_lock(&delayed_refs->lock);
2830                 node = rb_first(&delayed_refs->href_root);
2831                 if (!node) {
2832                         spin_unlock(&delayed_refs->lock);
2833                         goto out;
2834                 }
2835                 count = (unsigned long)-1;
2836
2837                 while (node) {
2838                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2839                                         href_node);
2840                         if (btrfs_delayed_ref_is_head(&head->node)) {
2841                                 struct btrfs_delayed_ref_node *ref;
2842
2843                                 ref = &head->node;
2844                                 atomic_inc(&ref->refs);
2845
2846                                 spin_unlock(&delayed_refs->lock);
2847                                 /*
2848                                  * Mutex was contended, block until it's
2849                                  * released and try again
2850                                  */
2851                                 mutex_lock(&head->mutex);
2852                                 mutex_unlock(&head->mutex);
2853
2854                                 btrfs_put_delayed_ref(ref);
2855                                 cond_resched();
2856                                 goto again;
2857                         } else {
2858                                 WARN_ON(1);
2859                         }
2860                         node = rb_next(node);
2861                 }
2862                 spin_unlock(&delayed_refs->lock);
2863                 cond_resched();
2864                 goto again;
2865         }
2866 out:
2867         ret = btrfs_delayed_qgroup_accounting(trans, root->fs_info);
2868         if (ret)
2869                 return ret;
2870         assert_qgroups_uptodate(trans);
2871         return 0;
2872 }
2873
2874 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2875                                 struct btrfs_root *root,
2876                                 u64 bytenr, u64 num_bytes, u64 flags,
2877                                 int level, int is_data)
2878 {
2879         struct btrfs_delayed_extent_op *extent_op;
2880         int ret;
2881
2882         extent_op = btrfs_alloc_delayed_extent_op();
2883         if (!extent_op)
2884                 return -ENOMEM;
2885
2886         extent_op->flags_to_set = flags;
2887         extent_op->update_flags = 1;
2888         extent_op->update_key = 0;
2889         extent_op->is_data = is_data ? 1 : 0;
2890         extent_op->level = level;
2891
2892         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2893                                           num_bytes, extent_op);
2894         if (ret)
2895                 btrfs_free_delayed_extent_op(extent_op);
2896         return ret;
2897 }
2898
2899 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2900                                       struct btrfs_root *root,
2901                                       struct btrfs_path *path,
2902                                       u64 objectid, u64 offset, u64 bytenr)
2903 {
2904         struct btrfs_delayed_ref_head *head;
2905         struct btrfs_delayed_ref_node *ref;
2906         struct btrfs_delayed_data_ref *data_ref;
2907         struct btrfs_delayed_ref_root *delayed_refs;
2908         struct rb_node *node;
2909         int ret = 0;
2910
2911         delayed_refs = &trans->transaction->delayed_refs;
2912         spin_lock(&delayed_refs->lock);
2913         head = btrfs_find_delayed_ref_head(trans, bytenr);
2914         if (!head) {
2915                 spin_unlock(&delayed_refs->lock);
2916                 return 0;
2917         }
2918
2919         if (!mutex_trylock(&head->mutex)) {
2920                 atomic_inc(&head->node.refs);
2921                 spin_unlock(&delayed_refs->lock);
2922
2923                 btrfs_release_path(path);
2924
2925                 /*
2926                  * Mutex was contended, block until it's released and let
2927                  * caller try again
2928                  */
2929                 mutex_lock(&head->mutex);
2930                 mutex_unlock(&head->mutex);
2931                 btrfs_put_delayed_ref(&head->node);
2932                 return -EAGAIN;
2933         }
2934         spin_unlock(&delayed_refs->lock);
2935
2936         spin_lock(&head->lock);
2937         node = rb_first(&head->ref_root);
2938         while (node) {
2939                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2940                 node = rb_next(node);
2941
2942                 /* If it's a shared ref we know a cross reference exists */
2943                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2944                         ret = 1;
2945                         break;
2946                 }
2947
2948                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2949
2950                 /*
2951                  * If our ref doesn't match the one we're currently looking at
2952                  * then we have a cross reference.
2953                  */
2954                 if (data_ref->root != root->root_key.objectid ||
2955                     data_ref->objectid != objectid ||
2956                     data_ref->offset != offset) {
2957                         ret = 1;
2958                         break;
2959                 }
2960         }
2961         spin_unlock(&head->lock);
2962         mutex_unlock(&head->mutex);
2963         return ret;
2964 }
2965
2966 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2967                                         struct btrfs_root *root,
2968                                         struct btrfs_path *path,
2969                                         u64 objectid, u64 offset, u64 bytenr)
2970 {
2971         struct btrfs_root *extent_root = root->fs_info->extent_root;
2972         struct extent_buffer *leaf;
2973         struct btrfs_extent_data_ref *ref;
2974         struct btrfs_extent_inline_ref *iref;
2975         struct btrfs_extent_item *ei;
2976         struct btrfs_key key;
2977         u32 item_size;
2978         int ret;
2979
2980         key.objectid = bytenr;
2981         key.offset = (u64)-1;
2982         key.type = BTRFS_EXTENT_ITEM_KEY;
2983
2984         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2985         if (ret < 0)
2986                 goto out;
2987         BUG_ON(ret == 0); /* Corruption */
2988
2989         ret = -ENOENT;
2990         if (path->slots[0] == 0)
2991                 goto out;
2992
2993         path->slots[0]--;
2994         leaf = path->nodes[0];
2995         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2996
2997         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2998                 goto out;
2999
3000         ret = 1;
3001         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3002 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3003         if (item_size < sizeof(*ei)) {
3004                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3005                 goto out;
3006         }
3007 #endif
3008         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3009
3010         if (item_size != sizeof(*ei) +
3011             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3012                 goto out;
3013
3014         if (btrfs_extent_generation(leaf, ei) <=
3015             btrfs_root_last_snapshot(&root->root_item))
3016                 goto out;
3017
3018         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3019         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3020             BTRFS_EXTENT_DATA_REF_KEY)
3021                 goto out;
3022
3023         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3024         if (btrfs_extent_refs(leaf, ei) !=
3025             btrfs_extent_data_ref_count(leaf, ref) ||
3026             btrfs_extent_data_ref_root(leaf, ref) !=
3027             root->root_key.objectid ||
3028             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3029             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3030                 goto out;
3031
3032         ret = 0;
3033 out:
3034         return ret;
3035 }
3036
3037 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3038                           struct btrfs_root *root,
3039                           u64 objectid, u64 offset, u64 bytenr)
3040 {
3041         struct btrfs_path *path;
3042         int ret;
3043         int ret2;
3044
3045         path = btrfs_alloc_path();
3046         if (!path)
3047                 return -ENOENT;
3048
3049         do {
3050                 ret = check_committed_ref(trans, root, path, objectid,
3051                                           offset, bytenr);
3052                 if (ret && ret != -ENOENT)
3053                         goto out;
3054
3055                 ret2 = check_delayed_ref(trans, root, path, objectid,
3056                                          offset, bytenr);
3057         } while (ret2 == -EAGAIN);
3058
3059         if (ret2 && ret2 != -ENOENT) {
3060                 ret = ret2;
3061                 goto out;
3062         }
3063
3064         if (ret != -ENOENT || ret2 != -ENOENT)
3065                 ret = 0;
3066 out:
3067         btrfs_free_path(path);
3068         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3069                 WARN_ON(ret > 0);
3070         return ret;
3071 }
3072
3073 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3074                            struct btrfs_root *root,
3075                            struct extent_buffer *buf,
3076                            int full_backref, int inc)
3077 {
3078         u64 bytenr;
3079         u64 num_bytes;
3080         u64 parent;
3081         u64 ref_root;
3082         u32 nritems;
3083         struct btrfs_key key;
3084         struct btrfs_file_extent_item *fi;
3085         int i;
3086         int level;
3087         int ret = 0;
3088         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3089                             u64, u64, u64, u64, u64, u64, int);
3090
3091
3092         if (btrfs_test_is_dummy_root(root))
3093                 return 0;
3094
3095         ref_root = btrfs_header_owner(buf);
3096         nritems = btrfs_header_nritems(buf);
3097         level = btrfs_header_level(buf);
3098
3099         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3100                 return 0;
3101
3102         if (inc)
3103                 process_func = btrfs_inc_extent_ref;
3104         else
3105                 process_func = btrfs_free_extent;
3106
3107         if (full_backref)
3108                 parent = buf->start;
3109         else
3110                 parent = 0;
3111
3112         for (i = 0; i < nritems; i++) {
3113                 if (level == 0) {
3114                         btrfs_item_key_to_cpu(buf, &key, i);
3115                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3116                                 continue;
3117                         fi = btrfs_item_ptr(buf, i,
3118                                             struct btrfs_file_extent_item);
3119                         if (btrfs_file_extent_type(buf, fi) ==
3120                             BTRFS_FILE_EXTENT_INLINE)
3121                                 continue;
3122                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3123                         if (bytenr == 0)
3124                                 continue;
3125
3126                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3127                         key.offset -= btrfs_file_extent_offset(buf, fi);
3128                         ret = process_func(trans, root, bytenr, num_bytes,
3129                                            parent, ref_root, key.objectid,
3130                                            key.offset, 1);
3131                         if (ret)
3132                                 goto fail;
3133                 } else {
3134                         bytenr = btrfs_node_blockptr(buf, i);
3135                         num_bytes = root->nodesize;
3136                         ret = process_func(trans, root, bytenr, num_bytes,
3137                                            parent, ref_root, level - 1, 0,
3138                                            1);
3139                         if (ret)
3140                                 goto fail;
3141                 }
3142         }
3143         return 0;
3144 fail:
3145         return ret;
3146 }
3147
3148 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3149                   struct extent_buffer *buf, int full_backref)
3150 {
3151         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3152 }
3153
3154 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3155                   struct extent_buffer *buf, int full_backref)
3156 {
3157         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3158 }
3159
3160 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3161                                  struct btrfs_root *root,
3162                                  struct btrfs_path *path,
3163                                  struct btrfs_block_group_cache *cache)
3164 {
3165         int ret;
3166         struct btrfs_root *extent_root = root->fs_info->extent_root;
3167         unsigned long bi;
3168         struct extent_buffer *leaf;
3169
3170         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3171         if (ret) {
3172                 if (ret > 0)
3173                         ret = -ENOENT;
3174                 goto fail;
3175         }
3176
3177         leaf = path->nodes[0];
3178         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3179         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3180         btrfs_mark_buffer_dirty(leaf);
3181         btrfs_release_path(path);
3182 fail:
3183         if (ret)
3184                 btrfs_abort_transaction(trans, root, ret);
3185         return ret;
3186
3187 }
3188
3189 static struct btrfs_block_group_cache *
3190 next_block_group(struct btrfs_root *root,
3191                  struct btrfs_block_group_cache *cache)
3192 {
3193         struct rb_node *node;
3194
3195         spin_lock(&root->fs_info->block_group_cache_lock);
3196
3197         /* If our block group was removed, we need a full search. */
3198         if (RB_EMPTY_NODE(&cache->cache_node)) {
3199                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3200
3201                 spin_unlock(&root->fs_info->block_group_cache_lock);
3202                 btrfs_put_block_group(cache);
3203                 cache = btrfs_lookup_first_block_group(root->fs_info,
3204                                                        next_bytenr);
3205                 return cache;
3206         }
3207         node = rb_next(&cache->cache_node);
3208         btrfs_put_block_group(cache);
3209         if (node) {
3210                 cache = rb_entry(node, struct btrfs_block_group_cache,
3211                                  cache_node);
3212                 btrfs_get_block_group(cache);
3213         } else
3214                 cache = NULL;
3215         spin_unlock(&root->fs_info->block_group_cache_lock);
3216         return cache;
3217 }
3218
3219 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3220                             struct btrfs_trans_handle *trans,
3221                             struct btrfs_path *path)
3222 {
3223         struct btrfs_root *root = block_group->fs_info->tree_root;
3224         struct inode *inode = NULL;
3225         u64 alloc_hint = 0;
3226         int dcs = BTRFS_DC_ERROR;
3227         u64 num_pages = 0;
3228         int retries = 0;
3229         int ret = 0;
3230
3231         /*
3232          * If this block group is smaller than 100 megs don't bother caching the
3233          * block group.
3234          */
3235         if (block_group->key.offset < (100 * 1024 * 1024)) {
3236                 spin_lock(&block_group->lock);
3237                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3238                 spin_unlock(&block_group->lock);
3239                 return 0;
3240         }
3241
3242         if (trans->aborted)
3243                 return 0;
3244 again:
3245         inode = lookup_free_space_inode(root, block_group, path);
3246         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3247                 ret = PTR_ERR(inode);
3248                 btrfs_release_path(path);
3249                 goto out;
3250         }
3251
3252         if (IS_ERR(inode)) {
3253                 BUG_ON(retries);
3254                 retries++;
3255
3256                 if (block_group->ro)
3257                         goto out_free;
3258
3259                 ret = create_free_space_inode(root, trans, block_group, path);
3260                 if (ret)
3261                         goto out_free;
3262                 goto again;
3263         }
3264
3265         /* We've already setup this transaction, go ahead and exit */
3266         if (block_group->cache_generation == trans->transid &&
3267             i_size_read(inode)) {
3268                 dcs = BTRFS_DC_SETUP;
3269                 goto out_put;
3270         }
3271
3272         /*
3273          * We want to set the generation to 0, that way if anything goes wrong
3274          * from here on out we know not to trust this cache when we load up next
3275          * time.
3276          */
3277         BTRFS_I(inode)->generation = 0;
3278         ret = btrfs_update_inode(trans, root, inode);
3279         if (ret) {
3280                 /*
3281                  * So theoretically we could recover from this, simply set the
3282                  * super cache generation to 0 so we know to invalidate the
3283                  * cache, but then we'd have to keep track of the block groups
3284                  * that fail this way so we know we _have_ to reset this cache
3285                  * before the next commit or risk reading stale cache.  So to
3286                  * limit our exposure to horrible edge cases lets just abort the
3287                  * transaction, this only happens in really bad situations
3288                  * anyway.
3289                  */
3290                 btrfs_abort_transaction(trans, root, ret);
3291                 goto out_put;
3292         }
3293         WARN_ON(ret);
3294
3295         if (i_size_read(inode) > 0) {
3296                 ret = btrfs_check_trunc_cache_free_space(root,
3297                                         &root->fs_info->global_block_rsv);
3298                 if (ret)
3299                         goto out_put;
3300
3301                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3302                 if (ret)
3303                         goto out_put;
3304         }
3305
3306         spin_lock(&block_group->lock);
3307         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3308             !btrfs_test_opt(root, SPACE_CACHE) ||
3309             block_group->delalloc_bytes) {
3310                 /*
3311                  * don't bother trying to write stuff out _if_
3312                  * a) we're not cached,
3313                  * b) we're with nospace_cache mount option.
3314                  */
3315                 dcs = BTRFS_DC_WRITTEN;
3316                 spin_unlock(&block_group->lock);
3317                 goto out_put;
3318         }
3319         spin_unlock(&block_group->lock);
3320
3321         /*
3322          * Try to preallocate enough space based on how big the block group is.
3323          * Keep in mind this has to include any pinned space which could end up
3324          * taking up quite a bit since it's not folded into the other space
3325          * cache.
3326          */
3327         num_pages = div_u64(block_group->key.offset, 256 * 1024 * 1024);
3328         if (!num_pages)
3329                 num_pages = 1;
3330
3331         num_pages *= 16;
3332         num_pages *= PAGE_CACHE_SIZE;
3333
3334         ret = btrfs_check_data_free_space(inode, num_pages);
3335         if (ret)
3336                 goto out_put;
3337
3338         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3339                                               num_pages, num_pages,
3340                                               &alloc_hint);
3341         if (!ret)
3342                 dcs = BTRFS_DC_SETUP;
3343         btrfs_free_reserved_data_space(inode, num_pages);
3344
3345 out_put:
3346         iput(inode);
3347 out_free:
3348         btrfs_release_path(path);
3349 out:
3350         spin_lock(&block_group->lock);
3351         if (!ret && dcs == BTRFS_DC_SETUP)
3352                 block_group->cache_generation = trans->transid;
3353         block_group->disk_cache_state = dcs;
3354         spin_unlock(&block_group->lock);
3355
3356         return ret;
3357 }
3358
3359 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3360                             struct btrfs_root *root)
3361 {
3362         struct btrfs_block_group_cache *cache, *tmp;
3363         struct btrfs_transaction *cur_trans = trans->transaction;
3364         struct btrfs_path *path;
3365
3366         if (list_empty(&cur_trans->dirty_bgs) ||
3367             !btrfs_test_opt(root, SPACE_CACHE))
3368                 return 0;
3369
3370         path = btrfs_alloc_path();
3371         if (!path)
3372                 return -ENOMEM;
3373
3374         /* Could add new block groups, use _safe just in case */
3375         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3376                                  dirty_list) {
3377                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3378                         cache_save_setup(cache, trans, path);
3379         }
3380
3381         btrfs_free_path(path);
3382         return 0;
3383 }
3384
3385 /*
3386  * transaction commit does final block group cache writeback during a
3387  * critical section where nothing is allowed to change the FS.  This is
3388  * required in order for the cache to actually match the block group,
3389  * but can introduce a lot of latency into the commit.
3390  *
3391  * So, btrfs_start_dirty_block_groups is here to kick off block group
3392  * cache IO.  There's a chance we'll have to redo some of it if the
3393  * block group changes again during the commit, but it greatly reduces
3394  * the commit latency by getting rid of the easy block groups while
3395  * we're still allowing others to join the commit.
3396  */
3397 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3398                                    struct btrfs_root *root)
3399 {
3400         struct btrfs_block_group_cache *cache;
3401         struct btrfs_transaction *cur_trans = trans->transaction;
3402         int ret = 0;
3403         int should_put;
3404         struct btrfs_path *path = NULL;
3405         LIST_HEAD(dirty);
3406         struct list_head *io = &cur_trans->io_bgs;
3407         int num_started = 0;
3408         int loops = 0;
3409
3410         spin_lock(&cur_trans->dirty_bgs_lock);
3411         if (!list_empty(&cur_trans->dirty_bgs)) {
3412                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3413         }
3414         spin_unlock(&cur_trans->dirty_bgs_lock);
3415
3416 again:
3417         if (list_empty(&dirty)) {
3418                 btrfs_free_path(path);
3419                 return 0;
3420         }
3421
3422         /*
3423          * make sure all the block groups on our dirty list actually
3424          * exist
3425          */
3426         btrfs_create_pending_block_groups(trans, root);
3427
3428         if (!path) {
3429                 path = btrfs_alloc_path();
3430                 if (!path)
3431                         return -ENOMEM;
3432         }
3433
3434         while (!list_empty(&dirty)) {
3435                 cache = list_first_entry(&dirty,
3436                                          struct btrfs_block_group_cache,
3437                                          dirty_list);
3438
3439                 /*
3440                  * cache_write_mutex is here only to save us from balance
3441                  * deleting this block group while we are writing out the
3442                  * cache
3443                  */
3444                 mutex_lock(&trans->transaction->cache_write_mutex);
3445
3446                 /*
3447                  * this can happen if something re-dirties a block
3448                  * group that is already under IO.  Just wait for it to
3449                  * finish and then do it all again
3450                  */
3451                 if (!list_empty(&cache->io_list)) {
3452                         list_del_init(&cache->io_list);
3453                         btrfs_wait_cache_io(root, trans, cache,
3454                                             &cache->io_ctl, path,
3455                                             cache->key.objectid);
3456                         btrfs_put_block_group(cache);
3457                 }
3458
3459
3460                 /*
3461                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3462                  * if it should update the cache_state.  Don't delete
3463                  * until after we wait.
3464                  *
3465                  * Since we're not running in the commit critical section
3466                  * we need the dirty_bgs_lock to protect from update_block_group
3467                  */
3468                 spin_lock(&cur_trans->dirty_bgs_lock);
3469                 list_del_init(&cache->dirty_list);
3470                 spin_unlock(&cur_trans->dirty_bgs_lock);
3471
3472                 should_put = 1;
3473
3474                 cache_save_setup(cache, trans, path);
3475
3476                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3477                         cache->io_ctl.inode = NULL;
3478                         ret = btrfs_write_out_cache(root, trans, cache, path);
3479                         if (ret == 0 && cache->io_ctl.inode) {
3480                                 num_started++;
3481                                 should_put = 0;
3482
3483                                 /*
3484                                  * the cache_write_mutex is protecting
3485                                  * the io_list
3486                                  */
3487                                 list_add_tail(&cache->io_list, io);
3488                         } else {
3489                                 /*
3490                                  * if we failed to write the cache, the
3491                                  * generation will be bad and life goes on
3492                                  */
3493                                 ret = 0;
3494                         }
3495                 }
3496                 if (!ret)
3497                         ret = write_one_cache_group(trans, root, path, cache);
3498                 mutex_unlock(&trans->transaction->cache_write_mutex);
3499
3500                 /* if its not on the io list, we need to put the block group */
3501                 if (should_put)
3502                         btrfs_put_block_group(cache);
3503
3504                 if (ret)
3505                         break;
3506         }
3507
3508         /*
3509          * go through delayed refs for all the stuff we've just kicked off
3510          * and then loop back (just once)
3511          */
3512         ret = btrfs_run_delayed_refs(trans, root, 0);
3513         if (!ret && loops == 0) {
3514                 loops++;
3515                 spin_lock(&cur_trans->dirty_bgs_lock);
3516                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3517                 spin_unlock(&cur_trans->dirty_bgs_lock);
3518                 goto again;
3519         }
3520
3521         btrfs_free_path(path);
3522         return ret;
3523 }
3524
3525 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3526                                    struct btrfs_root *root)
3527 {
3528         struct btrfs_block_group_cache *cache;
3529         struct btrfs_transaction *cur_trans = trans->transaction;
3530         int ret = 0;
3531         int should_put;
3532         struct btrfs_path *path;
3533         struct list_head *io = &cur_trans->io_bgs;
3534         int num_started = 0;
3535
3536         path = btrfs_alloc_path();
3537         if (!path)
3538                 return -ENOMEM;
3539
3540         /*
3541          * We don't need the lock here since we are protected by the transaction
3542          * commit.  We want to do the cache_save_setup first and then run the
3543          * delayed refs to make sure we have the best chance at doing this all
3544          * in one shot.
3545          */
3546         while (!list_empty(&cur_trans->dirty_bgs)) {
3547                 cache = list_first_entry(&cur_trans->dirty_bgs,
3548                                          struct btrfs_block_group_cache,
3549                                          dirty_list);
3550
3551                 /*
3552                  * this can happen if cache_save_setup re-dirties a block
3553                  * group that is already under IO.  Just wait for it to
3554                  * finish and then do it all again
3555                  */
3556                 if (!list_empty(&cache->io_list)) {
3557                         list_del_init(&cache->io_list);
3558                         btrfs_wait_cache_io(root, trans, cache,
3559                                             &cache->io_ctl, path,
3560                                             cache->key.objectid);
3561                         btrfs_put_block_group(cache);
3562                 }
3563
3564                 /*
3565                  * don't remove from the dirty list until after we've waited
3566                  * on any pending IO
3567                  */
3568                 list_del_init(&cache->dirty_list);
3569                 should_put = 1;
3570
3571                 cache_save_setup(cache, trans, path);
3572
3573                 if (!ret)
3574                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3575
3576                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3577                         cache->io_ctl.inode = NULL;
3578                         ret = btrfs_write_out_cache(root, trans, cache, path);
3579                         if (ret == 0 && cache->io_ctl.inode) {
3580                                 num_started++;
3581                                 should_put = 0;
3582                                 list_add_tail(&cache->io_list, io);
3583                         } else {
3584                                 /*
3585                                  * if we failed to write the cache, the
3586                                  * generation will be bad and life goes on
3587                                  */
3588                                 ret = 0;
3589                         }
3590                 }
3591                 if (!ret)
3592                         ret = write_one_cache_group(trans, root, path, cache);
3593
3594                 /* if its not on the io list, we need to put the block group */
3595                 if (should_put)
3596                         btrfs_put_block_group(cache);
3597         }
3598
3599         while (!list_empty(io)) {
3600                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3601                                          io_list);
3602                 list_del_init(&cache->io_list);
3603                 btrfs_wait_cache_io(root, trans, cache,
3604                                     &cache->io_ctl, path, cache->key.objectid);
3605                 btrfs_put_block_group(cache);
3606         }
3607
3608         btrfs_free_path(path);
3609         return ret;
3610 }
3611
3612 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3613 {
3614         struct btrfs_block_group_cache *block_group;
3615         int readonly = 0;
3616
3617         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3618         if (!block_group || block_group->ro)
3619                 readonly = 1;
3620         if (block_group)
3621                 btrfs_put_block_group(block_group);
3622         return readonly;
3623 }
3624
3625 static const char *alloc_name(u64 flags)
3626 {
3627         switch (flags) {
3628         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3629                 return "mixed";
3630         case BTRFS_BLOCK_GROUP_METADATA:
3631                 return "metadata";
3632         case BTRFS_BLOCK_GROUP_DATA:
3633                 return "data";
3634         case BTRFS_BLOCK_GROUP_SYSTEM:
3635                 return "system";
3636         default:
3637                 WARN_ON(1);
3638                 return "invalid-combination";
3639         };
3640 }
3641
3642 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3643                              u64 total_bytes, u64 bytes_used,
3644                              struct btrfs_space_info **space_info)
3645 {
3646         struct btrfs_space_info *found;
3647         int i;
3648         int factor;
3649         int ret;
3650
3651         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3652                      BTRFS_BLOCK_GROUP_RAID10))
3653                 factor = 2;
3654         else
3655                 factor = 1;
3656
3657         found = __find_space_info(info, flags);
3658         if (found) {
3659                 spin_lock(&found->lock);
3660                 found->total_bytes += total_bytes;
3661                 found->disk_total += total_bytes * factor;
3662                 found->bytes_used += bytes_used;
3663                 found->disk_used += bytes_used * factor;
3664                 found->full = 0;
3665                 spin_unlock(&found->lock);
3666                 *space_info = found;
3667                 return 0;
3668         }
3669         found = kzalloc(sizeof(*found), GFP_NOFS);
3670         if (!found)
3671                 return -ENOMEM;
3672
3673         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3674         if (ret) {
3675                 kfree(found);
3676                 return ret;
3677         }
3678
3679         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3680                 INIT_LIST_HEAD(&found->block_groups[i]);
3681         init_rwsem(&found->groups_sem);
3682         spin_lock_init(&found->lock);
3683         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3684         found->total_bytes = total_bytes;
3685         found->disk_total = total_bytes * factor;
3686         found->bytes_used = bytes_used;
3687         found->disk_used = bytes_used * factor;
3688         found->bytes_pinned = 0;
3689         found->bytes_reserved = 0;
3690         found->bytes_readonly = 0;
3691         found->bytes_may_use = 0;
3692         found->full = 0;
3693         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3694         found->chunk_alloc = 0;
3695         found->flush = 0;
3696         init_waitqueue_head(&found->wait);
3697         INIT_LIST_HEAD(&found->ro_bgs);
3698
3699         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3700                                     info->space_info_kobj, "%s",
3701                                     alloc_name(found->flags));
3702         if (ret) {
3703                 kfree(found);
3704                 return ret;
3705         }
3706
3707         *space_info = found;
3708         list_add_rcu(&found->list, &info->space_info);
3709         if (flags & BTRFS_BLOCK_GROUP_DATA)
3710                 info->data_sinfo = found;
3711
3712         return ret;
3713 }
3714
3715 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3716 {
3717         u64 extra_flags = chunk_to_extended(flags) &
3718                                 BTRFS_EXTENDED_PROFILE_MASK;
3719
3720         write_seqlock(&fs_info->profiles_lock);
3721         if (flags & BTRFS_BLOCK_GROUP_DATA)
3722                 fs_info->avail_data_alloc_bits |= extra_flags;
3723         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3724                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3725         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3726                 fs_info->avail_system_alloc_bits |= extra_flags;
3727         write_sequnlock(&fs_info->profiles_lock);
3728 }
3729
3730 /*
3731  * returns target flags in extended format or 0 if restripe for this
3732  * chunk_type is not in progress
3733  *
3734  * should be called with either volume_mutex or balance_lock held
3735  */
3736 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3737 {
3738         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3739         u64 target = 0;
3740
3741         if (!bctl)
3742                 return 0;
3743
3744         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3745             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3746                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3747         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3748                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3749                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3750         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3751                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3752                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3753         }
3754
3755         return target;
3756 }
3757
3758 /*
3759  * @flags: available profiles in extended format (see ctree.h)
3760  *
3761  * Returns reduced profile in chunk format.  If profile changing is in
3762  * progress (either running or paused) picks the target profile (if it's
3763  * already available), otherwise falls back to plain reducing.
3764  */
3765 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3766 {
3767         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3768         u64 target;
3769         u64 tmp;
3770
3771         /*
3772          * see if restripe for this chunk_type is in progress, if so
3773          * try to reduce to the target profile
3774          */
3775         spin_lock(&root->fs_info->balance_lock);
3776         target = get_restripe_target(root->fs_info, flags);
3777         if (target) {
3778                 /* pick target profile only if it's already available */
3779                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3780                         spin_unlock(&root->fs_info->balance_lock);
3781                         return extended_to_chunk(target);
3782                 }
3783         }
3784         spin_unlock(&root->fs_info->balance_lock);
3785
3786         /* First, mask out the RAID levels which aren't possible */
3787         if (num_devices == 1)
3788                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3789                            BTRFS_BLOCK_GROUP_RAID5);
3790         if (num_devices < 3)
3791                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3792         if (num_devices < 4)
3793                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3794
3795         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3796                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3797                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3798         flags &= ~tmp;
3799
3800         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3801                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3802         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3803                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3804         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3805                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3806         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3807                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3808         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3809                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3810
3811         return extended_to_chunk(flags | tmp);
3812 }
3813
3814 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3815 {
3816         unsigned seq;
3817         u64 flags;
3818
3819         do {
3820                 flags = orig_flags;
3821                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3822
3823                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3824                         flags |= root->fs_info->avail_data_alloc_bits;
3825                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3826                         flags |= root->fs_info->avail_system_alloc_bits;
3827                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3828                         flags |= root->fs_info->avail_metadata_alloc_bits;
3829         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3830
3831         return btrfs_reduce_alloc_profile(root, flags);
3832 }
3833
3834 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3835 {
3836         u64 flags;
3837         u64 ret;
3838
3839         if (data)
3840                 flags = BTRFS_BLOCK_GROUP_DATA;
3841         else if (root == root->fs_info->chunk_root)
3842                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3843         else
3844                 flags = BTRFS_BLOCK_GROUP_METADATA;
3845
3846         ret = get_alloc_profile(root, flags);
3847         return ret;
3848 }
3849
3850 /*
3851  * This will check the space that the inode allocates from to make sure we have
3852  * enough space for bytes.
3853  */
3854 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3855 {
3856         struct btrfs_space_info *data_sinfo;
3857         struct btrfs_root *root = BTRFS_I(inode)->root;
3858         struct btrfs_fs_info *fs_info = root->fs_info;
3859         u64 used;
3860         int ret = 0;
3861         int committed = 0;
3862         int have_pinned_space = 1;
3863
3864         /* make sure bytes are sectorsize aligned */
3865         bytes = ALIGN(bytes, root->sectorsize);
3866
3867         if (btrfs_is_free_space_inode(inode)) {
3868                 committed = 1;
3869                 ASSERT(current->journal_info);
3870         }
3871
3872         data_sinfo = fs_info->data_sinfo;
3873         if (!data_sinfo)
3874                 goto alloc;
3875
3876 again:
3877         /* make sure we have enough space to handle the data first */
3878         spin_lock(&data_sinfo->lock);
3879         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3880                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3881                 data_sinfo->bytes_may_use;
3882
3883         if (used + bytes > data_sinfo->total_bytes) {
3884                 struct btrfs_trans_handle *trans;
3885
3886                 /*
3887                  * if we don't have enough free bytes in this space then we need
3888                  * to alloc a new chunk.
3889                  */
3890                 if (!data_sinfo->full) {
3891                         u64 alloc_target;
3892
3893                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3894                         spin_unlock(&data_sinfo->lock);
3895 alloc:
3896                         alloc_target = btrfs_get_alloc_profile(root, 1);
3897                         /*
3898                          * It is ugly that we don't call nolock join
3899                          * transaction for the free space inode case here.
3900                          * But it is safe because we only do the data space
3901                          * reservation for the free space cache in the
3902                          * transaction context, the common join transaction
3903                          * just increase the counter of the current transaction
3904                          * handler, doesn't try to acquire the trans_lock of
3905                          * the fs.
3906                          */
3907                         trans = btrfs_join_transaction(root);
3908                         if (IS_ERR(trans))
3909                                 return PTR_ERR(trans);
3910
3911                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3912                                              alloc_target,
3913                                              CHUNK_ALLOC_NO_FORCE);
3914                         btrfs_end_transaction(trans, root);
3915                         if (ret < 0) {
3916                                 if (ret != -ENOSPC)
3917                                         return ret;
3918                                 else
3919                                         goto commit_trans;
3920                         }
3921
3922                         if (!data_sinfo)
3923                                 data_sinfo = fs_info->data_sinfo;
3924
3925                         goto again;
3926                 }
3927
3928                 /*
3929                  * If we don't have enough pinned space to deal with this
3930                  * allocation, and no removed chunk in current transaction,
3931                  * don't bother committing the transaction.
3932                  */
3933                 if (percpu_counter_compare(&data_sinfo->total_bytes_pinned,
3934                                            used + bytes -
3935                                            data_sinfo->total_bytes) < 0)
3936                         have_pinned_space = 0;
3937                 spin_unlock(&data_sinfo->lock);
3938
3939                 /* commit the current transaction and try again */
3940 commit_trans:
3941                 if (!committed &&
3942                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3943                         committed = 1;
3944
3945                         trans = btrfs_join_transaction(root);
3946                         if (IS_ERR(trans))
3947                                 return PTR_ERR(trans);
3948                         if (have_pinned_space ||
3949                             trans->transaction->have_free_bgs) {
3950                                 ret = btrfs_commit_transaction(trans, root);
3951                                 if (ret)
3952                                         return ret;
3953                                 /*
3954                                  * make sure that all running delayed iput are
3955                                  * done
3956                                  */
3957                                 down_write(&root->fs_info->delayed_iput_sem);
3958                                 up_write(&root->fs_info->delayed_iput_sem);
3959                                 goto again;
3960                         } else {
3961                                 btrfs_end_transaction(trans, root);
3962                         }
3963                 }
3964
3965                 trace_btrfs_space_reservation(root->fs_info,
3966                                               "space_info:enospc",
3967                                               data_sinfo->flags, bytes, 1);
3968                 return -ENOSPC;
3969         }
3970         data_sinfo->bytes_may_use += bytes;
3971         trace_btrfs_space_reservation(root->fs_info, "space_info",
3972                                       data_sinfo->flags, bytes, 1);
3973         spin_unlock(&data_sinfo->lock);
3974
3975         return 0;
3976 }
3977
3978 /*
3979  * Called if we need to clear a data reservation for this inode.
3980  */
3981 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3982 {
3983         struct btrfs_root *root = BTRFS_I(inode)->root;
3984         struct btrfs_space_info *data_sinfo;
3985
3986         /* make sure bytes are sectorsize aligned */
3987         bytes = ALIGN(bytes, root->sectorsize);
3988
3989         data_sinfo = root->fs_info->data_sinfo;
3990         spin_lock(&data_sinfo->lock);
3991         WARN_ON(data_sinfo->bytes_may_use < bytes);
3992         data_sinfo->bytes_may_use -= bytes;
3993         trace_btrfs_space_reservation(root->fs_info, "space_info",
3994                                       data_sinfo->flags, bytes, 0);
3995         spin_unlock(&data_sinfo->lock);
3996 }
3997
3998 static void force_metadata_allocation(struct btrfs_fs_info *info)
3999 {
4000         struct list_head *head = &info->space_info;
4001         struct btrfs_space_info *found;
4002
4003         rcu_read_lock();
4004         list_for_each_entry_rcu(found, head, list) {
4005                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4006                         found->force_alloc = CHUNK_ALLOC_FORCE;
4007         }
4008         rcu_read_unlock();
4009 }
4010
4011 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4012 {
4013         return (global->size << 1);
4014 }
4015
4016 static int should_alloc_chunk(struct btrfs_root *root,
4017                               struct btrfs_space_info *sinfo, int force)
4018 {
4019         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4020         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4021         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4022         u64 thresh;
4023
4024         if (force == CHUNK_ALLOC_FORCE)
4025                 return 1;
4026
4027         /*
4028          * We need to take into account the global rsv because for all intents
4029          * and purposes it's used space.  Don't worry about locking the
4030          * global_rsv, it doesn't change except when the transaction commits.
4031          */
4032         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4033                 num_allocated += calc_global_rsv_need_space(global_rsv);
4034
4035         /*
4036          * in limited mode, we want to have some free space up to
4037          * about 1% of the FS size.
4038          */
4039         if (force == CHUNK_ALLOC_LIMITED) {
4040                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4041                 thresh = max_t(u64, 64 * 1024 * 1024,
4042                                div_factor_fine(thresh, 1));
4043
4044                 if (num_bytes - num_allocated < thresh)
4045                         return 1;
4046         }
4047
4048         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
4049                 return 0;
4050         return 1;
4051 }
4052
4053 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
4054 {
4055         u64 num_dev;
4056
4057         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4058                     BTRFS_BLOCK_GROUP_RAID0 |
4059                     BTRFS_BLOCK_GROUP_RAID5 |
4060                     BTRFS_BLOCK_GROUP_RAID6))
4061                 num_dev = root->fs_info->fs_devices->rw_devices;
4062         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4063                 num_dev = 2;
4064         else
4065                 num_dev = 1;    /* DUP or single */
4066
4067         /* metadata for updaing devices and chunk tree */
4068         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
4069 }
4070
4071 static void check_system_chunk(struct btrfs_trans_handle *trans,
4072                                struct btrfs_root *root, u64 type)
4073 {
4074         struct btrfs_space_info *info;
4075         u64 left;
4076         u64 thresh;
4077
4078         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4079         spin_lock(&info->lock);
4080         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4081                 info->bytes_reserved - info->bytes_readonly;
4082         spin_unlock(&info->lock);
4083
4084         thresh = get_system_chunk_thresh(root, type);
4085         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4086                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4087                         left, thresh, type);
4088                 dump_space_info(info, 0, 0);
4089         }
4090
4091         if (left < thresh) {
4092                 u64 flags;
4093
4094                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4095                 btrfs_alloc_chunk(trans, root, flags);
4096         }
4097 }
4098
4099 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4100                           struct btrfs_root *extent_root, u64 flags, int force)
4101 {
4102         struct btrfs_space_info *space_info;
4103         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4104         int wait_for_alloc = 0;
4105         int ret = 0;
4106
4107         /* Don't re-enter if we're already allocating a chunk */
4108         if (trans->allocating_chunk)
4109                 return -ENOSPC;
4110
4111         space_info = __find_space_info(extent_root->fs_info, flags);
4112         if (!space_info) {
4113                 ret = update_space_info(extent_root->fs_info, flags,
4114                                         0, 0, &space_info);
4115                 BUG_ON(ret); /* -ENOMEM */
4116         }
4117         BUG_ON(!space_info); /* Logic error */
4118
4119 again:
4120         spin_lock(&space_info->lock);
4121         if (force < space_info->force_alloc)
4122                 force = space_info->force_alloc;
4123         if (space_info->full) {
4124                 if (should_alloc_chunk(extent_root, space_info, force))
4125                         ret = -ENOSPC;
4126                 else
4127                         ret = 0;
4128                 spin_unlock(&space_info->lock);
4129                 return ret;
4130         }
4131
4132         if (!should_alloc_chunk(extent_root, space_info, force)) {
4133                 spin_unlock(&space_info->lock);
4134                 return 0;
4135         } else if (space_info->chunk_alloc) {
4136                 wait_for_alloc = 1;
4137         } else {
4138                 space_info->chunk_alloc = 1;
4139         }
4140
4141         spin_unlock(&space_info->lock);
4142
4143         mutex_lock(&fs_info->chunk_mutex);
4144
4145         /*
4146          * The chunk_mutex is held throughout the entirety of a chunk
4147          * allocation, so once we've acquired the chunk_mutex we know that the
4148          * other guy is done and we need to recheck and see if we should
4149          * allocate.
4150          */
4151         if (wait_for_alloc) {
4152                 mutex_unlock(&fs_info->chunk_mutex);
4153                 wait_for_alloc = 0;
4154                 goto again;
4155         }
4156
4157         trans->allocating_chunk = true;
4158
4159         /*
4160          * If we have mixed data/metadata chunks we want to make sure we keep
4161          * allocating mixed chunks instead of individual chunks.
4162          */
4163         if (btrfs_mixed_space_info(space_info))
4164                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4165
4166         /*
4167          * if we're doing a data chunk, go ahead and make sure that
4168          * we keep a reasonable number of metadata chunks allocated in the
4169          * FS as well.
4170          */
4171         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4172                 fs_info->data_chunk_allocations++;
4173                 if (!(fs_info->data_chunk_allocations %
4174                       fs_info->metadata_ratio))
4175                         force_metadata_allocation(fs_info);
4176         }
4177
4178         /*
4179          * Check if we have enough space in SYSTEM chunk because we may need
4180          * to update devices.
4181          */
4182         check_system_chunk(trans, extent_root, flags);
4183
4184         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4185         trans->allocating_chunk = false;
4186
4187         spin_lock(&space_info->lock);
4188         if (ret < 0 && ret != -ENOSPC)
4189                 goto out;
4190         if (ret)
4191                 space_info->full = 1;
4192         else
4193                 ret = 1;
4194
4195         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4196 out:
4197         space_info->chunk_alloc = 0;
4198         spin_unlock(&space_info->lock);
4199         mutex_unlock(&fs_info->chunk_mutex);
4200         return ret;
4201 }
4202
4203 static int can_overcommit(struct btrfs_root *root,
4204                           struct btrfs_space_info *space_info, u64 bytes,
4205                           enum btrfs_reserve_flush_enum flush)
4206 {
4207         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4208         u64 profile = btrfs_get_alloc_profile(root, 0);
4209         u64 space_size;
4210         u64 avail;
4211         u64 used;
4212
4213         used = space_info->bytes_used + space_info->bytes_reserved +
4214                 space_info->bytes_pinned + space_info->bytes_readonly;
4215
4216         /*
4217          * We only want to allow over committing if we have lots of actual space
4218          * free, but if we don't have enough space to handle the global reserve
4219          * space then we could end up having a real enospc problem when trying
4220          * to allocate a chunk or some other such important allocation.
4221          */
4222         spin_lock(&global_rsv->lock);
4223         space_size = calc_global_rsv_need_space(global_rsv);
4224         spin_unlock(&global_rsv->lock);
4225         if (used + space_size >= space_info->total_bytes)
4226                 return 0;
4227
4228         used += space_info->bytes_may_use;
4229
4230         spin_lock(&root->fs_info->free_chunk_lock);
4231         avail = root->fs_info->free_chunk_space;
4232         spin_unlock(&root->fs_info->free_chunk_lock);
4233
4234         /*
4235          * If we have dup, raid1 or raid10 then only half of the free
4236          * space is actually useable.  For raid56, the space info used
4237          * doesn't include the parity drive, so we don't have to
4238          * change the math
4239          */
4240         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4241                        BTRFS_BLOCK_GROUP_RAID1 |
4242                        BTRFS_BLOCK_GROUP_RAID10))
4243                 avail >>= 1;
4244
4245         /*
4246          * If we aren't flushing all things, let us overcommit up to
4247          * 1/2th of the space. If we can flush, don't let us overcommit
4248          * too much, let it overcommit up to 1/8 of the space.
4249          */
4250         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4251                 avail >>= 3;
4252         else
4253                 avail >>= 1;
4254
4255         if (used + bytes < space_info->total_bytes + avail)
4256                 return 1;
4257         return 0;
4258 }
4259
4260 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4261                                          unsigned long nr_pages, int nr_items)
4262 {
4263         struct super_block *sb = root->fs_info->sb;
4264
4265         if (down_read_trylock(&sb->s_umount)) {
4266                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4267                 up_read(&sb->s_umount);
4268         } else {
4269                 /*
4270                  * We needn't worry the filesystem going from r/w to r/o though
4271                  * we don't acquire ->s_umount mutex, because the filesystem
4272                  * should guarantee the delalloc inodes list be empty after
4273                  * the filesystem is readonly(all dirty pages are written to
4274                  * the disk).
4275                  */
4276                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4277                 if (!current->journal_info)
4278                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4279         }
4280 }
4281
4282 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4283 {
4284         u64 bytes;
4285         int nr;
4286
4287         bytes = btrfs_calc_trans_metadata_size(root, 1);
4288         nr = (int)div64_u64(to_reclaim, bytes);
4289         if (!nr)
4290                 nr = 1;
4291         return nr;
4292 }
4293
4294 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4295
4296 /*
4297  * shrink metadata reservation for delalloc
4298  */
4299 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4300                             bool wait_ordered)
4301 {
4302         struct btrfs_block_rsv *block_rsv;
4303         struct btrfs_space_info *space_info;
4304         struct btrfs_trans_handle *trans;
4305         u64 delalloc_bytes;
4306         u64 max_reclaim;
4307         long time_left;
4308         unsigned long nr_pages;
4309         int loops;
4310         int items;
4311         enum btrfs_reserve_flush_enum flush;
4312
4313         /* Calc the number of the pages we need flush for space reservation */
4314         items = calc_reclaim_items_nr(root, to_reclaim);
4315         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4316
4317         trans = (struct btrfs_trans_handle *)current->journal_info;
4318         block_rsv = &root->fs_info->delalloc_block_rsv;
4319         space_info = block_rsv->space_info;
4320
4321         delalloc_bytes = percpu_counter_sum_positive(
4322                                                 &root->fs_info->delalloc_bytes);
4323         if (delalloc_bytes == 0) {
4324                 if (trans)
4325                         return;
4326                 if (wait_ordered)
4327                         btrfs_wait_ordered_roots(root->fs_info, items);
4328                 return;
4329         }
4330
4331         loops = 0;
4332         while (delalloc_bytes && loops < 3) {
4333                 max_reclaim = min(delalloc_bytes, to_reclaim);
4334                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4335                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4336                 /*
4337                  * We need to wait for the async pages to actually start before
4338                  * we do anything.
4339                  */
4340                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4341                 if (!max_reclaim)
4342                         goto skip_async;
4343
4344                 if (max_reclaim <= nr_pages)
4345                         max_reclaim = 0;
4346                 else
4347                         max_reclaim -= nr_pages;
4348
4349                 wait_event(root->fs_info->async_submit_wait,
4350                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4351                            (int)max_reclaim);
4352 skip_async:
4353                 if (!trans)
4354                         flush = BTRFS_RESERVE_FLUSH_ALL;
4355                 else
4356                         flush = BTRFS_RESERVE_NO_FLUSH;
4357                 spin_lock(&space_info->lock);
4358                 if (can_overcommit(root, space_info, orig, flush)) {
4359                         spin_unlock(&space_info->lock);
4360                         break;
4361                 }
4362                 spin_unlock(&space_info->lock);
4363
4364                 loops++;
4365                 if (wait_ordered && !trans) {
4366                         btrfs_wait_ordered_roots(root->fs_info, items);
4367                 } else {
4368                         time_left = schedule_timeout_killable(1);
4369                         if (time_left)
4370                                 break;
4371                 }
4372                 delalloc_bytes = percpu_counter_sum_positive(
4373                                                 &root->fs_info->delalloc_bytes);
4374         }
4375 }
4376
4377 /**
4378  * maybe_commit_transaction - possibly commit the transaction if its ok to
4379  * @root - the root we're allocating for
4380  * @bytes - the number of bytes we want to reserve
4381  * @force - force the commit
4382  *
4383  * This will check to make sure that committing the transaction will actually
4384  * get us somewhere and then commit the transaction if it does.  Otherwise it
4385  * will return -ENOSPC.
4386  */
4387 static int may_commit_transaction(struct btrfs_root *root,
4388                                   struct btrfs_space_info *space_info,
4389                                   u64 bytes, int force)
4390 {
4391         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4392         struct btrfs_trans_handle *trans;
4393
4394         trans = (struct btrfs_trans_handle *)current->journal_info;
4395         if (trans)
4396                 return -EAGAIN;
4397
4398         if (force)
4399                 goto commit;
4400
4401         /* See if there is enough pinned space to make this reservation */
4402         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4403                                    bytes) >= 0)
4404                 goto commit;
4405
4406         /*
4407          * See if there is some space in the delayed insertion reservation for
4408          * this reservation.
4409          */
4410         if (space_info != delayed_rsv->space_info)
4411                 return -ENOSPC;
4412
4413         spin_lock(&delayed_rsv->lock);
4414         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4415                                    bytes - delayed_rsv->size) >= 0) {
4416                 spin_unlock(&delayed_rsv->lock);
4417                 return -ENOSPC;
4418         }
4419         spin_unlock(&delayed_rsv->lock);
4420
4421 commit:
4422         trans = btrfs_join_transaction(root);
4423         if (IS_ERR(trans))
4424                 return -ENOSPC;
4425
4426         return btrfs_commit_transaction(trans, root);
4427 }
4428
4429 enum flush_state {
4430         FLUSH_DELAYED_ITEMS_NR  =       1,
4431         FLUSH_DELAYED_ITEMS     =       2,
4432         FLUSH_DELALLOC          =       3,
4433         FLUSH_DELALLOC_WAIT     =       4,
4434         ALLOC_CHUNK             =       5,
4435         COMMIT_TRANS            =       6,
4436 };
4437
4438 static int flush_space(struct btrfs_root *root,
4439                        struct btrfs_space_info *space_info, u64 num_bytes,
4440                        u64 orig_bytes, int state)
4441 {
4442         struct btrfs_trans_handle *trans;
4443         int nr;
4444         int ret = 0;
4445
4446         switch (state) {
4447         case FLUSH_DELAYED_ITEMS_NR:
4448         case FLUSH_DELAYED_ITEMS:
4449                 if (state == FLUSH_DELAYED_ITEMS_NR)
4450                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4451                 else
4452                         nr = -1;
4453
4454                 trans = btrfs_join_transaction(root);
4455                 if (IS_ERR(trans)) {
4456                         ret = PTR_ERR(trans);
4457                         break;
4458                 }
4459                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4460                 btrfs_end_transaction(trans, root);
4461                 break;
4462         case FLUSH_DELALLOC:
4463         case FLUSH_DELALLOC_WAIT:
4464                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4465                                 state == FLUSH_DELALLOC_WAIT);
4466                 break;
4467         case ALLOC_CHUNK:
4468                 trans = btrfs_join_transaction(root);
4469                 if (IS_ERR(trans)) {
4470                         ret = PTR_ERR(trans);
4471                         break;
4472                 }
4473                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4474                                      btrfs_get_alloc_profile(root, 0),
4475                                      CHUNK_ALLOC_NO_FORCE);
4476                 btrfs_end_transaction(trans, root);
4477                 if (ret == -ENOSPC)
4478                         ret = 0;
4479                 break;
4480         case COMMIT_TRANS:
4481                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4482                 break;
4483         default:
4484                 ret = -ENOSPC;
4485                 break;
4486         }
4487
4488         return ret;
4489 }
4490
4491 static inline u64
4492 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4493                                  struct btrfs_space_info *space_info)
4494 {
4495         u64 used;
4496         u64 expected;
4497         u64 to_reclaim;
4498
4499         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4500                                 16 * 1024 * 1024);
4501         spin_lock(&space_info->lock);
4502         if (can_overcommit(root, space_info, to_reclaim,
4503                            BTRFS_RESERVE_FLUSH_ALL)) {
4504                 to_reclaim = 0;
4505                 goto out;
4506         }
4507
4508         used = space_info->bytes_used + space_info->bytes_reserved +
4509                space_info->bytes_pinned + space_info->bytes_readonly +
4510                space_info->bytes_may_use;
4511         if (can_overcommit(root, space_info, 1024 * 1024,
4512                            BTRFS_RESERVE_FLUSH_ALL))
4513                 expected = div_factor_fine(space_info->total_bytes, 95);
4514         else
4515                 expected = div_factor_fine(space_info->total_bytes, 90);
4516
4517         if (used > expected)
4518                 to_reclaim = used - expected;
4519         else
4520                 to_reclaim = 0;
4521         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4522                                      space_info->bytes_reserved);
4523 out:
4524         spin_unlock(&space_info->lock);
4525
4526         return to_reclaim;
4527 }
4528
4529 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4530                                         struct btrfs_fs_info *fs_info, u64 used)
4531 {
4532         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4533
4534         /* If we're just plain full then async reclaim just slows us down. */
4535         if (space_info->bytes_used >= thresh)
4536                 return 0;
4537
4538         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4539                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4540 }
4541
4542 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4543                                        struct btrfs_fs_info *fs_info,
4544                                        int flush_state)
4545 {
4546         u64 used;
4547
4548         spin_lock(&space_info->lock);
4549         /*
4550          * We run out of space and have not got any free space via flush_space,
4551          * so don't bother doing async reclaim.
4552          */
4553         if (flush_state > COMMIT_TRANS && space_info->full) {
4554                 spin_unlock(&space_info->lock);
4555                 return 0;
4556         }
4557
4558         used = space_info->bytes_used + space_info->bytes_reserved +
4559                space_info->bytes_pinned + space_info->bytes_readonly +
4560                space_info->bytes_may_use;
4561         if (need_do_async_reclaim(space_info, fs_info, used)) {
4562                 spin_unlock(&space_info->lock);
4563                 return 1;
4564         }
4565         spin_unlock(&space_info->lock);
4566
4567         return 0;
4568 }
4569
4570 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4571 {
4572         struct btrfs_fs_info *fs_info;
4573         struct btrfs_space_info *space_info;
4574         u64 to_reclaim;
4575         int flush_state;
4576
4577         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4578         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4579
4580         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4581                                                       space_info);
4582         if (!to_reclaim)
4583                 return;
4584
4585         flush_state = FLUSH_DELAYED_ITEMS_NR;
4586         do {
4587                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4588                             to_reclaim, flush_state);
4589                 flush_state++;
4590                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4591                                                  flush_state))
4592                         return;
4593         } while (flush_state < COMMIT_TRANS);
4594 }
4595
4596 void btrfs_init_async_reclaim_work(struct work_struct *work)
4597 {
4598         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4599 }
4600
4601 /**
4602  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4603  * @root - the root we're allocating for
4604  * @block_rsv - the block_rsv we're allocating for
4605  * @orig_bytes - the number of bytes we want
4606  * @flush - whether or not we can flush to make our reservation
4607  *
4608  * This will reserve orgi_bytes number of bytes from the space info associated
4609  * with the block_rsv.  If there is not enough space it will make an attempt to
4610  * flush out space to make room.  It will do this by flushing delalloc if
4611  * possible or committing the transaction.  If flush is 0 then no attempts to
4612  * regain reservations will be made and this will fail if there is not enough
4613  * space already.
4614  */
4615 static int reserve_metadata_bytes(struct btrfs_root *root,
4616                                   struct btrfs_block_rsv *block_rsv,
4617                                   u64 orig_bytes,
4618                                   enum btrfs_reserve_flush_enum flush)
4619 {
4620         struct btrfs_space_info *space_info = block_rsv->space_info;
4621         u64 used;
4622         u64 num_bytes = orig_bytes;
4623         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4624         int ret = 0;
4625         bool flushing = false;
4626
4627 again:
4628         ret = 0;
4629         spin_lock(&space_info->lock);
4630         /*
4631          * We only want to wait if somebody other than us is flushing and we
4632          * are actually allowed to flush all things.
4633          */
4634         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4635                space_info->flush) {
4636                 spin_unlock(&space_info->lock);
4637                 /*
4638                  * If we have a trans handle we can't wait because the flusher
4639                  * may have to commit the transaction, which would mean we would
4640                  * deadlock since we are waiting for the flusher to finish, but
4641                  * hold the current transaction open.
4642                  */
4643                 if (current->journal_info)
4644                         return -EAGAIN;
4645                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4646                 /* Must have been killed, return */
4647                 if (ret)
4648                         return -EINTR;
4649
4650                 spin_lock(&space_info->lock);
4651         }
4652
4653         ret = -ENOSPC;
4654         used = space_info->bytes_used + space_info->bytes_reserved +
4655                 space_info->bytes_pinned + space_info->bytes_readonly +
4656                 space_info->bytes_may_use;
4657
4658         /*
4659          * The idea here is that we've not already over-reserved the block group
4660          * then we can go ahead and save our reservation first and then start
4661          * flushing if we need to.  Otherwise if we've already overcommitted
4662          * lets start flushing stuff first and then come back and try to make
4663          * our reservation.
4664          */
4665         if (used <= space_info->total_bytes) {
4666                 if (used + orig_bytes <= space_info->total_bytes) {
4667                         space_info->bytes_may_use += orig_bytes;
4668                         trace_btrfs_space_reservation(root->fs_info,
4669                                 "space_info", space_info->flags, orig_bytes, 1);
4670                         ret = 0;
4671                 } else {
4672                         /*
4673                          * Ok set num_bytes to orig_bytes since we aren't
4674                          * overocmmitted, this way we only try and reclaim what
4675                          * we need.
4676                          */
4677                         num_bytes = orig_bytes;
4678                 }
4679         } else {
4680                 /*
4681                  * Ok we're over committed, set num_bytes to the overcommitted
4682                  * amount plus the amount of bytes that we need for this
4683                  * reservation.
4684                  */
4685                 num_bytes = used - space_info->total_bytes +
4686                         (orig_bytes * 2);
4687         }
4688
4689         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4690                 space_info->bytes_may_use += orig_bytes;
4691                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4692                                               space_info->flags, orig_bytes,
4693                                               1);
4694                 ret = 0;
4695         }
4696
4697         /*
4698          * Couldn't make our reservation, save our place so while we're trying
4699          * to reclaim space we can actually use it instead of somebody else
4700          * stealing it from us.
4701          *
4702          * We make the other tasks wait for the flush only when we can flush
4703          * all things.
4704          */
4705         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4706                 flushing = true;
4707                 space_info->flush = 1;
4708         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4709                 used += orig_bytes;
4710                 /*
4711                  * We will do the space reservation dance during log replay,
4712                  * which means we won't have fs_info->fs_root set, so don't do
4713                  * the async reclaim as we will panic.
4714                  */
4715                 if (!root->fs_info->log_root_recovering &&
4716                     need_do_async_reclaim(space_info, root->fs_info, used) &&
4717                     !work_busy(&root->fs_info->async_reclaim_work))
4718                         queue_work(system_unbound_wq,
4719                                    &root->fs_info->async_reclaim_work);
4720         }
4721         spin_unlock(&space_info->lock);
4722
4723         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4724                 goto out;
4725
4726         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4727                           flush_state);
4728         flush_state++;
4729
4730         /*
4731          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4732          * would happen. So skip delalloc flush.
4733          */
4734         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4735             (flush_state == FLUSH_DELALLOC ||
4736              flush_state == FLUSH_DELALLOC_WAIT))
4737                 flush_state = ALLOC_CHUNK;
4738
4739         if (!ret)
4740                 goto again;
4741         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4742                  flush_state < COMMIT_TRANS)
4743                 goto again;
4744         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4745                  flush_state <= COMMIT_TRANS)
4746                 goto again;
4747
4748 out:
4749         if (ret == -ENOSPC &&
4750             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4751                 struct btrfs_block_rsv *global_rsv =
4752                         &root->fs_info->global_block_rsv;
4753
4754                 if (block_rsv != global_rsv &&
4755                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4756                         ret = 0;
4757         }
4758         if (ret == -ENOSPC)
4759                 trace_btrfs_space_reservation(root->fs_info,
4760                                               "space_info:enospc",
4761                                               space_info->flags, orig_bytes, 1);
4762         if (flushing) {
4763                 spin_lock(&space_info->lock);
4764                 space_info->flush = 0;
4765                 wake_up_all(&space_info->wait);
4766                 spin_unlock(&space_info->lock);
4767         }
4768         return ret;
4769 }
4770
4771 static struct btrfs_block_rsv *get_block_rsv(
4772                                         const struct btrfs_trans_handle *trans,
4773                                         const struct btrfs_root *root)
4774 {
4775         struct btrfs_block_rsv *block_rsv = NULL;
4776
4777         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4778                 block_rsv = trans->block_rsv;
4779
4780         if (root == root->fs_info->csum_root && trans->adding_csums)
4781                 block_rsv = trans->block_rsv;
4782
4783         if (root == root->fs_info->uuid_root)
4784                 block_rsv = trans->block_rsv;
4785
4786         if (!block_rsv)
4787                 block_rsv = root->block_rsv;
4788
4789         if (!block_rsv)
4790                 block_rsv = &root->fs_info->empty_block_rsv;
4791
4792         return block_rsv;
4793 }
4794
4795 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4796                                u64 num_bytes)
4797 {
4798         int ret = -ENOSPC;
4799         spin_lock(&block_rsv->lock);
4800         if (block_rsv->reserved >= num_bytes) {
4801                 block_rsv->reserved -= num_bytes;
4802                 if (block_rsv->reserved < block_rsv->size)
4803                         block_rsv->full = 0;
4804                 ret = 0;
4805         }
4806         spin_unlock(&block_rsv->lock);
4807         return ret;
4808 }
4809
4810 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4811                                 u64 num_bytes, int update_size)
4812 {
4813         spin_lock(&block_rsv->lock);
4814         block_rsv->reserved += num_bytes;
4815         if (update_size)
4816                 block_rsv->size += num_bytes;
4817         else if (block_rsv->reserved >= block_rsv->size)
4818                 block_rsv->full = 1;
4819         spin_unlock(&block_rsv->lock);
4820 }
4821
4822 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4823                              struct btrfs_block_rsv *dest, u64 num_bytes,
4824                              int min_factor)
4825 {
4826         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4827         u64 min_bytes;
4828
4829         if (global_rsv->space_info != dest->space_info)
4830                 return -ENOSPC;
4831
4832         spin_lock(&global_rsv->lock);
4833         min_bytes = div_factor(global_rsv->size, min_factor);
4834         if (global_rsv->reserved < min_bytes + num_bytes) {
4835                 spin_unlock(&global_rsv->lock);
4836                 return -ENOSPC;
4837         }
4838         global_rsv->reserved -= num_bytes;
4839         if (global_rsv->reserved < global_rsv->size)
4840                 global_rsv->full = 0;
4841         spin_unlock(&global_rsv->lock);
4842
4843         block_rsv_add_bytes(dest, num_bytes, 1);
4844         return 0;
4845 }
4846
4847 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4848                                     struct btrfs_block_rsv *block_rsv,
4849                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4850 {
4851         struct btrfs_space_info *space_info = block_rsv->space_info;
4852
4853         spin_lock(&block_rsv->lock);
4854         if (num_bytes == (u64)-1)
4855                 num_bytes = block_rsv->size;
4856         block_rsv->size -= num_bytes;
4857         if (block_rsv->reserved >= block_rsv->size) {
4858                 num_bytes = block_rsv->reserved - block_rsv->size;
4859                 block_rsv->reserved = block_rsv->size;
4860                 block_rsv->full = 1;
4861         } else {
4862                 num_bytes = 0;
4863         }
4864         spin_unlock(&block_rsv->lock);
4865
4866         if (num_bytes > 0) {
4867                 if (dest) {
4868                         spin_lock(&dest->lock);
4869                         if (!dest->full) {
4870                                 u64 bytes_to_add;
4871
4872                                 bytes_to_add = dest->size - dest->reserved;
4873                                 bytes_to_add = min(num_bytes, bytes_to_add);
4874                                 dest->reserved += bytes_to_add;
4875                                 if (dest->reserved >= dest->size)
4876                                         dest->full = 1;
4877                                 num_bytes -= bytes_to_add;
4878                         }
4879                         spin_unlock(&dest->lock);
4880                 }
4881                 if (num_bytes) {
4882                         spin_lock(&space_info->lock);
4883                         space_info->bytes_may_use -= num_bytes;
4884                         trace_btrfs_space_reservation(fs_info, "space_info",
4885                                         space_info->flags, num_bytes, 0);
4886                         spin_unlock(&space_info->lock);
4887                 }
4888         }
4889 }
4890
4891 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4892                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4893 {
4894         int ret;
4895
4896         ret = block_rsv_use_bytes(src, num_bytes);
4897         if (ret)
4898                 return ret;
4899
4900         block_rsv_add_bytes(dst, num_bytes, 1);
4901         return 0;
4902 }
4903
4904 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4905 {
4906         memset(rsv, 0, sizeof(*rsv));
4907         spin_lock_init(&rsv->lock);
4908         rsv->type = type;
4909 }
4910
4911 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4912                                               unsigned short type)
4913 {
4914         struct btrfs_block_rsv *block_rsv;
4915         struct btrfs_fs_info *fs_info = root->fs_info;
4916
4917         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4918         if (!block_rsv)
4919                 return NULL;
4920
4921         btrfs_init_block_rsv(block_rsv, type);
4922         block_rsv->space_info = __find_space_info(fs_info,
4923                                                   BTRFS_BLOCK_GROUP_METADATA);
4924         return block_rsv;
4925 }
4926
4927 void btrfs_free_block_rsv(struct btrfs_root *root,
4928                           struct btrfs_block_rsv *rsv)
4929 {
4930         if (!rsv)
4931                 return;
4932         btrfs_block_rsv_release(root, rsv, (u64)-1);
4933         kfree(rsv);
4934 }
4935
4936 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
4937 {
4938         kfree(rsv);
4939 }
4940
4941 int btrfs_block_rsv_add(struct btrfs_root *root,
4942                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4943                         enum btrfs_reserve_flush_enum flush)
4944 {
4945         int ret;
4946
4947         if (num_bytes == 0)
4948                 return 0;
4949
4950         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4951         if (!ret) {
4952                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4953                 return 0;
4954         }
4955
4956         return ret;
4957 }
4958
4959 int btrfs_block_rsv_check(struct btrfs_root *root,
4960                           struct btrfs_block_rsv *block_rsv, int min_factor)
4961 {
4962         u64 num_bytes = 0;
4963         int ret = -ENOSPC;
4964
4965         if (!block_rsv)
4966                 return 0;
4967
4968         spin_lock(&block_rsv->lock);
4969         num_bytes = div_factor(block_rsv->size, min_factor);
4970         if (block_rsv->reserved >= num_bytes)
4971                 ret = 0;
4972         spin_unlock(&block_rsv->lock);
4973
4974         return ret;
4975 }
4976
4977 int btrfs_block_rsv_refill(struct btrfs_root *root,
4978                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4979                            enum btrfs_reserve_flush_enum flush)
4980 {
4981         u64 num_bytes = 0;
4982         int ret = -ENOSPC;
4983
4984         if (!block_rsv)
4985                 return 0;
4986
4987         spin_lock(&block_rsv->lock);
4988         num_bytes = min_reserved;
4989         if (block_rsv->reserved >= num_bytes)
4990                 ret = 0;
4991         else
4992                 num_bytes -= block_rsv->reserved;
4993         spin_unlock(&block_rsv->lock);
4994
4995         if (!ret)
4996                 return 0;
4997
4998         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4999         if (!ret) {
5000                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5001                 return 0;
5002         }
5003
5004         return ret;
5005 }
5006
5007 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5008                             struct btrfs_block_rsv *dst_rsv,
5009                             u64 num_bytes)
5010 {
5011         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5012 }
5013
5014 void btrfs_block_rsv_release(struct btrfs_root *root,
5015                              struct btrfs_block_rsv *block_rsv,
5016                              u64 num_bytes)
5017 {
5018         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5019         if (global_rsv == block_rsv ||
5020             block_rsv->space_info != global_rsv->space_info)
5021                 global_rsv = NULL;
5022         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5023                                 num_bytes);
5024 }
5025
5026 /*
5027  * helper to calculate size of global block reservation.
5028  * the desired value is sum of space used by extent tree,
5029  * checksum tree and root tree
5030  */
5031 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5032 {
5033         struct btrfs_space_info *sinfo;
5034         u64 num_bytes;
5035         u64 meta_used;
5036         u64 data_used;
5037         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5038
5039         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5040         spin_lock(&sinfo->lock);
5041         data_used = sinfo->bytes_used;
5042         spin_unlock(&sinfo->lock);
5043
5044         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5045         spin_lock(&sinfo->lock);
5046         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5047                 data_used = 0;
5048         meta_used = sinfo->bytes_used;
5049         spin_unlock(&sinfo->lock);
5050
5051         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5052                     csum_size * 2;
5053         num_bytes += div_u64(data_used + meta_used, 50);
5054
5055         if (num_bytes * 3 > meta_used)
5056                 num_bytes = div_u64(meta_used, 3);
5057
5058         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5059 }
5060
5061 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5062 {
5063         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5064         struct btrfs_space_info *sinfo = block_rsv->space_info;
5065         u64 num_bytes;
5066
5067         num_bytes = calc_global_metadata_size(fs_info);
5068
5069         spin_lock(&sinfo->lock);
5070         spin_lock(&block_rsv->lock);
5071
5072         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
5073
5074         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5075                     sinfo->bytes_reserved + sinfo->bytes_readonly +
5076                     sinfo->bytes_may_use;
5077
5078         if (sinfo->total_bytes > num_bytes) {
5079                 num_bytes = sinfo->total_bytes - num_bytes;
5080                 block_rsv->reserved += num_bytes;
5081                 sinfo->bytes_may_use += num_bytes;
5082                 trace_btrfs_space_reservation(fs_info, "space_info",
5083                                       sinfo->flags, num_bytes, 1);
5084         }
5085
5086         if (block_rsv->reserved >= block_rsv->size) {
5087                 num_bytes = block_rsv->reserved - block_rsv->size;
5088                 sinfo->bytes_may_use -= num_bytes;
5089                 trace_btrfs_space_reservation(fs_info, "space_info",
5090                                       sinfo->flags, num_bytes, 0);
5091                 block_rsv->reserved = block_rsv->size;
5092                 block_rsv->full = 1;
5093         }
5094
5095         spin_unlock(&block_rsv->lock);
5096         spin_unlock(&sinfo->lock);
5097 }
5098
5099 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5100 {
5101         struct btrfs_space_info *space_info;
5102
5103         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5104         fs_info->chunk_block_rsv.space_info = space_info;
5105
5106         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5107         fs_info->global_block_rsv.space_info = space_info;
5108         fs_info->delalloc_block_rsv.space_info = space_info;
5109         fs_info->trans_block_rsv.space_info = space_info;
5110         fs_info->empty_block_rsv.space_info = space_info;
5111         fs_info->delayed_block_rsv.space_info = space_info;
5112
5113         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5114         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5115         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5116         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5117         if (fs_info->quota_root)
5118                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5119         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5120
5121         update_global_block_rsv(fs_info);
5122 }
5123
5124 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5125 {
5126         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5127                                 (u64)-1);
5128         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5129         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5130         WARN_ON(fs_info->trans_block_rsv.size > 0);
5131         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5132         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5133         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5134         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5135         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5136 }
5137
5138 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5139                                   struct btrfs_root *root)
5140 {
5141         if (!trans->block_rsv)
5142                 return;
5143
5144         if (!trans->bytes_reserved)
5145                 return;
5146
5147         trace_btrfs_space_reservation(root->fs_info, "transaction",
5148                                       trans->transid, trans->bytes_reserved, 0);
5149         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5150         trans->bytes_reserved = 0;
5151 }
5152
5153 /* Can only return 0 or -ENOSPC */
5154 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5155                                   struct inode *inode)
5156 {
5157         struct btrfs_root *root = BTRFS_I(inode)->root;
5158         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5159         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5160
5161         /*
5162          * We need to hold space in order to delete our orphan item once we've
5163          * added it, so this takes the reservation so we can release it later
5164          * when we are truly done with the orphan item.
5165          */
5166         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5167         trace_btrfs_space_reservation(root->fs_info, "orphan",
5168                                       btrfs_ino(inode), num_bytes, 1);
5169         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5170 }
5171
5172 void btrfs_orphan_release_metadata(struct inode *inode)
5173 {
5174         struct btrfs_root *root = BTRFS_I(inode)->root;
5175         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5176         trace_btrfs_space_reservation(root->fs_info, "orphan",
5177                                       btrfs_ino(inode), num_bytes, 0);
5178         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5179 }
5180
5181 /*
5182  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5183  * root: the root of the parent directory
5184  * rsv: block reservation
5185  * items: the number of items that we need do reservation
5186  * qgroup_reserved: used to return the reserved size in qgroup
5187  *
5188  * This function is used to reserve the space for snapshot/subvolume
5189  * creation and deletion. Those operations are different with the
5190  * common file/directory operations, they change two fs/file trees
5191  * and root tree, the number of items that the qgroup reserves is
5192  * different with the free space reservation. So we can not use
5193  * the space reseravtion mechanism in start_transaction().
5194  */
5195 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5196                                      struct btrfs_block_rsv *rsv,
5197                                      int items,
5198                                      u64 *qgroup_reserved,
5199                                      bool use_global_rsv)
5200 {
5201         u64 num_bytes;
5202         int ret;
5203         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5204
5205         if (root->fs_info->quota_enabled) {
5206                 /* One for parent inode, two for dir entries */
5207                 num_bytes = 3 * root->nodesize;
5208                 ret = btrfs_qgroup_reserve(root, num_bytes);
5209                 if (ret)
5210                         return ret;
5211         } else {
5212                 num_bytes = 0;
5213         }
5214
5215         *qgroup_reserved = num_bytes;
5216
5217         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5218         rsv->space_info = __find_space_info(root->fs_info,
5219                                             BTRFS_BLOCK_GROUP_METADATA);
5220         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5221                                   BTRFS_RESERVE_FLUSH_ALL);
5222
5223         if (ret == -ENOSPC && use_global_rsv)
5224                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5225
5226         if (ret) {
5227                 if (*qgroup_reserved)
5228                         btrfs_qgroup_free(root, *qgroup_reserved);
5229         }
5230
5231         return ret;
5232 }
5233
5234 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5235                                       struct btrfs_block_rsv *rsv,
5236                                       u64 qgroup_reserved)
5237 {
5238         btrfs_block_rsv_release(root, rsv, (u64)-1);
5239         if (qgroup_reserved)
5240                 btrfs_qgroup_free(root, qgroup_reserved);
5241 }
5242
5243 /**
5244  * drop_outstanding_extent - drop an outstanding extent
5245  * @inode: the inode we're dropping the extent for
5246  * @num_bytes: the number of bytes we're relaseing.
5247  *
5248  * This is called when we are freeing up an outstanding extent, either called
5249  * after an error or after an extent is written.  This will return the number of
5250  * reserved extents that need to be freed.  This must be called with
5251  * BTRFS_I(inode)->lock held.
5252  */
5253 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5254 {
5255         unsigned drop_inode_space = 0;
5256         unsigned dropped_extents = 0;
5257         unsigned num_extents = 0;
5258
5259         num_extents = (unsigned)div64_u64(num_bytes +
5260                                           BTRFS_MAX_EXTENT_SIZE - 1,
5261                                           BTRFS_MAX_EXTENT_SIZE);
5262         ASSERT(num_extents);
5263         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5264         BTRFS_I(inode)->outstanding_extents -= num_extents;
5265
5266         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5267             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5268                                &BTRFS_I(inode)->runtime_flags))
5269                 drop_inode_space = 1;
5270
5271         /*
5272          * If we have more or the same amount of outsanding extents than we have
5273          * reserved then we need to leave the reserved extents count alone.
5274          */
5275         if (BTRFS_I(inode)->outstanding_extents >=
5276             BTRFS_I(inode)->reserved_extents)
5277                 return drop_inode_space;
5278
5279         dropped_extents = BTRFS_I(inode)->reserved_extents -
5280                 BTRFS_I(inode)->outstanding_extents;
5281         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5282         return dropped_extents + drop_inode_space;
5283 }
5284
5285 /**
5286  * calc_csum_metadata_size - return the amount of metada space that must be
5287  *      reserved/free'd for the given bytes.
5288  * @inode: the inode we're manipulating
5289  * @num_bytes: the number of bytes in question
5290  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5291  *
5292  * This adjusts the number of csum_bytes in the inode and then returns the
5293  * correct amount of metadata that must either be reserved or freed.  We
5294  * calculate how many checksums we can fit into one leaf and then divide the
5295  * number of bytes that will need to be checksumed by this value to figure out
5296  * how many checksums will be required.  If we are adding bytes then the number
5297  * may go up and we will return the number of additional bytes that must be
5298  * reserved.  If it is going down we will return the number of bytes that must
5299  * be freed.
5300  *
5301  * This must be called with BTRFS_I(inode)->lock held.
5302  */
5303 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5304                                    int reserve)
5305 {
5306         struct btrfs_root *root = BTRFS_I(inode)->root;
5307         u64 old_csums, num_csums;
5308
5309         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5310             BTRFS_I(inode)->csum_bytes == 0)
5311                 return 0;
5312
5313         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5314         if (reserve)
5315                 BTRFS_I(inode)->csum_bytes += num_bytes;
5316         else
5317                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5318         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5319
5320         /* No change, no need to reserve more */
5321         if (old_csums == num_csums)
5322                 return 0;
5323
5324         if (reserve)
5325                 return btrfs_calc_trans_metadata_size(root,
5326                                                       num_csums - old_csums);
5327
5328         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5329 }
5330
5331 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5332 {
5333         struct btrfs_root *root = BTRFS_I(inode)->root;
5334         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5335         u64 to_reserve = 0;
5336         u64 csum_bytes;
5337         unsigned nr_extents = 0;
5338         int extra_reserve = 0;
5339         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5340         int ret = 0;
5341         bool delalloc_lock = true;
5342         u64 to_free = 0;
5343         unsigned dropped;
5344
5345         /* If we are a free space inode we need to not flush since we will be in
5346          * the middle of a transaction commit.  We also don't need the delalloc
5347          * mutex since we won't race with anybody.  We need this mostly to make
5348          * lockdep shut its filthy mouth.
5349          */
5350         if (btrfs_is_free_space_inode(inode)) {
5351                 flush = BTRFS_RESERVE_NO_FLUSH;
5352                 delalloc_lock = false;
5353         }
5354
5355         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5356             btrfs_transaction_in_commit(root->fs_info))
5357                 schedule_timeout(1);
5358
5359         if (delalloc_lock)
5360                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5361
5362         num_bytes = ALIGN(num_bytes, root->sectorsize);
5363
5364         spin_lock(&BTRFS_I(inode)->lock);
5365         nr_extents = (unsigned)div64_u64(num_bytes +
5366                                          BTRFS_MAX_EXTENT_SIZE - 1,
5367                                          BTRFS_MAX_EXTENT_SIZE);
5368         BTRFS_I(inode)->outstanding_extents += nr_extents;
5369         nr_extents = 0;
5370
5371         if (BTRFS_I(inode)->outstanding_extents >
5372             BTRFS_I(inode)->reserved_extents)
5373                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5374                         BTRFS_I(inode)->reserved_extents;
5375
5376         /*
5377          * Add an item to reserve for updating the inode when we complete the
5378          * delalloc io.
5379          */
5380         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5381                       &BTRFS_I(inode)->runtime_flags)) {
5382                 nr_extents++;
5383                 extra_reserve = 1;
5384         }
5385
5386         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5387         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5388         csum_bytes = BTRFS_I(inode)->csum_bytes;
5389         spin_unlock(&BTRFS_I(inode)->lock);
5390
5391         if (root->fs_info->quota_enabled) {
5392                 ret = btrfs_qgroup_reserve(root, num_bytes +
5393                                            nr_extents * root->nodesize);
5394                 if (ret)
5395                         goto out_fail;
5396         }
5397
5398         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5399         if (unlikely(ret)) {
5400                 if (root->fs_info->quota_enabled)
5401                         btrfs_qgroup_free(root, num_bytes +
5402                                                 nr_extents * root->nodesize);
5403                 goto out_fail;
5404         }
5405
5406         spin_lock(&BTRFS_I(inode)->lock);
5407         if (extra_reserve) {
5408                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5409                         &BTRFS_I(inode)->runtime_flags);
5410                 nr_extents--;
5411         }
5412         BTRFS_I(inode)->reserved_extents += nr_extents;
5413         spin_unlock(&BTRFS_I(inode)->lock);
5414
5415         if (delalloc_lock)
5416                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5417
5418         if (to_reserve)
5419                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5420                                               btrfs_ino(inode), to_reserve, 1);
5421         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5422
5423         return 0;
5424
5425 out_fail:
5426         spin_lock(&BTRFS_I(inode)->lock);
5427         dropped = drop_outstanding_extent(inode, num_bytes);
5428         /*
5429          * If the inodes csum_bytes is the same as the original
5430          * csum_bytes then we know we haven't raced with any free()ers
5431          * so we can just reduce our inodes csum bytes and carry on.
5432          */
5433         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5434                 calc_csum_metadata_size(inode, num_bytes, 0);
5435         } else {
5436                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5437                 u64 bytes;
5438
5439                 /*
5440                  * This is tricky, but first we need to figure out how much we
5441                  * free'd from any free-ers that occured during this
5442                  * reservation, so we reset ->csum_bytes to the csum_bytes
5443                  * before we dropped our lock, and then call the free for the
5444                  * number of bytes that were freed while we were trying our
5445                  * reservation.
5446                  */
5447                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5448                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5449                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5450
5451
5452                 /*
5453                  * Now we need to see how much we would have freed had we not
5454                  * been making this reservation and our ->csum_bytes were not
5455                  * artificially inflated.
5456                  */
5457                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5458                 bytes = csum_bytes - orig_csum_bytes;
5459                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5460
5461                 /*
5462                  * Now reset ->csum_bytes to what it should be.  If bytes is
5463                  * more than to_free then we would have free'd more space had we
5464                  * not had an artificially high ->csum_bytes, so we need to free
5465                  * the remainder.  If bytes is the same or less then we don't
5466                  * need to do anything, the other free-ers did the correct
5467                  * thing.
5468                  */
5469                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5470                 if (bytes > to_free)
5471                         to_free = bytes - to_free;
5472                 else
5473                         to_free = 0;
5474         }
5475         spin_unlock(&BTRFS_I(inode)->lock);
5476         if (dropped)
5477                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5478
5479         if (to_free) {
5480                 btrfs_block_rsv_release(root, block_rsv, to_free);
5481                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5482                                               btrfs_ino(inode), to_free, 0);
5483         }
5484         if (delalloc_lock)
5485                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5486         return ret;
5487 }
5488
5489 /**
5490  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5491  * @inode: the inode to release the reservation for
5492  * @num_bytes: the number of bytes we're releasing
5493  *
5494  * This will release the metadata reservation for an inode.  This can be called
5495  * once we complete IO for a given set of bytes to release their metadata
5496  * reservations.
5497  */
5498 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5499 {
5500         struct btrfs_root *root = BTRFS_I(inode)->root;
5501         u64 to_free = 0;
5502         unsigned dropped;
5503
5504         num_bytes = ALIGN(num_bytes, root->sectorsize);
5505         spin_lock(&BTRFS_I(inode)->lock);
5506         dropped = drop_outstanding_extent(inode, num_bytes);
5507
5508         if (num_bytes)
5509                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5510         spin_unlock(&BTRFS_I(inode)->lock);
5511         if (dropped > 0)
5512                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5513
5514         if (btrfs_test_is_dummy_root(root))
5515                 return;
5516
5517         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5518                                       btrfs_ino(inode), to_free, 0);
5519         if (root->fs_info->quota_enabled) {
5520                 btrfs_qgroup_free(root, num_bytes +
5521                                         dropped * root->nodesize);
5522         }
5523
5524         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5525                                 to_free);
5526 }
5527
5528 /**
5529  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5530  * @inode: inode we're writing to
5531  * @num_bytes: the number of bytes we want to allocate
5532  *
5533  * This will do the following things
5534  *
5535  * o reserve space in the data space info for num_bytes
5536  * o reserve space in the metadata space info based on number of outstanding
5537  *   extents and how much csums will be needed
5538  * o add to the inodes ->delalloc_bytes
5539  * o add it to the fs_info's delalloc inodes list.
5540  *
5541  * This will return 0 for success and -ENOSPC if there is no space left.
5542  */
5543 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5544 {
5545         int ret;
5546
5547         ret = btrfs_check_data_free_space(inode, num_bytes);
5548         if (ret)
5549                 return ret;
5550
5551         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5552         if (ret) {
5553                 btrfs_free_reserved_data_space(inode, num_bytes);
5554                 return ret;
5555         }
5556
5557         return 0;
5558 }
5559
5560 /**
5561  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5562  * @inode: inode we're releasing space for
5563  * @num_bytes: the number of bytes we want to free up
5564  *
5565  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5566  * called in the case that we don't need the metadata AND data reservations
5567  * anymore.  So if there is an error or we insert an inline extent.
5568  *
5569  * This function will release the metadata space that was not used and will
5570  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5571  * list if there are no delalloc bytes left.
5572  */
5573 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5574 {
5575         btrfs_delalloc_release_metadata(inode, num_bytes);
5576         btrfs_free_reserved_data_space(inode, num_bytes);
5577 }
5578
5579 static int update_block_group(struct btrfs_trans_handle *trans,
5580                               struct btrfs_root *root, u64 bytenr,
5581                               u64 num_bytes, int alloc)
5582 {
5583         struct btrfs_block_group_cache *cache = NULL;
5584         struct btrfs_fs_info *info = root->fs_info;
5585         u64 total = num_bytes;
5586         u64 old_val;
5587         u64 byte_in_group;
5588         int factor;
5589
5590         /* block accounting for super block */
5591         spin_lock(&info->delalloc_root_lock);
5592         old_val = btrfs_super_bytes_used(info->super_copy);
5593         if (alloc)
5594                 old_val += num_bytes;
5595         else
5596                 old_val -= num_bytes;
5597         btrfs_set_super_bytes_used(info->super_copy, old_val);
5598         spin_unlock(&info->delalloc_root_lock);
5599
5600         while (total) {
5601                 cache = btrfs_lookup_block_group(info, bytenr);
5602                 if (!cache)
5603                         return -ENOENT;
5604                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5605                                     BTRFS_BLOCK_GROUP_RAID1 |
5606                                     BTRFS_BLOCK_GROUP_RAID10))
5607                         factor = 2;
5608                 else
5609                         factor = 1;
5610                 /*
5611                  * If this block group has free space cache written out, we
5612                  * need to make sure to load it if we are removing space.  This
5613                  * is because we need the unpinning stage to actually add the
5614                  * space back to the block group, otherwise we will leak space.
5615                  */
5616                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5617                         cache_block_group(cache, 1);
5618
5619                 byte_in_group = bytenr - cache->key.objectid;
5620                 WARN_ON(byte_in_group > cache->key.offset);
5621
5622                 spin_lock(&cache->space_info->lock);
5623                 spin_lock(&cache->lock);
5624
5625                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5626                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5627                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5628
5629                 old_val = btrfs_block_group_used(&cache->item);
5630                 num_bytes = min(total, cache->key.offset - byte_in_group);
5631                 if (alloc) {
5632                         old_val += num_bytes;
5633                         btrfs_set_block_group_used(&cache->item, old_val);
5634                         cache->reserved -= num_bytes;
5635                         cache->space_info->bytes_reserved -= num_bytes;
5636                         cache->space_info->bytes_used += num_bytes;
5637                         cache->space_info->disk_used += num_bytes * factor;
5638                         spin_unlock(&cache->lock);
5639                         spin_unlock(&cache->space_info->lock);
5640                 } else {
5641                         old_val -= num_bytes;
5642                         btrfs_set_block_group_used(&cache->item, old_val);
5643                         cache->pinned += num_bytes;
5644                         cache->space_info->bytes_pinned += num_bytes;
5645                         cache->space_info->bytes_used -= num_bytes;
5646                         cache->space_info->disk_used -= num_bytes * factor;
5647                         spin_unlock(&cache->lock);
5648                         spin_unlock(&cache->space_info->lock);
5649
5650                         set_extent_dirty(info->pinned_extents,
5651                                          bytenr, bytenr + num_bytes - 1,
5652                                          GFP_NOFS | __GFP_NOFAIL);
5653                         /*
5654                          * No longer have used bytes in this block group, queue
5655                          * it for deletion.
5656                          */
5657                         if (old_val == 0) {
5658                                 spin_lock(&info->unused_bgs_lock);
5659                                 if (list_empty(&cache->bg_list)) {
5660                                         btrfs_get_block_group(cache);
5661                                         list_add_tail(&cache->bg_list,
5662                                                       &info->unused_bgs);
5663                                 }
5664                                 spin_unlock(&info->unused_bgs_lock);
5665                         }
5666                 }
5667
5668                 spin_lock(&trans->transaction->dirty_bgs_lock);
5669                 if (list_empty(&cache->dirty_list)) {
5670                         list_add_tail(&cache->dirty_list,
5671                                       &trans->transaction->dirty_bgs);
5672                                 trans->transaction->num_dirty_bgs++;
5673                         btrfs_get_block_group(cache);
5674                 }
5675                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5676
5677                 btrfs_put_block_group(cache);
5678                 total -= num_bytes;
5679                 bytenr += num_bytes;
5680         }
5681         return 0;
5682 }
5683
5684 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5685 {
5686         struct btrfs_block_group_cache *cache;
5687         u64 bytenr;
5688
5689         spin_lock(&root->fs_info->block_group_cache_lock);
5690         bytenr = root->fs_info->first_logical_byte;
5691         spin_unlock(&root->fs_info->block_group_cache_lock);
5692
5693         if (bytenr < (u64)-1)
5694                 return bytenr;
5695
5696         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5697         if (!cache)
5698                 return 0;
5699
5700         bytenr = cache->key.objectid;
5701         btrfs_put_block_group(cache);
5702
5703         return bytenr;
5704 }
5705
5706 static int pin_down_extent(struct btrfs_root *root,
5707                            struct btrfs_block_group_cache *cache,
5708                            u64 bytenr, u64 num_bytes, int reserved)
5709 {
5710         spin_lock(&cache->space_info->lock);
5711         spin_lock(&cache->lock);
5712         cache->pinned += num_bytes;
5713         cache->space_info->bytes_pinned += num_bytes;
5714         if (reserved) {
5715                 cache->reserved -= num_bytes;
5716                 cache->space_info->bytes_reserved -= num_bytes;
5717         }
5718         spin_unlock(&cache->lock);
5719         spin_unlock(&cache->space_info->lock);
5720
5721         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5722                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5723         if (reserved)
5724                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5725         return 0;
5726 }
5727
5728 /*
5729  * this function must be called within transaction
5730  */
5731 int btrfs_pin_extent(struct btrfs_root *root,
5732                      u64 bytenr, u64 num_bytes, int reserved)
5733 {
5734         struct btrfs_block_group_cache *cache;
5735
5736         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5737         BUG_ON(!cache); /* Logic error */
5738
5739         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5740
5741         btrfs_put_block_group(cache);
5742         return 0;
5743 }
5744
5745 /*
5746  * this function must be called within transaction
5747  */
5748 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5749                                     u64 bytenr, u64 num_bytes)
5750 {
5751         struct btrfs_block_group_cache *cache;
5752         int ret;
5753
5754         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5755         if (!cache)
5756                 return -EINVAL;
5757
5758         /*
5759          * pull in the free space cache (if any) so that our pin
5760          * removes the free space from the cache.  We have load_only set
5761          * to one because the slow code to read in the free extents does check
5762          * the pinned extents.
5763          */
5764         cache_block_group(cache, 1);
5765
5766         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5767
5768         /* remove us from the free space cache (if we're there at all) */
5769         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5770         btrfs_put_block_group(cache);
5771         return ret;
5772 }
5773
5774 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5775 {
5776         int ret;
5777         struct btrfs_block_group_cache *block_group;
5778         struct btrfs_caching_control *caching_ctl;
5779
5780         block_group = btrfs_lookup_block_group(root->fs_info, start);
5781         if (!block_group)
5782                 return -EINVAL;
5783
5784         cache_block_group(block_group, 0);
5785         caching_ctl = get_caching_control(block_group);
5786
5787         if (!caching_ctl) {
5788                 /* Logic error */
5789                 BUG_ON(!block_group_cache_done(block_group));
5790                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5791         } else {
5792                 mutex_lock(&caching_ctl->mutex);
5793
5794                 if (start >= caching_ctl->progress) {
5795                         ret = add_excluded_extent(root, start, num_bytes);
5796                 } else if (start + num_bytes <= caching_ctl->progress) {
5797                         ret = btrfs_remove_free_space(block_group,
5798                                                       start, num_bytes);
5799                 } else {
5800                         num_bytes = caching_ctl->progress - start;
5801                         ret = btrfs_remove_free_space(block_group,
5802                                                       start, num_bytes);
5803                         if (ret)
5804                                 goto out_lock;
5805
5806                         num_bytes = (start + num_bytes) -
5807                                 caching_ctl->progress;
5808                         start = caching_ctl->progress;
5809                         ret = add_excluded_extent(root, start, num_bytes);
5810                 }
5811 out_lock:
5812                 mutex_unlock(&caching_ctl->mutex);
5813                 put_caching_control(caching_ctl);
5814         }
5815         btrfs_put_block_group(block_group);
5816         return ret;
5817 }
5818
5819 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5820                                  struct extent_buffer *eb)
5821 {
5822         struct btrfs_file_extent_item *item;
5823         struct btrfs_key key;
5824         int found_type;
5825         int i;
5826
5827         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5828                 return 0;
5829
5830         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5831                 btrfs_item_key_to_cpu(eb, &key, i);
5832                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5833                         continue;
5834                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5835                 found_type = btrfs_file_extent_type(eb, item);
5836                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5837                         continue;
5838                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5839                         continue;
5840                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5841                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5842                 __exclude_logged_extent(log, key.objectid, key.offset);
5843         }
5844
5845         return 0;
5846 }
5847
5848 /**
5849  * btrfs_update_reserved_bytes - update the block_group and space info counters
5850  * @cache:      The cache we are manipulating
5851  * @num_bytes:  The number of bytes in question
5852  * @reserve:    One of the reservation enums
5853  * @delalloc:   The blocks are allocated for the delalloc write
5854  *
5855  * This is called by the allocator when it reserves space, or by somebody who is
5856  * freeing space that was never actually used on disk.  For example if you
5857  * reserve some space for a new leaf in transaction A and before transaction A
5858  * commits you free that leaf, you call this with reserve set to 0 in order to
5859  * clear the reservation.
5860  *
5861  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5862  * ENOSPC accounting.  For data we handle the reservation through clearing the
5863  * delalloc bits in the io_tree.  We have to do this since we could end up
5864  * allocating less disk space for the amount of data we have reserved in the
5865  * case of compression.
5866  *
5867  * If this is a reservation and the block group has become read only we cannot
5868  * make the reservation and return -EAGAIN, otherwise this function always
5869  * succeeds.
5870  */
5871 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5872                                        u64 num_bytes, int reserve, int delalloc)
5873 {
5874         struct btrfs_space_info *space_info = cache->space_info;
5875         int ret = 0;
5876
5877         spin_lock(&space_info->lock);
5878         spin_lock(&cache->lock);
5879         if (reserve != RESERVE_FREE) {
5880                 if (cache->ro) {
5881                         ret = -EAGAIN;
5882                 } else {
5883                         cache->reserved += num_bytes;
5884                         space_info->bytes_reserved += num_bytes;
5885                         if (reserve == RESERVE_ALLOC) {
5886                                 trace_btrfs_space_reservation(cache->fs_info,
5887                                                 "space_info", space_info->flags,
5888                                                 num_bytes, 0);
5889                                 space_info->bytes_may_use -= num_bytes;
5890                         }
5891
5892                         if (delalloc)
5893                                 cache->delalloc_bytes += num_bytes;
5894                 }
5895         } else {
5896                 if (cache->ro)
5897                         space_info->bytes_readonly += num_bytes;
5898                 cache->reserved -= num_bytes;
5899                 space_info->bytes_reserved -= num_bytes;
5900
5901                 if (delalloc)
5902                         cache->delalloc_bytes -= num_bytes;
5903         }
5904         spin_unlock(&cache->lock);
5905         spin_unlock(&space_info->lock);
5906         return ret;
5907 }
5908
5909 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5910                                 struct btrfs_root *root)
5911 {
5912         struct btrfs_fs_info *fs_info = root->fs_info;
5913         struct btrfs_caching_control *next;
5914         struct btrfs_caching_control *caching_ctl;
5915         struct btrfs_block_group_cache *cache;
5916
5917         down_write(&fs_info->commit_root_sem);
5918
5919         list_for_each_entry_safe(caching_ctl, next,
5920                                  &fs_info->caching_block_groups, list) {
5921                 cache = caching_ctl->block_group;
5922                 if (block_group_cache_done(cache)) {
5923                         cache->last_byte_to_unpin = (u64)-1;
5924                         list_del_init(&caching_ctl->list);
5925                         put_caching_control(caching_ctl);
5926                 } else {
5927                         cache->last_byte_to_unpin = caching_ctl->progress;
5928                 }
5929         }
5930
5931         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5932                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5933         else
5934                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5935
5936         up_write(&fs_info->commit_root_sem);
5937
5938         update_global_block_rsv(fs_info);
5939 }
5940
5941 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
5942                               const bool return_free_space)
5943 {
5944         struct btrfs_fs_info *fs_info = root->fs_info;
5945         struct btrfs_block_group_cache *cache = NULL;
5946         struct btrfs_space_info *space_info;
5947         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5948         u64 len;
5949         bool readonly;
5950
5951         while (start <= end) {
5952                 readonly = false;
5953                 if (!cache ||
5954                     start >= cache->key.objectid + cache->key.offset) {
5955                         if (cache)
5956                                 btrfs_put_block_group(cache);
5957                         cache = btrfs_lookup_block_group(fs_info, start);
5958                         BUG_ON(!cache); /* Logic error */
5959                 }
5960
5961                 len = cache->key.objectid + cache->key.offset - start;
5962                 len = min(len, end + 1 - start);
5963
5964                 if (start < cache->last_byte_to_unpin) {
5965                         len = min(len, cache->last_byte_to_unpin - start);
5966                         if (return_free_space)
5967                                 btrfs_add_free_space(cache, start, len);
5968                 }
5969
5970                 start += len;
5971                 space_info = cache->space_info;
5972
5973                 spin_lock(&space_info->lock);
5974                 spin_lock(&cache->lock);
5975                 cache->pinned -= len;
5976                 space_info->bytes_pinned -= len;
5977                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
5978                 if (cache->ro) {
5979                         space_info->bytes_readonly += len;
5980                         readonly = true;
5981                 }
5982                 spin_unlock(&cache->lock);
5983                 if (!readonly && global_rsv->space_info == space_info) {
5984                         spin_lock(&global_rsv->lock);
5985                         if (!global_rsv->full) {
5986                                 len = min(len, global_rsv->size -
5987                                           global_rsv->reserved);
5988                                 global_rsv->reserved += len;
5989                                 space_info->bytes_may_use += len;
5990                                 if (global_rsv->reserved >= global_rsv->size)
5991                                         global_rsv->full = 1;
5992                         }
5993                         spin_unlock(&global_rsv->lock);
5994                 }
5995                 spin_unlock(&space_info->lock);
5996         }
5997
5998         if (cache)
5999                 btrfs_put_block_group(cache);
6000         return 0;
6001 }
6002
6003 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6004                                struct btrfs_root *root)
6005 {
6006         struct btrfs_fs_info *fs_info = root->fs_info;
6007         struct extent_io_tree *unpin;
6008         u64 start;
6009         u64 end;
6010         int ret;
6011
6012         if (trans->aborted)
6013                 return 0;
6014
6015         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6016                 unpin = &fs_info->freed_extents[1];
6017         else
6018                 unpin = &fs_info->freed_extents[0];
6019
6020         while (1) {
6021                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6022                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6023                                             EXTENT_DIRTY, NULL);
6024                 if (ret) {
6025                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6026                         break;
6027                 }
6028
6029                 if (btrfs_test_opt(root, DISCARD))
6030                         ret = btrfs_discard_extent(root, start,
6031                                                    end + 1 - start, NULL);
6032
6033                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
6034                 unpin_extent_range(root, start, end, true);
6035                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6036                 cond_resched();
6037         }
6038
6039         return 0;
6040 }
6041
6042 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6043                              u64 owner, u64 root_objectid)
6044 {
6045         struct btrfs_space_info *space_info;
6046         u64 flags;
6047
6048         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6049                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6050                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6051                 else
6052                         flags = BTRFS_BLOCK_GROUP_METADATA;
6053         } else {
6054                 flags = BTRFS_BLOCK_GROUP_DATA;
6055         }
6056
6057         space_info = __find_space_info(fs_info, flags);
6058         BUG_ON(!space_info); /* Logic bug */
6059         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6060 }
6061
6062
6063 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6064                                 struct btrfs_root *root,
6065                                 u64 bytenr, u64 num_bytes, u64 parent,
6066                                 u64 root_objectid, u64 owner_objectid,
6067                                 u64 owner_offset, int refs_to_drop,
6068                                 struct btrfs_delayed_extent_op *extent_op,
6069                                 int no_quota)
6070 {
6071         struct btrfs_key key;
6072         struct btrfs_path *path;
6073         struct btrfs_fs_info *info = root->fs_info;
6074         struct btrfs_root *extent_root = info->extent_root;
6075         struct extent_buffer *leaf;
6076         struct btrfs_extent_item *ei;
6077         struct btrfs_extent_inline_ref *iref;
6078         int ret;
6079         int is_data;
6080         int extent_slot = 0;
6081         int found_extent = 0;
6082         int num_to_del = 1;
6083         u32 item_size;
6084         u64 refs;
6085         int last_ref = 0;
6086         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_SUB_EXCL;
6087         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6088                                                  SKINNY_METADATA);
6089
6090         if (!info->quota_enabled || !is_fstree(root_objectid))
6091                 no_quota = 1;
6092
6093         path = btrfs_alloc_path();
6094         if (!path)
6095                 return -ENOMEM;
6096
6097         path->reada = 1;
6098         path->leave_spinning = 1;
6099
6100         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6101         BUG_ON(!is_data && refs_to_drop != 1);
6102
6103         if (is_data)
6104                 skinny_metadata = 0;
6105
6106         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6107                                     bytenr, num_bytes, parent,
6108                                     root_objectid, owner_objectid,
6109                                     owner_offset);
6110         if (ret == 0) {
6111                 extent_slot = path->slots[0];
6112                 while (extent_slot >= 0) {
6113                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6114                                               extent_slot);
6115                         if (key.objectid != bytenr)
6116                                 break;
6117                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6118                             key.offset == num_bytes) {
6119                                 found_extent = 1;
6120                                 break;
6121                         }
6122                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6123                             key.offset == owner_objectid) {
6124                                 found_extent = 1;
6125                                 break;
6126                         }
6127                         if (path->slots[0] - extent_slot > 5)
6128                                 break;
6129                         extent_slot--;
6130                 }
6131 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6132                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6133                 if (found_extent && item_size < sizeof(*ei))
6134                         found_extent = 0;
6135 #endif
6136                 if (!found_extent) {
6137                         BUG_ON(iref);
6138                         ret = remove_extent_backref(trans, extent_root, path,
6139                                                     NULL, refs_to_drop,
6140                                                     is_data, &last_ref);
6141                         if (ret) {
6142                                 btrfs_abort_transaction(trans, extent_root, ret);
6143                                 goto out;
6144                         }
6145                         btrfs_release_path(path);
6146                         path->leave_spinning = 1;
6147
6148                         key.objectid = bytenr;
6149                         key.type = BTRFS_EXTENT_ITEM_KEY;
6150                         key.offset = num_bytes;
6151
6152                         if (!is_data && skinny_metadata) {
6153                                 key.type = BTRFS_METADATA_ITEM_KEY;
6154                                 key.offset = owner_objectid;
6155                         }
6156
6157                         ret = btrfs_search_slot(trans, extent_root,
6158                                                 &key, path, -1, 1);
6159                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6160                                 /*
6161                                  * Couldn't find our skinny metadata item,
6162                                  * see if we have ye olde extent item.
6163                                  */
6164                                 path->slots[0]--;
6165                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6166                                                       path->slots[0]);
6167                                 if (key.objectid == bytenr &&
6168                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6169                                     key.offset == num_bytes)
6170                                         ret = 0;
6171                         }
6172
6173                         if (ret > 0 && skinny_metadata) {
6174                                 skinny_metadata = false;
6175                                 key.objectid = bytenr;
6176                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6177                                 key.offset = num_bytes;
6178                                 btrfs_release_path(path);
6179                                 ret = btrfs_search_slot(trans, extent_root,
6180                                                         &key, path, -1, 1);
6181                         }
6182
6183                         if (ret) {
6184                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6185                                         ret, bytenr);
6186                                 if (ret > 0)
6187                                         btrfs_print_leaf(extent_root,
6188                                                          path->nodes[0]);
6189                         }
6190                         if (ret < 0) {
6191                                 btrfs_abort_transaction(trans, extent_root, ret);
6192                                 goto out;
6193                         }
6194                         extent_slot = path->slots[0];
6195                 }
6196         } else if (WARN_ON(ret == -ENOENT)) {
6197                 btrfs_print_leaf(extent_root, path->nodes[0]);
6198                 btrfs_err(info,
6199                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6200                         bytenr, parent, root_objectid, owner_objectid,
6201                         owner_offset);
6202                 btrfs_abort_transaction(trans, extent_root, ret);
6203                 goto out;
6204         } else {
6205                 btrfs_abort_transaction(trans, extent_root, ret);
6206                 goto out;
6207         }
6208
6209         leaf = path->nodes[0];
6210         item_size = btrfs_item_size_nr(leaf, extent_slot);
6211 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6212         if (item_size < sizeof(*ei)) {
6213                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6214                 ret = convert_extent_item_v0(trans, extent_root, path,
6215                                              owner_objectid, 0);
6216                 if (ret < 0) {
6217                         btrfs_abort_transaction(trans, extent_root, ret);
6218                         goto out;
6219                 }
6220
6221                 btrfs_release_path(path);
6222                 path->leave_spinning = 1;
6223
6224                 key.objectid = bytenr;
6225                 key.type = BTRFS_EXTENT_ITEM_KEY;
6226                 key.offset = num_bytes;
6227
6228                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6229                                         -1, 1);
6230                 if (ret) {
6231                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6232                                 ret, bytenr);
6233                         btrfs_print_leaf(extent_root, path->nodes[0]);
6234                 }
6235                 if (ret < 0) {
6236                         btrfs_abort_transaction(trans, extent_root, ret);
6237                         goto out;
6238                 }
6239
6240                 extent_slot = path->slots[0];
6241                 leaf = path->nodes[0];
6242                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6243         }
6244 #endif
6245         BUG_ON(item_size < sizeof(*ei));
6246         ei = btrfs_item_ptr(leaf, extent_slot,
6247                             struct btrfs_extent_item);
6248         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6249             key.type == BTRFS_EXTENT_ITEM_KEY) {
6250                 struct btrfs_tree_block_info *bi;
6251                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6252                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6253                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6254         }
6255
6256         refs = btrfs_extent_refs(leaf, ei);
6257         if (refs < refs_to_drop) {
6258                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6259                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6260                 ret = -EINVAL;
6261                 btrfs_abort_transaction(trans, extent_root, ret);
6262                 goto out;
6263         }
6264         refs -= refs_to_drop;
6265
6266         if (refs > 0) {
6267                 type = BTRFS_QGROUP_OPER_SUB_SHARED;
6268                 if (extent_op)
6269                         __run_delayed_extent_op(extent_op, leaf, ei);
6270                 /*
6271                  * In the case of inline back ref, reference count will
6272                  * be updated by remove_extent_backref
6273                  */
6274                 if (iref) {
6275                         BUG_ON(!found_extent);
6276                 } else {
6277                         btrfs_set_extent_refs(leaf, ei, refs);
6278                         btrfs_mark_buffer_dirty(leaf);
6279                 }
6280                 if (found_extent) {
6281                         ret = remove_extent_backref(trans, extent_root, path,
6282                                                     iref, refs_to_drop,
6283                                                     is_data, &last_ref);
6284                         if (ret) {
6285                                 btrfs_abort_transaction(trans, extent_root, ret);
6286                                 goto out;
6287                         }
6288                 }
6289                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6290                                  root_objectid);
6291         } else {
6292                 if (found_extent) {
6293                         BUG_ON(is_data && refs_to_drop !=
6294                                extent_data_ref_count(root, path, iref));
6295                         if (iref) {
6296                                 BUG_ON(path->slots[0] != extent_slot);
6297                         } else {
6298                                 BUG_ON(path->slots[0] != extent_slot + 1);
6299                                 path->slots[0] = extent_slot;
6300                                 num_to_del = 2;
6301                         }
6302                 }
6303
6304                 last_ref = 1;
6305                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6306                                       num_to_del);
6307                 if (ret) {
6308                         btrfs_abort_transaction(trans, extent_root, ret);
6309                         goto out;
6310                 }
6311                 btrfs_release_path(path);
6312
6313                 if (is_data) {
6314                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6315                         if (ret) {
6316                                 btrfs_abort_transaction(trans, extent_root, ret);
6317                                 goto out;
6318                         }
6319                 }
6320
6321                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6322                 if (ret) {
6323                         btrfs_abort_transaction(trans, extent_root, ret);
6324                         goto out;
6325                 }
6326         }
6327         btrfs_release_path(path);
6328
6329         /* Deal with the quota accounting */
6330         if (!ret && last_ref && !no_quota) {
6331                 int mod_seq = 0;
6332
6333                 if (owner_objectid >= BTRFS_FIRST_FREE_OBJECTID &&
6334                     type == BTRFS_QGROUP_OPER_SUB_SHARED)
6335                         mod_seq = 1;
6336
6337                 ret = btrfs_qgroup_record_ref(trans, info, root_objectid,
6338                                               bytenr, num_bytes, type,
6339                                               mod_seq);
6340         }
6341 out:
6342         btrfs_free_path(path);
6343         return ret;
6344 }
6345
6346 /*
6347  * when we free an block, it is possible (and likely) that we free the last
6348  * delayed ref for that extent as well.  This searches the delayed ref tree for
6349  * a given extent, and if there are no other delayed refs to be processed, it
6350  * removes it from the tree.
6351  */
6352 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6353                                       struct btrfs_root *root, u64 bytenr)
6354 {
6355         struct btrfs_delayed_ref_head *head;
6356         struct btrfs_delayed_ref_root *delayed_refs;
6357         int ret = 0;
6358
6359         delayed_refs = &trans->transaction->delayed_refs;
6360         spin_lock(&delayed_refs->lock);
6361         head = btrfs_find_delayed_ref_head(trans, bytenr);
6362         if (!head)
6363                 goto out_delayed_unlock;
6364
6365         spin_lock(&head->lock);
6366         if (rb_first(&head->ref_root))
6367                 goto out;
6368
6369         if (head->extent_op) {
6370                 if (!head->must_insert_reserved)
6371                         goto out;
6372                 btrfs_free_delayed_extent_op(head->extent_op);
6373                 head->extent_op = NULL;
6374         }
6375
6376         /*
6377          * waiting for the lock here would deadlock.  If someone else has it
6378          * locked they are already in the process of dropping it anyway
6379          */
6380         if (!mutex_trylock(&head->mutex))
6381                 goto out;
6382
6383         /*
6384          * at this point we have a head with no other entries.  Go
6385          * ahead and process it.
6386          */
6387         head->node.in_tree = 0;
6388         rb_erase(&head->href_node, &delayed_refs->href_root);
6389
6390         atomic_dec(&delayed_refs->num_entries);
6391
6392         /*
6393          * we don't take a ref on the node because we're removing it from the
6394          * tree, so we just steal the ref the tree was holding.
6395          */
6396         delayed_refs->num_heads--;
6397         if (head->processing == 0)
6398                 delayed_refs->num_heads_ready--;
6399         head->processing = 0;
6400         spin_unlock(&head->lock);
6401         spin_unlock(&delayed_refs->lock);
6402
6403         BUG_ON(head->extent_op);
6404         if (head->must_insert_reserved)
6405                 ret = 1;
6406
6407         mutex_unlock(&head->mutex);
6408         btrfs_put_delayed_ref(&head->node);
6409         return ret;
6410 out:
6411         spin_unlock(&head->lock);
6412
6413 out_delayed_unlock:
6414         spin_unlock(&delayed_refs->lock);
6415         return 0;
6416 }
6417
6418 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6419                            struct btrfs_root *root,
6420                            struct extent_buffer *buf,
6421                            u64 parent, int last_ref)
6422 {
6423         int pin = 1;
6424         int ret;
6425
6426         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6427                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6428                                         buf->start, buf->len,
6429                                         parent, root->root_key.objectid,
6430                                         btrfs_header_level(buf),
6431                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6432                 BUG_ON(ret); /* -ENOMEM */
6433         }
6434
6435         if (!last_ref)
6436                 return;
6437
6438         if (btrfs_header_generation(buf) == trans->transid) {
6439                 struct btrfs_block_group_cache *cache;
6440
6441                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6442                         ret = check_ref_cleanup(trans, root, buf->start);
6443                         if (!ret)
6444                                 goto out;
6445                 }
6446
6447                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6448
6449                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6450                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6451                         btrfs_put_block_group(cache);
6452                         goto out;
6453                 }
6454
6455                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6456
6457                 btrfs_add_free_space(cache, buf->start, buf->len);
6458                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6459                 btrfs_put_block_group(cache);
6460                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6461                 pin = 0;
6462         }
6463 out:
6464         if (pin)
6465                 add_pinned_bytes(root->fs_info, buf->len,
6466                                  btrfs_header_level(buf),
6467                                  root->root_key.objectid);
6468
6469         /*
6470          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6471          * anymore.
6472          */
6473         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6474 }
6475
6476 /* Can return -ENOMEM */
6477 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6478                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6479                       u64 owner, u64 offset, int no_quota)
6480 {
6481         int ret;
6482         struct btrfs_fs_info *fs_info = root->fs_info;
6483
6484         if (btrfs_test_is_dummy_root(root))
6485                 return 0;
6486
6487         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6488
6489         /*
6490          * tree log blocks never actually go into the extent allocation
6491          * tree, just update pinning info and exit early.
6492          */
6493         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6494                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6495                 /* unlocks the pinned mutex */
6496                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6497                 ret = 0;
6498         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6499                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6500                                         num_bytes,
6501                                         parent, root_objectid, (int)owner,
6502                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6503         } else {
6504                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6505                                                 num_bytes,
6506                                                 parent, root_objectid, owner,
6507                                                 offset, BTRFS_DROP_DELAYED_REF,
6508                                                 NULL, no_quota);
6509         }
6510         return ret;
6511 }
6512
6513 /*
6514  * when we wait for progress in the block group caching, its because
6515  * our allocation attempt failed at least once.  So, we must sleep
6516  * and let some progress happen before we try again.
6517  *
6518  * This function will sleep at least once waiting for new free space to
6519  * show up, and then it will check the block group free space numbers
6520  * for our min num_bytes.  Another option is to have it go ahead
6521  * and look in the rbtree for a free extent of a given size, but this
6522  * is a good start.
6523  *
6524  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6525  * any of the information in this block group.
6526  */
6527 static noinline void
6528 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6529                                 u64 num_bytes)
6530 {
6531         struct btrfs_caching_control *caching_ctl;
6532
6533         caching_ctl = get_caching_control(cache);
6534         if (!caching_ctl)
6535                 return;
6536
6537         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6538                    (cache->free_space_ctl->free_space >= num_bytes));
6539
6540         put_caching_control(caching_ctl);
6541 }
6542
6543 static noinline int
6544 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6545 {
6546         struct btrfs_caching_control *caching_ctl;
6547         int ret = 0;
6548
6549         caching_ctl = get_caching_control(cache);
6550         if (!caching_ctl)
6551                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6552
6553         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6554         if (cache->cached == BTRFS_CACHE_ERROR)
6555                 ret = -EIO;
6556         put_caching_control(caching_ctl);
6557         return ret;
6558 }
6559
6560 int __get_raid_index(u64 flags)
6561 {
6562         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6563                 return BTRFS_RAID_RAID10;
6564         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6565                 return BTRFS_RAID_RAID1;
6566         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6567                 return BTRFS_RAID_DUP;
6568         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6569                 return BTRFS_RAID_RAID0;
6570         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6571                 return BTRFS_RAID_RAID5;
6572         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6573                 return BTRFS_RAID_RAID6;
6574
6575         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6576 }
6577
6578 int get_block_group_index(struct btrfs_block_group_cache *cache)
6579 {
6580         return __get_raid_index(cache->flags);
6581 }
6582
6583 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6584         [BTRFS_RAID_RAID10]     = "raid10",
6585         [BTRFS_RAID_RAID1]      = "raid1",
6586         [BTRFS_RAID_DUP]        = "dup",
6587         [BTRFS_RAID_RAID0]      = "raid0",
6588         [BTRFS_RAID_SINGLE]     = "single",
6589         [BTRFS_RAID_RAID5]      = "raid5",
6590         [BTRFS_RAID_RAID6]      = "raid6",
6591 };
6592
6593 static const char *get_raid_name(enum btrfs_raid_types type)
6594 {
6595         if (type >= BTRFS_NR_RAID_TYPES)
6596                 return NULL;
6597
6598         return btrfs_raid_type_names[type];
6599 }
6600
6601 enum btrfs_loop_type {
6602         LOOP_CACHING_NOWAIT = 0,
6603         LOOP_CACHING_WAIT = 1,
6604         LOOP_ALLOC_CHUNK = 2,
6605         LOOP_NO_EMPTY_SIZE = 3,
6606 };
6607
6608 static inline void
6609 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6610                        int delalloc)
6611 {
6612         if (delalloc)
6613                 down_read(&cache->data_rwsem);
6614 }
6615
6616 static inline void
6617 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6618                        int delalloc)
6619 {
6620         btrfs_get_block_group(cache);
6621         if (delalloc)
6622                 down_read(&cache->data_rwsem);
6623 }
6624
6625 static struct btrfs_block_group_cache *
6626 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6627                    struct btrfs_free_cluster *cluster,
6628                    int delalloc)
6629 {
6630         struct btrfs_block_group_cache *used_bg;
6631         bool locked = false;
6632 again:
6633         spin_lock(&cluster->refill_lock);
6634         if (locked) {
6635                 if (used_bg == cluster->block_group)
6636                         return used_bg;
6637
6638                 up_read(&used_bg->data_rwsem);
6639                 btrfs_put_block_group(used_bg);
6640         }
6641
6642         used_bg = cluster->block_group;
6643         if (!used_bg)
6644                 return NULL;
6645
6646         if (used_bg == block_group)
6647                 return used_bg;
6648
6649         btrfs_get_block_group(used_bg);
6650
6651         if (!delalloc)
6652                 return used_bg;
6653
6654         if (down_read_trylock(&used_bg->data_rwsem))
6655                 return used_bg;
6656
6657         spin_unlock(&cluster->refill_lock);
6658         down_read(&used_bg->data_rwsem);
6659         locked = true;
6660         goto again;
6661 }
6662
6663 static inline void
6664 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6665                          int delalloc)
6666 {
6667         if (delalloc)
6668                 up_read(&cache->data_rwsem);
6669         btrfs_put_block_group(cache);
6670 }
6671
6672 /*
6673  * walks the btree of allocated extents and find a hole of a given size.
6674  * The key ins is changed to record the hole:
6675  * ins->objectid == start position
6676  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6677  * ins->offset == the size of the hole.
6678  * Any available blocks before search_start are skipped.
6679  *
6680  * If there is no suitable free space, we will record the max size of
6681  * the free space extent currently.
6682  */
6683 static noinline int find_free_extent(struct btrfs_root *orig_root,
6684                                      u64 num_bytes, u64 empty_size,
6685                                      u64 hint_byte, struct btrfs_key *ins,
6686                                      u64 flags, int delalloc)
6687 {
6688         int ret = 0;
6689         struct btrfs_root *root = orig_root->fs_info->extent_root;
6690         struct btrfs_free_cluster *last_ptr = NULL;
6691         struct btrfs_block_group_cache *block_group = NULL;
6692         u64 search_start = 0;
6693         u64 max_extent_size = 0;
6694         int empty_cluster = 2 * 1024 * 1024;
6695         struct btrfs_space_info *space_info;
6696         int loop = 0;
6697         int index = __get_raid_index(flags);
6698         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6699                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6700         bool failed_cluster_refill = false;
6701         bool failed_alloc = false;
6702         bool use_cluster = true;
6703         bool have_caching_bg = false;
6704
6705         WARN_ON(num_bytes < root->sectorsize);
6706         ins->type = BTRFS_EXTENT_ITEM_KEY;
6707         ins->objectid = 0;
6708         ins->offset = 0;
6709
6710         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6711
6712         space_info = __find_space_info(root->fs_info, flags);
6713         if (!space_info) {
6714                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6715                 return -ENOSPC;
6716         }
6717
6718         /*
6719          * If the space info is for both data and metadata it means we have a
6720          * small filesystem and we can't use the clustering stuff.
6721          */
6722         if (btrfs_mixed_space_info(space_info))
6723                 use_cluster = false;
6724
6725         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6726                 last_ptr = &root->fs_info->meta_alloc_cluster;
6727                 if (!btrfs_test_opt(root, SSD))
6728                         empty_cluster = 64 * 1024;
6729         }
6730
6731         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6732             btrfs_test_opt(root, SSD)) {
6733                 last_ptr = &root->fs_info->data_alloc_cluster;
6734         }
6735
6736         if (last_ptr) {
6737                 spin_lock(&last_ptr->lock);
6738                 if (last_ptr->block_group)
6739                         hint_byte = last_ptr->window_start;
6740                 spin_unlock(&last_ptr->lock);
6741         }
6742
6743         search_start = max(search_start, first_logical_byte(root, 0));
6744         search_start = max(search_start, hint_byte);
6745
6746         if (!last_ptr)
6747                 empty_cluster = 0;
6748
6749         if (search_start == hint_byte) {
6750                 block_group = btrfs_lookup_block_group(root->fs_info,
6751                                                        search_start);
6752                 /*
6753                  * we don't want to use the block group if it doesn't match our
6754                  * allocation bits, or if its not cached.
6755                  *
6756                  * However if we are re-searching with an ideal block group
6757                  * picked out then we don't care that the block group is cached.
6758                  */
6759                 if (block_group && block_group_bits(block_group, flags) &&
6760                     block_group->cached != BTRFS_CACHE_NO) {
6761                         down_read(&space_info->groups_sem);
6762                         if (list_empty(&block_group->list) ||
6763                             block_group->ro) {
6764                                 /*
6765                                  * someone is removing this block group,
6766                                  * we can't jump into the have_block_group
6767                                  * target because our list pointers are not
6768                                  * valid
6769                                  */
6770                                 btrfs_put_block_group(block_group);
6771                                 up_read(&space_info->groups_sem);
6772                         } else {
6773                                 index = get_block_group_index(block_group);
6774                                 btrfs_lock_block_group(block_group, delalloc);
6775                                 goto have_block_group;
6776                         }
6777                 } else if (block_group) {
6778                         btrfs_put_block_group(block_group);
6779                 }
6780         }
6781 search:
6782         have_caching_bg = false;
6783         down_read(&space_info->groups_sem);
6784         list_for_each_entry(block_group, &space_info->block_groups[index],
6785                             list) {
6786                 u64 offset;
6787                 int cached;
6788
6789                 btrfs_grab_block_group(block_group, delalloc);
6790                 search_start = block_group->key.objectid;
6791
6792                 /*
6793                  * this can happen if we end up cycling through all the
6794                  * raid types, but we want to make sure we only allocate
6795                  * for the proper type.
6796                  */
6797                 if (!block_group_bits(block_group, flags)) {
6798                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6799                                 BTRFS_BLOCK_GROUP_RAID1 |
6800                                 BTRFS_BLOCK_GROUP_RAID5 |
6801                                 BTRFS_BLOCK_GROUP_RAID6 |
6802                                 BTRFS_BLOCK_GROUP_RAID10;
6803
6804                         /*
6805                          * if they asked for extra copies and this block group
6806                          * doesn't provide them, bail.  This does allow us to
6807                          * fill raid0 from raid1.
6808                          */
6809                         if ((flags & extra) && !(block_group->flags & extra))
6810                                 goto loop;
6811                 }
6812
6813 have_block_group:
6814                 cached = block_group_cache_done(block_group);
6815                 if (unlikely(!cached)) {
6816                         ret = cache_block_group(block_group, 0);
6817                         BUG_ON(ret < 0);
6818                         ret = 0;
6819                 }
6820
6821                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6822                         goto loop;
6823                 if (unlikely(block_group->ro))
6824                         goto loop;
6825
6826                 /*
6827                  * Ok we want to try and use the cluster allocator, so
6828                  * lets look there
6829                  */
6830                 if (last_ptr) {
6831                         struct btrfs_block_group_cache *used_block_group;
6832                         unsigned long aligned_cluster;
6833                         /*
6834                          * the refill lock keeps out other
6835                          * people trying to start a new cluster
6836                          */
6837                         used_block_group = btrfs_lock_cluster(block_group,
6838                                                               last_ptr,
6839                                                               delalloc);
6840                         if (!used_block_group)
6841                                 goto refill_cluster;
6842
6843                         if (used_block_group != block_group &&
6844                             (used_block_group->ro ||
6845                              !block_group_bits(used_block_group, flags)))
6846                                 goto release_cluster;
6847
6848                         offset = btrfs_alloc_from_cluster(used_block_group,
6849                                                 last_ptr,
6850                                                 num_bytes,
6851                                                 used_block_group->key.objectid,
6852                                                 &max_extent_size);
6853                         if (offset) {
6854                                 /* we have a block, we're done */
6855                                 spin_unlock(&last_ptr->refill_lock);
6856                                 trace_btrfs_reserve_extent_cluster(root,
6857                                                 used_block_group,
6858                                                 search_start, num_bytes);
6859                                 if (used_block_group != block_group) {
6860                                         btrfs_release_block_group(block_group,
6861                                                                   delalloc);
6862                                         block_group = used_block_group;
6863                                 }
6864                                 goto checks;
6865                         }
6866
6867                         WARN_ON(last_ptr->block_group != used_block_group);
6868 release_cluster:
6869                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6870                          * set up a new clusters, so lets just skip it
6871                          * and let the allocator find whatever block
6872                          * it can find.  If we reach this point, we
6873                          * will have tried the cluster allocator
6874                          * plenty of times and not have found
6875                          * anything, so we are likely way too
6876                          * fragmented for the clustering stuff to find
6877                          * anything.
6878                          *
6879                          * However, if the cluster is taken from the
6880                          * current block group, release the cluster
6881                          * first, so that we stand a better chance of
6882                          * succeeding in the unclustered
6883                          * allocation.  */
6884                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6885                             used_block_group != block_group) {
6886                                 spin_unlock(&last_ptr->refill_lock);
6887                                 btrfs_release_block_group(used_block_group,
6888                                                           delalloc);
6889                                 goto unclustered_alloc;
6890                         }
6891
6892                         /*
6893                          * this cluster didn't work out, free it and
6894                          * start over
6895                          */
6896                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6897
6898                         if (used_block_group != block_group)
6899                                 btrfs_release_block_group(used_block_group,
6900                                                           delalloc);
6901 refill_cluster:
6902                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6903                                 spin_unlock(&last_ptr->refill_lock);
6904                                 goto unclustered_alloc;
6905                         }
6906
6907                         aligned_cluster = max_t(unsigned long,
6908                                                 empty_cluster + empty_size,
6909                                               block_group->full_stripe_len);
6910
6911                         /* allocate a cluster in this block group */
6912                         ret = btrfs_find_space_cluster(root, block_group,
6913                                                        last_ptr, search_start,
6914                                                        num_bytes,
6915                                                        aligned_cluster);
6916                         if (ret == 0) {
6917                                 /*
6918                                  * now pull our allocation out of this
6919                                  * cluster
6920                                  */
6921                                 offset = btrfs_alloc_from_cluster(block_group,
6922                                                         last_ptr,
6923                                                         num_bytes,
6924                                                         search_start,
6925                                                         &max_extent_size);
6926                                 if (offset) {
6927                                         /* we found one, proceed */
6928                                         spin_unlock(&last_ptr->refill_lock);
6929                                         trace_btrfs_reserve_extent_cluster(root,
6930                                                 block_group, search_start,
6931                                                 num_bytes);
6932                                         goto checks;
6933                                 }
6934                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6935                                    && !failed_cluster_refill) {
6936                                 spin_unlock(&last_ptr->refill_lock);
6937
6938                                 failed_cluster_refill = true;
6939                                 wait_block_group_cache_progress(block_group,
6940                                        num_bytes + empty_cluster + empty_size);
6941                                 goto have_block_group;
6942                         }
6943
6944                         /*
6945                          * at this point we either didn't find a cluster
6946                          * or we weren't able to allocate a block from our
6947                          * cluster.  Free the cluster we've been trying
6948                          * to use, and go to the next block group
6949                          */
6950                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6951                         spin_unlock(&last_ptr->refill_lock);
6952                         goto loop;
6953                 }
6954
6955 unclustered_alloc:
6956                 spin_lock(&block_group->free_space_ctl->tree_lock);
6957                 if (cached &&
6958                     block_group->free_space_ctl->free_space <
6959                     num_bytes + empty_cluster + empty_size) {
6960                         if (block_group->free_space_ctl->free_space >
6961                             max_extent_size)
6962                                 max_extent_size =
6963                                         block_group->free_space_ctl->free_space;
6964                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6965                         goto loop;
6966                 }
6967                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6968
6969                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6970                                                     num_bytes, empty_size,
6971                                                     &max_extent_size);
6972                 /*
6973                  * If we didn't find a chunk, and we haven't failed on this
6974                  * block group before, and this block group is in the middle of
6975                  * caching and we are ok with waiting, then go ahead and wait
6976                  * for progress to be made, and set failed_alloc to true.
6977                  *
6978                  * If failed_alloc is true then we've already waited on this
6979                  * block group once and should move on to the next block group.
6980                  */
6981                 if (!offset && !failed_alloc && !cached &&
6982                     loop > LOOP_CACHING_NOWAIT) {
6983                         wait_block_group_cache_progress(block_group,
6984                                                 num_bytes + empty_size);
6985                         failed_alloc = true;
6986                         goto have_block_group;
6987                 } else if (!offset) {
6988                         if (!cached)
6989                                 have_caching_bg = true;
6990                         goto loop;
6991                 }
6992 checks:
6993                 search_start = ALIGN(offset, root->stripesize);
6994
6995                 /* move on to the next group */
6996                 if (search_start + num_bytes >
6997                     block_group->key.objectid + block_group->key.offset) {
6998                         btrfs_add_free_space(block_group, offset, num_bytes);
6999                         goto loop;
7000                 }
7001
7002                 if (offset < search_start)
7003                         btrfs_add_free_space(block_group, offset,
7004                                              search_start - offset);
7005                 BUG_ON(offset > search_start);
7006
7007                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7008                                                   alloc_type, delalloc);
7009                 if (ret == -EAGAIN) {
7010                         btrfs_add_free_space(block_group, offset, num_bytes);
7011                         goto loop;
7012                 }
7013
7014                 /* we are all good, lets return */
7015                 ins->objectid = search_start;
7016                 ins->offset = num_bytes;
7017
7018                 trace_btrfs_reserve_extent(orig_root, block_group,
7019                                            search_start, num_bytes);
7020                 btrfs_release_block_group(block_group, delalloc);
7021                 break;
7022 loop:
7023                 failed_cluster_refill = false;
7024                 failed_alloc = false;
7025                 BUG_ON(index != get_block_group_index(block_group));
7026                 btrfs_release_block_group(block_group, delalloc);
7027         }
7028         up_read(&space_info->groups_sem);
7029
7030         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7031                 goto search;
7032
7033         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7034                 goto search;
7035
7036         /*
7037          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7038          *                      caching kthreads as we move along
7039          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7040          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7041          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7042          *                      again
7043          */
7044         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7045                 index = 0;
7046                 loop++;
7047                 if (loop == LOOP_ALLOC_CHUNK) {
7048                         struct btrfs_trans_handle *trans;
7049                         int exist = 0;
7050
7051                         trans = current->journal_info;
7052                         if (trans)
7053                                 exist = 1;
7054                         else
7055                                 trans = btrfs_join_transaction(root);
7056
7057                         if (IS_ERR(trans)) {
7058                                 ret = PTR_ERR(trans);
7059                                 goto out;
7060                         }
7061
7062                         ret = do_chunk_alloc(trans, root, flags,
7063                                              CHUNK_ALLOC_FORCE);
7064                         /*
7065                          * Do not bail out on ENOSPC since we
7066                          * can do more things.
7067                          */
7068                         if (ret < 0 && ret != -ENOSPC)
7069                                 btrfs_abort_transaction(trans,
7070                                                         root, ret);
7071                         else
7072                                 ret = 0;
7073                         if (!exist)
7074                                 btrfs_end_transaction(trans, root);
7075                         if (ret)
7076                                 goto out;
7077                 }
7078
7079                 if (loop == LOOP_NO_EMPTY_SIZE) {
7080                         empty_size = 0;
7081                         empty_cluster = 0;
7082                 }
7083
7084                 goto search;
7085         } else if (!ins->objectid) {
7086                 ret = -ENOSPC;
7087         } else if (ins->objectid) {
7088                 ret = 0;
7089         }
7090 out:
7091         if (ret == -ENOSPC)
7092                 ins->offset = max_extent_size;
7093         return ret;
7094 }
7095
7096 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7097                             int dump_block_groups)
7098 {
7099         struct btrfs_block_group_cache *cache;
7100         int index = 0;
7101
7102         spin_lock(&info->lock);
7103         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7104                info->flags,
7105                info->total_bytes - info->bytes_used - info->bytes_pinned -
7106                info->bytes_reserved - info->bytes_readonly,
7107                (info->full) ? "" : "not ");
7108         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7109                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7110                info->total_bytes, info->bytes_used, info->bytes_pinned,
7111                info->bytes_reserved, info->bytes_may_use,
7112                info->bytes_readonly);
7113         spin_unlock(&info->lock);
7114
7115         if (!dump_block_groups)
7116                 return;
7117
7118         down_read(&info->groups_sem);
7119 again:
7120         list_for_each_entry(cache, &info->block_groups[index], list) {
7121                 spin_lock(&cache->lock);
7122                 printk(KERN_INFO "BTRFS: "
7123                            "block group %llu has %llu bytes, "
7124                            "%llu used %llu pinned %llu reserved %s\n",
7125                        cache->key.objectid, cache->key.offset,
7126                        btrfs_block_group_used(&cache->item), cache->pinned,
7127                        cache->reserved, cache->ro ? "[readonly]" : "");
7128                 btrfs_dump_free_space(cache, bytes);
7129                 spin_unlock(&cache->lock);
7130         }
7131         if (++index < BTRFS_NR_RAID_TYPES)
7132                 goto again;
7133         up_read(&info->groups_sem);
7134 }
7135
7136 int btrfs_reserve_extent(struct btrfs_root *root,
7137                          u64 num_bytes, u64 min_alloc_size,
7138                          u64 empty_size, u64 hint_byte,
7139                          struct btrfs_key *ins, int is_data, int delalloc)
7140 {
7141         bool final_tried = false;
7142         u64 flags;
7143         int ret;
7144
7145         flags = btrfs_get_alloc_profile(root, is_data);
7146 again:
7147         WARN_ON(num_bytes < root->sectorsize);
7148         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7149                                flags, delalloc);
7150
7151         if (ret == -ENOSPC) {
7152                 if (!final_tried && ins->offset) {
7153                         num_bytes = min(num_bytes >> 1, ins->offset);
7154                         num_bytes = round_down(num_bytes, root->sectorsize);
7155                         num_bytes = max(num_bytes, min_alloc_size);
7156                         if (num_bytes == min_alloc_size)
7157                                 final_tried = true;
7158                         goto again;
7159                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7160                         struct btrfs_space_info *sinfo;
7161
7162                         sinfo = __find_space_info(root->fs_info, flags);
7163                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7164                                 flags, num_bytes);
7165                         if (sinfo)
7166                                 dump_space_info(sinfo, num_bytes, 1);
7167                 }
7168         }
7169
7170         return ret;
7171 }
7172
7173 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7174                                         u64 start, u64 len,
7175                                         int pin, int delalloc)
7176 {
7177         struct btrfs_block_group_cache *cache;
7178         int ret = 0;
7179
7180         cache = btrfs_lookup_block_group(root->fs_info, start);
7181         if (!cache) {
7182                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7183                         start);
7184                 return -ENOSPC;
7185         }
7186
7187         if (pin)
7188                 pin_down_extent(root, cache, start, len, 1);
7189         else {
7190                 if (btrfs_test_opt(root, DISCARD))
7191                         ret = btrfs_discard_extent(root, start, len, NULL);
7192                 btrfs_add_free_space(cache, start, len);
7193                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7194         }
7195         btrfs_put_block_group(cache);
7196
7197         trace_btrfs_reserved_extent_free(root, start, len);
7198
7199         return ret;
7200 }
7201
7202 int btrfs_free_reserved_extent(struct btrfs_root *root,
7203                                u64 start, u64 len, int delalloc)
7204 {
7205         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7206 }
7207
7208 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7209                                        u64 start, u64 len)
7210 {
7211         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7212 }
7213
7214 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7215                                       struct btrfs_root *root,
7216                                       u64 parent, u64 root_objectid,
7217                                       u64 flags, u64 owner, u64 offset,
7218                                       struct btrfs_key *ins, int ref_mod)
7219 {
7220         int ret;
7221         struct btrfs_fs_info *fs_info = root->fs_info;
7222         struct btrfs_extent_item *extent_item;
7223         struct btrfs_extent_inline_ref *iref;
7224         struct btrfs_path *path;
7225         struct extent_buffer *leaf;
7226         int type;
7227         u32 size;
7228
7229         if (parent > 0)
7230                 type = BTRFS_SHARED_DATA_REF_KEY;
7231         else
7232                 type = BTRFS_EXTENT_DATA_REF_KEY;
7233
7234         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7235
7236         path = btrfs_alloc_path();
7237         if (!path)
7238                 return -ENOMEM;
7239
7240         path->leave_spinning = 1;
7241         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7242                                       ins, size);
7243         if (ret) {
7244                 btrfs_free_path(path);
7245                 return ret;
7246         }
7247
7248         leaf = path->nodes[0];
7249         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7250                                      struct btrfs_extent_item);
7251         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7252         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7253         btrfs_set_extent_flags(leaf, extent_item,
7254                                flags | BTRFS_EXTENT_FLAG_DATA);
7255
7256         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7257         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7258         if (parent > 0) {
7259                 struct btrfs_shared_data_ref *ref;
7260                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7261                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7262                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7263         } else {
7264                 struct btrfs_extent_data_ref *ref;
7265                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7266                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7267                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7268                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7269                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7270         }
7271
7272         btrfs_mark_buffer_dirty(path->nodes[0]);
7273         btrfs_free_path(path);
7274
7275         /* Always set parent to 0 here since its exclusive anyway. */
7276         ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7277                                       ins->objectid, ins->offset,
7278                                       BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7279         if (ret)
7280                 return ret;
7281
7282         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7283         if (ret) { /* -ENOENT, logic error */
7284                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7285                         ins->objectid, ins->offset);
7286                 BUG();
7287         }
7288         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7289         return ret;
7290 }
7291
7292 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7293                                      struct btrfs_root *root,
7294                                      u64 parent, u64 root_objectid,
7295                                      u64 flags, struct btrfs_disk_key *key,
7296                                      int level, struct btrfs_key *ins,
7297                                      int no_quota)
7298 {
7299         int ret;
7300         struct btrfs_fs_info *fs_info = root->fs_info;
7301         struct btrfs_extent_item *extent_item;
7302         struct btrfs_tree_block_info *block_info;
7303         struct btrfs_extent_inline_ref *iref;
7304         struct btrfs_path *path;
7305         struct extent_buffer *leaf;
7306         u32 size = sizeof(*extent_item) + sizeof(*iref);
7307         u64 num_bytes = ins->offset;
7308         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7309                                                  SKINNY_METADATA);
7310
7311         if (!skinny_metadata)
7312                 size += sizeof(*block_info);
7313
7314         path = btrfs_alloc_path();
7315         if (!path) {
7316                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7317                                                    root->nodesize);
7318                 return -ENOMEM;
7319         }
7320
7321         path->leave_spinning = 1;
7322         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7323                                       ins, size);
7324         if (ret) {
7325                 btrfs_free_path(path);
7326                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7327                                                    root->nodesize);
7328                 return ret;
7329         }
7330
7331         leaf = path->nodes[0];
7332         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7333                                      struct btrfs_extent_item);
7334         btrfs_set_extent_refs(leaf, extent_item, 1);
7335         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7336         btrfs_set_extent_flags(leaf, extent_item,
7337                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7338
7339         if (skinny_metadata) {
7340                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7341                 num_bytes = root->nodesize;
7342         } else {
7343                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7344                 btrfs_set_tree_block_key(leaf, block_info, key);
7345                 btrfs_set_tree_block_level(leaf, block_info, level);
7346                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7347         }
7348
7349         if (parent > 0) {
7350                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7351                 btrfs_set_extent_inline_ref_type(leaf, iref,
7352                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7353                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7354         } else {
7355                 btrfs_set_extent_inline_ref_type(leaf, iref,
7356                                                  BTRFS_TREE_BLOCK_REF_KEY);
7357                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7358         }
7359
7360         btrfs_mark_buffer_dirty(leaf);
7361         btrfs_free_path(path);
7362
7363         if (!no_quota) {
7364                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7365                                               ins->objectid, num_bytes,
7366                                               BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7367                 if (ret)
7368                         return ret;
7369         }
7370
7371         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7372                                  1);
7373         if (ret) { /* -ENOENT, logic error */
7374                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7375                         ins->objectid, ins->offset);
7376                 BUG();
7377         }
7378
7379         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7380         return ret;
7381 }
7382
7383 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7384                                      struct btrfs_root *root,
7385                                      u64 root_objectid, u64 owner,
7386                                      u64 offset, struct btrfs_key *ins)
7387 {
7388         int ret;
7389
7390         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7391
7392         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7393                                          ins->offset, 0,
7394                                          root_objectid, owner, offset,
7395                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7396         return ret;
7397 }
7398
7399 /*
7400  * this is used by the tree logging recovery code.  It records that
7401  * an extent has been allocated and makes sure to clear the free
7402  * space cache bits as well
7403  */
7404 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7405                                    struct btrfs_root *root,
7406                                    u64 root_objectid, u64 owner, u64 offset,
7407                                    struct btrfs_key *ins)
7408 {
7409         int ret;
7410         struct btrfs_block_group_cache *block_group;
7411
7412         /*
7413          * Mixed block groups will exclude before processing the log so we only
7414          * need to do the exlude dance if this fs isn't mixed.
7415          */
7416         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7417                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7418                 if (ret)
7419                         return ret;
7420         }
7421
7422         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7423         if (!block_group)
7424                 return -EINVAL;
7425
7426         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7427                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7428         BUG_ON(ret); /* logic error */
7429         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7430                                          0, owner, offset, ins, 1);
7431         btrfs_put_block_group(block_group);
7432         return ret;
7433 }
7434
7435 static struct extent_buffer *
7436 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7437                       u64 bytenr, int level)
7438 {
7439         struct extent_buffer *buf;
7440
7441         buf = btrfs_find_create_tree_block(root, bytenr);
7442         if (!buf)
7443                 return ERR_PTR(-ENOMEM);
7444         btrfs_set_header_generation(buf, trans->transid);
7445         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7446         btrfs_tree_lock(buf);
7447         clean_tree_block(trans, root->fs_info, buf);
7448         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7449
7450         btrfs_set_lock_blocking(buf);
7451         btrfs_set_buffer_uptodate(buf);
7452
7453         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7454                 buf->log_index = root->log_transid % 2;
7455                 /*
7456                  * we allow two log transactions at a time, use different
7457                  * EXENT bit to differentiate dirty pages.
7458                  */
7459                 if (buf->log_index == 0)
7460                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7461                                         buf->start + buf->len - 1, GFP_NOFS);
7462                 else
7463                         set_extent_new(&root->dirty_log_pages, buf->start,
7464                                         buf->start + buf->len - 1, GFP_NOFS);
7465         } else {
7466                 buf->log_index = -1;
7467                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7468                          buf->start + buf->len - 1, GFP_NOFS);
7469         }
7470         trans->blocks_used++;
7471         /* this returns a buffer locked for blocking */
7472         return buf;
7473 }
7474
7475 static struct btrfs_block_rsv *
7476 use_block_rsv(struct btrfs_trans_handle *trans,
7477               struct btrfs_root *root, u32 blocksize)
7478 {
7479         struct btrfs_block_rsv *block_rsv;
7480         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7481         int ret;
7482         bool global_updated = false;
7483
7484         block_rsv = get_block_rsv(trans, root);
7485
7486         if (unlikely(block_rsv->size == 0))
7487                 goto try_reserve;
7488 again:
7489         ret = block_rsv_use_bytes(block_rsv, blocksize);
7490         if (!ret)
7491                 return block_rsv;
7492
7493         if (block_rsv->failfast)
7494                 return ERR_PTR(ret);
7495
7496         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7497                 global_updated = true;
7498                 update_global_block_rsv(root->fs_info);
7499                 goto again;
7500         }
7501
7502         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7503                 static DEFINE_RATELIMIT_STATE(_rs,
7504                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7505                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7506                 if (__ratelimit(&_rs))
7507                         WARN(1, KERN_DEBUG
7508                                 "BTRFS: block rsv returned %d\n", ret);
7509         }
7510 try_reserve:
7511         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7512                                      BTRFS_RESERVE_NO_FLUSH);
7513         if (!ret)
7514                 return block_rsv;
7515         /*
7516          * If we couldn't reserve metadata bytes try and use some from
7517          * the global reserve if its space type is the same as the global
7518          * reservation.
7519          */
7520         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7521             block_rsv->space_info == global_rsv->space_info) {
7522                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7523                 if (!ret)
7524                         return global_rsv;
7525         }
7526         return ERR_PTR(ret);
7527 }
7528
7529 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7530                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7531 {
7532         block_rsv_add_bytes(block_rsv, blocksize, 0);
7533         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7534 }
7535
7536 /*
7537  * finds a free extent and does all the dirty work required for allocation
7538  * returns the key for the extent through ins, and a tree buffer for
7539  * the first block of the extent through buf.
7540  *
7541  * returns the tree buffer or NULL.
7542  */
7543 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7544                                         struct btrfs_root *root,
7545                                         u64 parent, u64 root_objectid,
7546                                         struct btrfs_disk_key *key, int level,
7547                                         u64 hint, u64 empty_size)
7548 {
7549         struct btrfs_key ins;
7550         struct btrfs_block_rsv *block_rsv;
7551         struct extent_buffer *buf;
7552         u64 flags = 0;
7553         int ret;
7554         u32 blocksize = root->nodesize;
7555         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7556                                                  SKINNY_METADATA);
7557
7558         if (btrfs_test_is_dummy_root(root)) {
7559                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7560                                             level);
7561                 if (!IS_ERR(buf))
7562                         root->alloc_bytenr += blocksize;
7563                 return buf;
7564         }
7565
7566         block_rsv = use_block_rsv(trans, root, blocksize);
7567         if (IS_ERR(block_rsv))
7568                 return ERR_CAST(block_rsv);
7569
7570         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7571                                    empty_size, hint, &ins, 0, 0);
7572         if (ret) {
7573                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7574                 return ERR_PTR(ret);
7575         }
7576
7577         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
7578         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
7579
7580         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7581                 if (parent == 0)
7582                         parent = ins.objectid;
7583                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7584         } else
7585                 BUG_ON(parent > 0);
7586
7587         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7588                 struct btrfs_delayed_extent_op *extent_op;
7589                 extent_op = btrfs_alloc_delayed_extent_op();
7590                 BUG_ON(!extent_op); /* -ENOMEM */
7591                 if (key)
7592                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7593                 else
7594                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7595                 extent_op->flags_to_set = flags;
7596                 if (skinny_metadata)
7597                         extent_op->update_key = 0;
7598                 else
7599                         extent_op->update_key = 1;
7600                 extent_op->update_flags = 1;
7601                 extent_op->is_data = 0;
7602                 extent_op->level = level;
7603
7604                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7605                                         ins.objectid,
7606                                         ins.offset, parent, root_objectid,
7607                                         level, BTRFS_ADD_DELAYED_EXTENT,
7608                                         extent_op, 0);
7609                 BUG_ON(ret); /* -ENOMEM */
7610         }
7611         return buf;
7612 }
7613
7614 struct walk_control {
7615         u64 refs[BTRFS_MAX_LEVEL];
7616         u64 flags[BTRFS_MAX_LEVEL];
7617         struct btrfs_key update_progress;
7618         int stage;
7619         int level;
7620         int shared_level;
7621         int update_ref;
7622         int keep_locks;
7623         int reada_slot;
7624         int reada_count;
7625         int for_reloc;
7626 };
7627
7628 #define DROP_REFERENCE  1
7629 #define UPDATE_BACKREF  2
7630
7631 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7632                                      struct btrfs_root *root,
7633                                      struct walk_control *wc,
7634                                      struct btrfs_path *path)
7635 {
7636         u64 bytenr;
7637         u64 generation;
7638         u64 refs;
7639         u64 flags;
7640         u32 nritems;
7641         u32 blocksize;
7642         struct btrfs_key key;
7643         struct extent_buffer *eb;
7644         int ret;
7645         int slot;
7646         int nread = 0;
7647
7648         if (path->slots[wc->level] < wc->reada_slot) {
7649                 wc->reada_count = wc->reada_count * 2 / 3;
7650                 wc->reada_count = max(wc->reada_count, 2);
7651         } else {
7652                 wc->reada_count = wc->reada_count * 3 / 2;
7653                 wc->reada_count = min_t(int, wc->reada_count,
7654                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7655         }
7656
7657         eb = path->nodes[wc->level];
7658         nritems = btrfs_header_nritems(eb);
7659         blocksize = root->nodesize;
7660
7661         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7662                 if (nread >= wc->reada_count)
7663                         break;
7664
7665                 cond_resched();
7666                 bytenr = btrfs_node_blockptr(eb, slot);
7667                 generation = btrfs_node_ptr_generation(eb, slot);
7668
7669                 if (slot == path->slots[wc->level])
7670                         goto reada;
7671
7672                 if (wc->stage == UPDATE_BACKREF &&
7673                     generation <= root->root_key.offset)
7674                         continue;
7675
7676                 /* We don't lock the tree block, it's OK to be racy here */
7677                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7678                                                wc->level - 1, 1, &refs,
7679                                                &flags);
7680                 /* We don't care about errors in readahead. */
7681                 if (ret < 0)
7682                         continue;
7683                 BUG_ON(refs == 0);
7684
7685                 if (wc->stage == DROP_REFERENCE) {
7686                         if (refs == 1)
7687                                 goto reada;
7688
7689                         if (wc->level == 1 &&
7690                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7691                                 continue;
7692                         if (!wc->update_ref ||
7693                             generation <= root->root_key.offset)
7694                                 continue;
7695                         btrfs_node_key_to_cpu(eb, &key, slot);
7696                         ret = btrfs_comp_cpu_keys(&key,
7697                                                   &wc->update_progress);
7698                         if (ret < 0)
7699                                 continue;
7700                 } else {
7701                         if (wc->level == 1 &&
7702                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7703                                 continue;
7704                 }
7705 reada:
7706                 readahead_tree_block(root, bytenr);
7707                 nread++;
7708         }
7709         wc->reada_slot = slot;
7710 }
7711
7712 static int account_leaf_items(struct btrfs_trans_handle *trans,
7713                               struct btrfs_root *root,
7714                               struct extent_buffer *eb)
7715 {
7716         int nr = btrfs_header_nritems(eb);
7717         int i, extent_type, ret;
7718         struct btrfs_key key;
7719         struct btrfs_file_extent_item *fi;
7720         u64 bytenr, num_bytes;
7721
7722         for (i = 0; i < nr; i++) {
7723                 btrfs_item_key_to_cpu(eb, &key, i);
7724
7725                 if (key.type != BTRFS_EXTENT_DATA_KEY)
7726                         continue;
7727
7728                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
7729                 /* filter out non qgroup-accountable extents  */
7730                 extent_type = btrfs_file_extent_type(eb, fi);
7731
7732                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
7733                         continue;
7734
7735                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
7736                 if (!bytenr)
7737                         continue;
7738
7739                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
7740
7741                 ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7742                                               root->objectid,
7743                                               bytenr, num_bytes,
7744                                               BTRFS_QGROUP_OPER_SUB_SUBTREE, 0);
7745                 if (ret)
7746                         return ret;
7747         }
7748         return 0;
7749 }
7750
7751 /*
7752  * Walk up the tree from the bottom, freeing leaves and any interior
7753  * nodes which have had all slots visited. If a node (leaf or
7754  * interior) is freed, the node above it will have it's slot
7755  * incremented. The root node will never be freed.
7756  *
7757  * At the end of this function, we should have a path which has all
7758  * slots incremented to the next position for a search. If we need to
7759  * read a new node it will be NULL and the node above it will have the
7760  * correct slot selected for a later read.
7761  *
7762  * If we increment the root nodes slot counter past the number of
7763  * elements, 1 is returned to signal completion of the search.
7764  */
7765 static int adjust_slots_upwards(struct btrfs_root *root,
7766                                 struct btrfs_path *path, int root_level)
7767 {
7768         int level = 0;
7769         int nr, slot;
7770         struct extent_buffer *eb;
7771
7772         if (root_level == 0)
7773                 return 1;
7774
7775         while (level <= root_level) {
7776                 eb = path->nodes[level];
7777                 nr = btrfs_header_nritems(eb);
7778                 path->slots[level]++;
7779                 slot = path->slots[level];
7780                 if (slot >= nr || level == 0) {
7781                         /*
7782                          * Don't free the root -  we will detect this
7783                          * condition after our loop and return a
7784                          * positive value for caller to stop walking the tree.
7785                          */
7786                         if (level != root_level) {
7787                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7788                                 path->locks[level] = 0;
7789
7790                                 free_extent_buffer(eb);
7791                                 path->nodes[level] = NULL;
7792                                 path->slots[level] = 0;
7793                         }
7794                 } else {
7795                         /*
7796                          * We have a valid slot to walk back down
7797                          * from. Stop here so caller can process these
7798                          * new nodes.
7799                          */
7800                         break;
7801                 }
7802
7803                 level++;
7804         }
7805
7806         eb = path->nodes[root_level];
7807         if (path->slots[root_level] >= btrfs_header_nritems(eb))
7808                 return 1;
7809
7810         return 0;
7811 }
7812
7813 /*
7814  * root_eb is the subtree root and is locked before this function is called.
7815  */
7816 static int account_shared_subtree(struct btrfs_trans_handle *trans,
7817                                   struct btrfs_root *root,
7818                                   struct extent_buffer *root_eb,
7819                                   u64 root_gen,
7820                                   int root_level)
7821 {
7822         int ret = 0;
7823         int level;
7824         struct extent_buffer *eb = root_eb;
7825         struct btrfs_path *path = NULL;
7826
7827         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
7828         BUG_ON(root_eb == NULL);
7829
7830         if (!root->fs_info->quota_enabled)
7831                 return 0;
7832
7833         if (!extent_buffer_uptodate(root_eb)) {
7834                 ret = btrfs_read_buffer(root_eb, root_gen);
7835                 if (ret)
7836                         goto out;
7837         }
7838
7839         if (root_level == 0) {
7840                 ret = account_leaf_items(trans, root, root_eb);
7841                 goto out;
7842         }
7843
7844         path = btrfs_alloc_path();
7845         if (!path)
7846                 return -ENOMEM;
7847
7848         /*
7849          * Walk down the tree.  Missing extent blocks are filled in as
7850          * we go. Metadata is accounted every time we read a new
7851          * extent block.
7852          *
7853          * When we reach a leaf, we account for file extent items in it,
7854          * walk back up the tree (adjusting slot pointers as we go)
7855          * and restart the search process.
7856          */
7857         extent_buffer_get(root_eb); /* For path */
7858         path->nodes[root_level] = root_eb;
7859         path->slots[root_level] = 0;
7860         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
7861 walk_down:
7862         level = root_level;
7863         while (level >= 0) {
7864                 if (path->nodes[level] == NULL) {
7865                         int parent_slot;
7866                         u64 child_gen;
7867                         u64 child_bytenr;
7868
7869                         /* We need to get child blockptr/gen from
7870                          * parent before we can read it. */
7871                         eb = path->nodes[level + 1];
7872                         parent_slot = path->slots[level + 1];
7873                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
7874                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
7875
7876                         eb = read_tree_block(root, child_bytenr, child_gen);
7877                         if (!eb || !extent_buffer_uptodate(eb)) {
7878                                 ret = -EIO;
7879                                 goto out;
7880                         }
7881
7882                         path->nodes[level] = eb;
7883                         path->slots[level] = 0;
7884
7885                         btrfs_tree_read_lock(eb);
7886                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
7887                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
7888
7889                         ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7890                                                 root->objectid,
7891                                                 child_bytenr,
7892                                                 root->nodesize,
7893                                                 BTRFS_QGROUP_OPER_SUB_SUBTREE,
7894                                                 0);
7895                         if (ret)
7896                                 goto out;
7897
7898                 }
7899
7900                 if (level == 0) {
7901                         ret = account_leaf_items(trans, root, path->nodes[level]);
7902                         if (ret)
7903                                 goto out;
7904
7905                         /* Nonzero return here means we completed our search */
7906                         ret = adjust_slots_upwards(root, path, root_level);
7907                         if (ret)
7908                                 break;
7909
7910                         /* Restart search with new slots */
7911                         goto walk_down;
7912                 }
7913
7914                 level--;
7915         }
7916
7917         ret = 0;
7918 out:
7919         btrfs_free_path(path);
7920
7921         return ret;
7922 }
7923
7924 /*
7925  * helper to process tree block while walking down the tree.
7926  *
7927  * when wc->stage == UPDATE_BACKREF, this function updates
7928  * back refs for pointers in the block.
7929  *
7930  * NOTE: return value 1 means we should stop walking down.
7931  */
7932 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7933                                    struct btrfs_root *root,
7934                                    struct btrfs_path *path,
7935                                    struct walk_control *wc, int lookup_info)
7936 {
7937         int level = wc->level;
7938         struct extent_buffer *eb = path->nodes[level];
7939         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7940         int ret;
7941
7942         if (wc->stage == UPDATE_BACKREF &&
7943             btrfs_header_owner(eb) != root->root_key.objectid)
7944                 return 1;
7945
7946         /*
7947          * when reference count of tree block is 1, it won't increase
7948          * again. once full backref flag is set, we never clear it.
7949          */
7950         if (lookup_info &&
7951             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7952              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7953                 BUG_ON(!path->locks[level]);
7954                 ret = btrfs_lookup_extent_info(trans, root,
7955                                                eb->start, level, 1,
7956                                                &wc->refs[level],
7957                                                &wc->flags[level]);
7958                 BUG_ON(ret == -ENOMEM);
7959                 if (ret)
7960                         return ret;
7961                 BUG_ON(wc->refs[level] == 0);
7962         }
7963
7964         if (wc->stage == DROP_REFERENCE) {
7965                 if (wc->refs[level] > 1)
7966                         return 1;
7967
7968                 if (path->locks[level] && !wc->keep_locks) {
7969                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7970                         path->locks[level] = 0;
7971                 }
7972                 return 0;
7973         }
7974
7975         /* wc->stage == UPDATE_BACKREF */
7976         if (!(wc->flags[level] & flag)) {
7977                 BUG_ON(!path->locks[level]);
7978                 ret = btrfs_inc_ref(trans, root, eb, 1);
7979                 BUG_ON(ret); /* -ENOMEM */
7980                 ret = btrfs_dec_ref(trans, root, eb, 0);
7981                 BUG_ON(ret); /* -ENOMEM */
7982                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7983                                                   eb->len, flag,
7984                                                   btrfs_header_level(eb), 0);
7985                 BUG_ON(ret); /* -ENOMEM */
7986                 wc->flags[level] |= flag;
7987         }
7988
7989         /*
7990          * the block is shared by multiple trees, so it's not good to
7991          * keep the tree lock
7992          */
7993         if (path->locks[level] && level > 0) {
7994                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7995                 path->locks[level] = 0;
7996         }
7997         return 0;
7998 }
7999
8000 /*
8001  * helper to process tree block pointer.
8002  *
8003  * when wc->stage == DROP_REFERENCE, this function checks
8004  * reference count of the block pointed to. if the block
8005  * is shared and we need update back refs for the subtree
8006  * rooted at the block, this function changes wc->stage to
8007  * UPDATE_BACKREF. if the block is shared and there is no
8008  * need to update back, this function drops the reference
8009  * to the block.
8010  *
8011  * NOTE: return value 1 means we should stop walking down.
8012  */
8013 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8014                                  struct btrfs_root *root,
8015                                  struct btrfs_path *path,
8016                                  struct walk_control *wc, int *lookup_info)
8017 {
8018         u64 bytenr;
8019         u64 generation;
8020         u64 parent;
8021         u32 blocksize;
8022         struct btrfs_key key;
8023         struct extent_buffer *next;
8024         int level = wc->level;
8025         int reada = 0;
8026         int ret = 0;
8027         bool need_account = false;
8028
8029         generation = btrfs_node_ptr_generation(path->nodes[level],
8030                                                path->slots[level]);
8031         /*
8032          * if the lower level block was created before the snapshot
8033          * was created, we know there is no need to update back refs
8034          * for the subtree
8035          */
8036         if (wc->stage == UPDATE_BACKREF &&
8037             generation <= root->root_key.offset) {
8038                 *lookup_info = 1;
8039                 return 1;
8040         }
8041
8042         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8043         blocksize = root->nodesize;
8044
8045         next = btrfs_find_tree_block(root->fs_info, bytenr);
8046         if (!next) {
8047                 next = btrfs_find_create_tree_block(root, bytenr);
8048                 if (!next)
8049                         return -ENOMEM;
8050                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8051                                                level - 1);
8052                 reada = 1;
8053         }
8054         btrfs_tree_lock(next);
8055         btrfs_set_lock_blocking(next);
8056
8057         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8058                                        &wc->refs[level - 1],
8059                                        &wc->flags[level - 1]);
8060         if (ret < 0) {
8061                 btrfs_tree_unlock(next);
8062                 return ret;
8063         }
8064
8065         if (unlikely(wc->refs[level - 1] == 0)) {
8066                 btrfs_err(root->fs_info, "Missing references.");
8067                 BUG();
8068         }
8069         *lookup_info = 0;
8070
8071         if (wc->stage == DROP_REFERENCE) {
8072                 if (wc->refs[level - 1] > 1) {
8073                         need_account = true;
8074                         if (level == 1 &&
8075                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8076                                 goto skip;
8077
8078                         if (!wc->update_ref ||
8079                             generation <= root->root_key.offset)
8080                                 goto skip;
8081
8082                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8083                                               path->slots[level]);
8084                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8085                         if (ret < 0)
8086                                 goto skip;
8087
8088                         wc->stage = UPDATE_BACKREF;
8089                         wc->shared_level = level - 1;
8090                 }
8091         } else {
8092                 if (level == 1 &&
8093                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8094                         goto skip;
8095         }
8096
8097         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8098                 btrfs_tree_unlock(next);
8099                 free_extent_buffer(next);
8100                 next = NULL;
8101                 *lookup_info = 1;
8102         }
8103
8104         if (!next) {
8105                 if (reada && level == 1)
8106                         reada_walk_down(trans, root, wc, path);
8107                 next = read_tree_block(root, bytenr, generation);
8108                 if (!next || !extent_buffer_uptodate(next)) {
8109                         free_extent_buffer(next);
8110                         return -EIO;
8111                 }
8112                 btrfs_tree_lock(next);
8113                 btrfs_set_lock_blocking(next);
8114         }
8115
8116         level--;
8117         BUG_ON(level != btrfs_header_level(next));
8118         path->nodes[level] = next;
8119         path->slots[level] = 0;
8120         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8121         wc->level = level;
8122         if (wc->level == 1)
8123                 wc->reada_slot = 0;
8124         return 0;
8125 skip:
8126         wc->refs[level - 1] = 0;
8127         wc->flags[level - 1] = 0;
8128         if (wc->stage == DROP_REFERENCE) {
8129                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8130                         parent = path->nodes[level]->start;
8131                 } else {
8132                         BUG_ON(root->root_key.objectid !=
8133                                btrfs_header_owner(path->nodes[level]));
8134                         parent = 0;
8135                 }
8136
8137                 if (need_account) {
8138                         ret = account_shared_subtree(trans, root, next,
8139                                                      generation, level - 1);
8140                         if (ret) {
8141                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8142                                         "%d accounting shared subtree. Quota "
8143                                         "is out of sync, rescan required.\n",
8144                                         root->fs_info->sb->s_id, ret);
8145                         }
8146                 }
8147                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8148                                 root->root_key.objectid, level - 1, 0, 0);
8149                 BUG_ON(ret); /* -ENOMEM */
8150         }
8151         btrfs_tree_unlock(next);
8152         free_extent_buffer(next);
8153         *lookup_info = 1;
8154         return 1;
8155 }
8156
8157 /*
8158  * helper to process tree block while walking up the tree.
8159  *
8160  * when wc->stage == DROP_REFERENCE, this function drops
8161  * reference count on the block.
8162  *
8163  * when wc->stage == UPDATE_BACKREF, this function changes
8164  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8165  * to UPDATE_BACKREF previously while processing the block.
8166  *
8167  * NOTE: return value 1 means we should stop walking up.
8168  */
8169 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8170                                  struct btrfs_root *root,
8171                                  struct btrfs_path *path,
8172                                  struct walk_control *wc)
8173 {
8174         int ret;
8175         int level = wc->level;
8176         struct extent_buffer *eb = path->nodes[level];
8177         u64 parent = 0;
8178
8179         if (wc->stage == UPDATE_BACKREF) {
8180                 BUG_ON(wc->shared_level < level);
8181                 if (level < wc->shared_level)
8182                         goto out;
8183
8184                 ret = find_next_key(path, level + 1, &wc->update_progress);
8185                 if (ret > 0)
8186                         wc->update_ref = 0;
8187
8188                 wc->stage = DROP_REFERENCE;
8189                 wc->shared_level = -1;
8190                 path->slots[level] = 0;
8191
8192                 /*
8193                  * check reference count again if the block isn't locked.
8194                  * we should start walking down the tree again if reference
8195                  * count is one.
8196                  */
8197                 if (!path->locks[level]) {
8198                         BUG_ON(level == 0);
8199                         btrfs_tree_lock(eb);
8200                         btrfs_set_lock_blocking(eb);
8201                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8202
8203                         ret = btrfs_lookup_extent_info(trans, root,
8204                                                        eb->start, level, 1,
8205                                                        &wc->refs[level],
8206                                                        &wc->flags[level]);
8207                         if (ret < 0) {
8208                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8209                                 path->locks[level] = 0;
8210                                 return ret;
8211                         }
8212                         BUG_ON(wc->refs[level] == 0);
8213                         if (wc->refs[level] == 1) {
8214                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8215                                 path->locks[level] = 0;
8216                                 return 1;
8217                         }
8218                 }
8219         }
8220
8221         /* wc->stage == DROP_REFERENCE */
8222         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8223
8224         if (wc->refs[level] == 1) {
8225                 if (level == 0) {
8226                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8227                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8228                         else
8229                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8230                         BUG_ON(ret); /* -ENOMEM */
8231                         ret = account_leaf_items(trans, root, eb);
8232                         if (ret) {
8233                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8234                                         "%d accounting leaf items. Quota "
8235                                         "is out of sync, rescan required.\n",
8236                                         root->fs_info->sb->s_id, ret);
8237                         }
8238                 }
8239                 /* make block locked assertion in clean_tree_block happy */
8240                 if (!path->locks[level] &&
8241                     btrfs_header_generation(eb) == trans->transid) {
8242                         btrfs_tree_lock(eb);
8243                         btrfs_set_lock_blocking(eb);
8244                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8245                 }
8246                 clean_tree_block(trans, root->fs_info, eb);
8247         }
8248
8249         if (eb == root->node) {
8250                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8251                         parent = eb->start;
8252                 else
8253                         BUG_ON(root->root_key.objectid !=
8254                                btrfs_header_owner(eb));
8255         } else {
8256                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8257                         parent = path->nodes[level + 1]->start;
8258                 else
8259                         BUG_ON(root->root_key.objectid !=
8260                                btrfs_header_owner(path->nodes[level + 1]));
8261         }
8262
8263         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8264 out:
8265         wc->refs[level] = 0;
8266         wc->flags[level] = 0;
8267         return 0;
8268 }
8269
8270 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8271                                    struct btrfs_root *root,
8272                                    struct btrfs_path *path,
8273                                    struct walk_control *wc)
8274 {
8275         int level = wc->level;
8276         int lookup_info = 1;
8277         int ret;
8278
8279         while (level >= 0) {
8280                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8281                 if (ret > 0)
8282                         break;
8283
8284                 if (level == 0)
8285                         break;
8286
8287                 if (path->slots[level] >=
8288                     btrfs_header_nritems(path->nodes[level]))
8289                         break;
8290
8291                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8292                 if (ret > 0) {
8293                         path->slots[level]++;
8294                         continue;
8295                 } else if (ret < 0)
8296                         return ret;
8297                 level = wc->level;
8298         }
8299         return 0;
8300 }
8301
8302 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8303                                  struct btrfs_root *root,
8304                                  struct btrfs_path *path,
8305                                  struct walk_control *wc, int max_level)
8306 {
8307         int level = wc->level;
8308         int ret;
8309
8310         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8311         while (level < max_level && path->nodes[level]) {
8312                 wc->level = level;
8313                 if (path->slots[level] + 1 <
8314                     btrfs_header_nritems(path->nodes[level])) {
8315                         path->slots[level]++;
8316                         return 0;
8317                 } else {
8318                         ret = walk_up_proc(trans, root, path, wc);
8319                         if (ret > 0)
8320                                 return 0;
8321
8322                         if (path->locks[level]) {
8323                                 btrfs_tree_unlock_rw(path->nodes[level],
8324                                                      path->locks[level]);
8325                                 path->locks[level] = 0;
8326                         }
8327                         free_extent_buffer(path->nodes[level]);
8328                         path->nodes[level] = NULL;
8329                         level++;
8330                 }
8331         }
8332         return 1;
8333 }
8334
8335 /*
8336  * drop a subvolume tree.
8337  *
8338  * this function traverses the tree freeing any blocks that only
8339  * referenced by the tree.
8340  *
8341  * when a shared tree block is found. this function decreases its
8342  * reference count by one. if update_ref is true, this function
8343  * also make sure backrefs for the shared block and all lower level
8344  * blocks are properly updated.
8345  *
8346  * If called with for_reloc == 0, may exit early with -EAGAIN
8347  */
8348 int btrfs_drop_snapshot(struct btrfs_root *root,
8349                          struct btrfs_block_rsv *block_rsv, int update_ref,
8350                          int for_reloc)
8351 {
8352         struct btrfs_path *path;
8353         struct btrfs_trans_handle *trans;
8354         struct btrfs_root *tree_root = root->fs_info->tree_root;
8355         struct btrfs_root_item *root_item = &root->root_item;
8356         struct walk_control *wc;
8357         struct btrfs_key key;
8358         int err = 0;
8359         int ret;
8360         int level;
8361         bool root_dropped = false;
8362
8363         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8364
8365         path = btrfs_alloc_path();
8366         if (!path) {
8367                 err = -ENOMEM;
8368                 goto out;
8369         }
8370
8371         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8372         if (!wc) {
8373                 btrfs_free_path(path);
8374                 err = -ENOMEM;
8375                 goto out;
8376         }
8377
8378         trans = btrfs_start_transaction(tree_root, 0);
8379         if (IS_ERR(trans)) {
8380                 err = PTR_ERR(trans);
8381                 goto out_free;
8382         }
8383
8384         if (block_rsv)
8385                 trans->block_rsv = block_rsv;
8386
8387         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8388                 level = btrfs_header_level(root->node);
8389                 path->nodes[level] = btrfs_lock_root_node(root);
8390                 btrfs_set_lock_blocking(path->nodes[level]);
8391                 path->slots[level] = 0;
8392                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8393                 memset(&wc->update_progress, 0,
8394                        sizeof(wc->update_progress));
8395         } else {
8396                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8397                 memcpy(&wc->update_progress, &key,
8398                        sizeof(wc->update_progress));
8399
8400                 level = root_item->drop_level;
8401                 BUG_ON(level == 0);
8402                 path->lowest_level = level;
8403                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8404                 path->lowest_level = 0;
8405                 if (ret < 0) {
8406                         err = ret;
8407                         goto out_end_trans;
8408                 }
8409                 WARN_ON(ret > 0);
8410
8411                 /*
8412                  * unlock our path, this is safe because only this
8413                  * function is allowed to delete this snapshot
8414                  */
8415                 btrfs_unlock_up_safe(path, 0);
8416
8417                 level = btrfs_header_level(root->node);
8418                 while (1) {
8419                         btrfs_tree_lock(path->nodes[level]);
8420                         btrfs_set_lock_blocking(path->nodes[level]);
8421                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8422
8423                         ret = btrfs_lookup_extent_info(trans, root,
8424                                                 path->nodes[level]->start,
8425                                                 level, 1, &wc->refs[level],
8426                                                 &wc->flags[level]);
8427                         if (ret < 0) {
8428                                 err = ret;
8429                                 goto out_end_trans;
8430                         }
8431                         BUG_ON(wc->refs[level] == 0);
8432
8433                         if (level == root_item->drop_level)
8434                                 break;
8435
8436                         btrfs_tree_unlock(path->nodes[level]);
8437                         path->locks[level] = 0;
8438                         WARN_ON(wc->refs[level] != 1);
8439                         level--;
8440                 }
8441         }
8442
8443         wc->level = level;
8444         wc->shared_level = -1;
8445         wc->stage = DROP_REFERENCE;
8446         wc->update_ref = update_ref;
8447         wc->keep_locks = 0;
8448         wc->for_reloc = for_reloc;
8449         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8450
8451         while (1) {
8452
8453                 ret = walk_down_tree(trans, root, path, wc);
8454                 if (ret < 0) {
8455                         err = ret;
8456                         break;
8457                 }
8458
8459                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8460                 if (ret < 0) {
8461                         err = ret;
8462                         break;
8463                 }
8464
8465                 if (ret > 0) {
8466                         BUG_ON(wc->stage != DROP_REFERENCE);
8467                         break;
8468                 }
8469
8470                 if (wc->stage == DROP_REFERENCE) {
8471                         level = wc->level;
8472                         btrfs_node_key(path->nodes[level],
8473                                        &root_item->drop_progress,
8474                                        path->slots[level]);
8475                         root_item->drop_level = level;
8476                 }
8477
8478                 BUG_ON(wc->level == 0);
8479                 if (btrfs_should_end_transaction(trans, tree_root) ||
8480                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8481                         ret = btrfs_update_root(trans, tree_root,
8482                                                 &root->root_key,
8483                                                 root_item);
8484                         if (ret) {
8485                                 btrfs_abort_transaction(trans, tree_root, ret);
8486                                 err = ret;
8487                                 goto out_end_trans;
8488                         }
8489
8490                         /*
8491                          * Qgroup update accounting is run from
8492                          * delayed ref handling. This usually works
8493                          * out because delayed refs are normally the
8494                          * only way qgroup updates are added. However,
8495                          * we may have added updates during our tree
8496                          * walk so run qgroups here to make sure we
8497                          * don't lose any updates.
8498                          */
8499                         ret = btrfs_delayed_qgroup_accounting(trans,
8500                                                               root->fs_info);
8501                         if (ret)
8502                                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8503                                                    "running qgroup updates "
8504                                                    "during snapshot delete. "
8505                                                    "Quota is out of sync, "
8506                                                    "rescan required.\n", ret);
8507
8508                         btrfs_end_transaction_throttle(trans, tree_root);
8509                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8510                                 pr_debug("BTRFS: drop snapshot early exit\n");
8511                                 err = -EAGAIN;
8512                                 goto out_free;
8513                         }
8514
8515                         trans = btrfs_start_transaction(tree_root, 0);
8516                         if (IS_ERR(trans)) {
8517                                 err = PTR_ERR(trans);
8518                                 goto out_free;
8519                         }
8520                         if (block_rsv)
8521                                 trans->block_rsv = block_rsv;
8522                 }
8523         }
8524         btrfs_release_path(path);
8525         if (err)
8526                 goto out_end_trans;
8527
8528         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8529         if (ret) {
8530                 btrfs_abort_transaction(trans, tree_root, ret);
8531                 goto out_end_trans;
8532         }
8533
8534         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8535                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8536                                       NULL, NULL);
8537                 if (ret < 0) {
8538                         btrfs_abort_transaction(trans, tree_root, ret);
8539                         err = ret;
8540                         goto out_end_trans;
8541                 } else if (ret > 0) {
8542                         /* if we fail to delete the orphan item this time
8543                          * around, it'll get picked up the next time.
8544                          *
8545                          * The most common failure here is just -ENOENT.
8546                          */
8547                         btrfs_del_orphan_item(trans, tree_root,
8548                                               root->root_key.objectid);
8549                 }
8550         }
8551
8552         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8553                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
8554         } else {
8555                 free_extent_buffer(root->node);
8556                 free_extent_buffer(root->commit_root);
8557                 btrfs_put_fs_root(root);
8558         }
8559         root_dropped = true;
8560 out_end_trans:
8561         ret = btrfs_delayed_qgroup_accounting(trans, tree_root->fs_info);
8562         if (ret)
8563                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8564                                    "running qgroup updates "
8565                                    "during snapshot delete. "
8566                                    "Quota is out of sync, "
8567                                    "rescan required.\n", ret);
8568
8569         btrfs_end_transaction_throttle(trans, tree_root);
8570 out_free:
8571         kfree(wc);
8572         btrfs_free_path(path);
8573 out:
8574         /*
8575          * So if we need to stop dropping the snapshot for whatever reason we
8576          * need to make sure to add it back to the dead root list so that we
8577          * keep trying to do the work later.  This also cleans up roots if we
8578          * don't have it in the radix (like when we recover after a power fail
8579          * or unmount) so we don't leak memory.
8580          */
8581         if (!for_reloc && root_dropped == false)
8582                 btrfs_add_dead_root(root);
8583         if (err && err != -EAGAIN)
8584                 btrfs_std_error(root->fs_info, err);
8585         return err;
8586 }
8587
8588 /*
8589  * drop subtree rooted at tree block 'node'.
8590  *
8591  * NOTE: this function will unlock and release tree block 'node'
8592  * only used by relocation code
8593  */
8594 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8595                         struct btrfs_root *root,
8596                         struct extent_buffer *node,
8597                         struct extent_buffer *parent)
8598 {
8599         struct btrfs_path *path;
8600         struct walk_control *wc;
8601         int level;
8602         int parent_level;
8603         int ret = 0;
8604         int wret;
8605
8606         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8607
8608         path = btrfs_alloc_path();
8609         if (!path)
8610                 return -ENOMEM;
8611
8612         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8613         if (!wc) {
8614                 btrfs_free_path(path);
8615                 return -ENOMEM;
8616         }
8617
8618         btrfs_assert_tree_locked(parent);
8619         parent_level = btrfs_header_level(parent);
8620         extent_buffer_get(parent);
8621         path->nodes[parent_level] = parent;
8622         path->slots[parent_level] = btrfs_header_nritems(parent);
8623
8624         btrfs_assert_tree_locked(node);
8625         level = btrfs_header_level(node);
8626         path->nodes[level] = node;
8627         path->slots[level] = 0;
8628         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8629
8630         wc->refs[parent_level] = 1;
8631         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8632         wc->level = level;
8633         wc->shared_level = -1;
8634         wc->stage = DROP_REFERENCE;
8635         wc->update_ref = 0;
8636         wc->keep_locks = 1;
8637         wc->for_reloc = 1;
8638         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8639
8640         while (1) {
8641                 wret = walk_down_tree(trans, root, path, wc);
8642                 if (wret < 0) {
8643                         ret = wret;
8644                         break;
8645                 }
8646
8647                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8648                 if (wret < 0)
8649                         ret = wret;
8650                 if (wret != 0)
8651                         break;
8652         }
8653
8654         kfree(wc);
8655         btrfs_free_path(path);
8656         return ret;
8657 }
8658
8659 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8660 {
8661         u64 num_devices;
8662         u64 stripped;
8663
8664         /*
8665          * if restripe for this chunk_type is on pick target profile and
8666          * return, otherwise do the usual balance
8667          */
8668         stripped = get_restripe_target(root->fs_info, flags);
8669         if (stripped)
8670                 return extended_to_chunk(stripped);
8671
8672         num_devices = root->fs_info->fs_devices->rw_devices;
8673
8674         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8675                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8676                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8677
8678         if (num_devices == 1) {
8679                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8680                 stripped = flags & ~stripped;
8681
8682                 /* turn raid0 into single device chunks */
8683                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8684                         return stripped;
8685
8686                 /* turn mirroring into duplication */
8687                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8688                              BTRFS_BLOCK_GROUP_RAID10))
8689                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8690         } else {
8691                 /* they already had raid on here, just return */
8692                 if (flags & stripped)
8693                         return flags;
8694
8695                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8696                 stripped = flags & ~stripped;
8697
8698                 /* switch duplicated blocks with raid1 */
8699                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8700                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8701
8702                 /* this is drive concat, leave it alone */
8703         }
8704
8705         return flags;
8706 }
8707
8708 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8709 {
8710         struct btrfs_space_info *sinfo = cache->space_info;
8711         u64 num_bytes;
8712         u64 min_allocable_bytes;
8713         int ret = -ENOSPC;
8714
8715
8716         /*
8717          * We need some metadata space and system metadata space for
8718          * allocating chunks in some corner cases until we force to set
8719          * it to be readonly.
8720          */
8721         if ((sinfo->flags &
8722              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8723             !force)
8724                 min_allocable_bytes = 1 * 1024 * 1024;
8725         else
8726                 min_allocable_bytes = 0;
8727
8728         spin_lock(&sinfo->lock);
8729         spin_lock(&cache->lock);
8730
8731         if (cache->ro) {
8732                 ret = 0;
8733                 goto out;
8734         }
8735
8736         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8737                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8738
8739         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8740             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8741             min_allocable_bytes <= sinfo->total_bytes) {
8742                 sinfo->bytes_readonly += num_bytes;
8743                 cache->ro = 1;
8744                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
8745                 ret = 0;
8746         }
8747 out:
8748         spin_unlock(&cache->lock);
8749         spin_unlock(&sinfo->lock);
8750         return ret;
8751 }
8752
8753 int btrfs_set_block_group_ro(struct btrfs_root *root,
8754                              struct btrfs_block_group_cache *cache)
8755
8756 {
8757         struct btrfs_trans_handle *trans;
8758         u64 alloc_flags;
8759         int ret;
8760
8761         BUG_ON(cache->ro);
8762
8763 again:
8764         trans = btrfs_join_transaction(root);
8765         if (IS_ERR(trans))
8766                 return PTR_ERR(trans);
8767
8768         /*
8769          * we're not allowed to set block groups readonly after the dirty
8770          * block groups cache has started writing.  If it already started,
8771          * back off and let this transaction commit
8772          */
8773         mutex_lock(&root->fs_info->ro_block_group_mutex);
8774         if (trans->transaction->dirty_bg_run) {
8775                 u64 transid = trans->transid;
8776
8777                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
8778                 btrfs_end_transaction(trans, root);
8779
8780                 ret = btrfs_wait_for_commit(root, transid);
8781                 if (ret)
8782                         return ret;
8783                 goto again;
8784         }
8785
8786
8787         ret = set_block_group_ro(cache, 0);
8788         if (!ret)
8789                 goto out;
8790         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8791         ret = do_chunk_alloc(trans, root, alloc_flags,
8792                              CHUNK_ALLOC_FORCE);
8793         if (ret < 0)
8794                 goto out;
8795         ret = set_block_group_ro(cache, 0);
8796 out:
8797         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
8798                 alloc_flags = update_block_group_flags(root, cache->flags);
8799                 check_system_chunk(trans, root, alloc_flags);
8800         }
8801         mutex_unlock(&root->fs_info->ro_block_group_mutex);
8802
8803         btrfs_end_transaction(trans, root);
8804         return ret;
8805 }
8806
8807 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8808                             struct btrfs_root *root, u64 type)
8809 {
8810         u64 alloc_flags = get_alloc_profile(root, type);
8811         return do_chunk_alloc(trans, root, alloc_flags,
8812                               CHUNK_ALLOC_FORCE);
8813 }
8814
8815 /*
8816  * helper to account the unused space of all the readonly block group in the
8817  * space_info. takes mirrors into account.
8818  */
8819 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8820 {
8821         struct btrfs_block_group_cache *block_group;
8822         u64 free_bytes = 0;
8823         int factor;
8824
8825         /* It's df, we don't care if it's racey */
8826         if (list_empty(&sinfo->ro_bgs))
8827                 return 0;
8828
8829         spin_lock(&sinfo->lock);
8830         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
8831                 spin_lock(&block_group->lock);
8832
8833                 if (!block_group->ro) {
8834                         spin_unlock(&block_group->lock);
8835                         continue;
8836                 }
8837
8838                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8839                                           BTRFS_BLOCK_GROUP_RAID10 |
8840                                           BTRFS_BLOCK_GROUP_DUP))
8841                         factor = 2;
8842                 else
8843                         factor = 1;
8844
8845                 free_bytes += (block_group->key.offset -
8846                                btrfs_block_group_used(&block_group->item)) *
8847                                factor;
8848
8849                 spin_unlock(&block_group->lock);
8850         }
8851         spin_unlock(&sinfo->lock);
8852
8853         return free_bytes;
8854 }
8855
8856 void btrfs_set_block_group_rw(struct btrfs_root *root,
8857                               struct btrfs_block_group_cache *cache)
8858 {
8859         struct btrfs_space_info *sinfo = cache->space_info;
8860         u64 num_bytes;
8861
8862         BUG_ON(!cache->ro);
8863
8864         spin_lock(&sinfo->lock);
8865         spin_lock(&cache->lock);
8866         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8867                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8868         sinfo->bytes_readonly -= num_bytes;
8869         cache->ro = 0;
8870         list_del_init(&cache->ro_list);
8871         spin_unlock(&cache->lock);
8872         spin_unlock(&sinfo->lock);
8873 }
8874
8875 /*
8876  * checks to see if its even possible to relocate this block group.
8877  *
8878  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8879  * ok to go ahead and try.
8880  */
8881 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8882 {
8883         struct btrfs_block_group_cache *block_group;
8884         struct btrfs_space_info *space_info;
8885         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8886         struct btrfs_device *device;
8887         struct btrfs_trans_handle *trans;
8888         u64 min_free;
8889         u64 dev_min = 1;
8890         u64 dev_nr = 0;
8891         u64 target;
8892         int index;
8893         int full = 0;
8894         int ret = 0;
8895
8896         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8897
8898         /* odd, couldn't find the block group, leave it alone */
8899         if (!block_group)
8900                 return -1;
8901
8902         min_free = btrfs_block_group_used(&block_group->item);
8903
8904         /* no bytes used, we're good */
8905         if (!min_free)
8906                 goto out;
8907
8908         space_info = block_group->space_info;
8909         spin_lock(&space_info->lock);
8910
8911         full = space_info->full;
8912
8913         /*
8914          * if this is the last block group we have in this space, we can't
8915          * relocate it unless we're able to allocate a new chunk below.
8916          *
8917          * Otherwise, we need to make sure we have room in the space to handle
8918          * all of the extents from this block group.  If we can, we're good
8919          */
8920         if ((space_info->total_bytes != block_group->key.offset) &&
8921             (space_info->bytes_used + space_info->bytes_reserved +
8922              space_info->bytes_pinned + space_info->bytes_readonly +
8923              min_free < space_info->total_bytes)) {
8924                 spin_unlock(&space_info->lock);
8925                 goto out;
8926         }
8927         spin_unlock(&space_info->lock);
8928
8929         /*
8930          * ok we don't have enough space, but maybe we have free space on our
8931          * devices to allocate new chunks for relocation, so loop through our
8932          * alloc devices and guess if we have enough space.  if this block
8933          * group is going to be restriped, run checks against the target
8934          * profile instead of the current one.
8935          */
8936         ret = -1;
8937
8938         /*
8939          * index:
8940          *      0: raid10
8941          *      1: raid1
8942          *      2: dup
8943          *      3: raid0
8944          *      4: single
8945          */
8946         target = get_restripe_target(root->fs_info, block_group->flags);
8947         if (target) {
8948                 index = __get_raid_index(extended_to_chunk(target));
8949         } else {
8950                 /*
8951                  * this is just a balance, so if we were marked as full
8952                  * we know there is no space for a new chunk
8953                  */
8954                 if (full)
8955                         goto out;
8956
8957                 index = get_block_group_index(block_group);
8958         }
8959
8960         if (index == BTRFS_RAID_RAID10) {
8961                 dev_min = 4;
8962                 /* Divide by 2 */
8963                 min_free >>= 1;
8964         } else if (index == BTRFS_RAID_RAID1) {
8965                 dev_min = 2;
8966         } else if (index == BTRFS_RAID_DUP) {
8967                 /* Multiply by 2 */
8968                 min_free <<= 1;
8969         } else if (index == BTRFS_RAID_RAID0) {
8970                 dev_min = fs_devices->rw_devices;
8971                 min_free = div64_u64(min_free, dev_min);
8972         }
8973
8974         /* We need to do this so that we can look at pending chunks */
8975         trans = btrfs_join_transaction(root);
8976         if (IS_ERR(trans)) {
8977                 ret = PTR_ERR(trans);
8978                 goto out;
8979         }
8980
8981         mutex_lock(&root->fs_info->chunk_mutex);
8982         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8983                 u64 dev_offset;
8984
8985                 /*
8986                  * check to make sure we can actually find a chunk with enough
8987                  * space to fit our block group in.
8988                  */
8989                 if (device->total_bytes > device->bytes_used + min_free &&
8990                     !device->is_tgtdev_for_dev_replace) {
8991                         ret = find_free_dev_extent(trans, device, min_free,
8992                                                    &dev_offset, NULL);
8993                         if (!ret)
8994                                 dev_nr++;
8995
8996                         if (dev_nr >= dev_min)
8997                                 break;
8998
8999                         ret = -1;
9000                 }
9001         }
9002         mutex_unlock(&root->fs_info->chunk_mutex);
9003         btrfs_end_transaction(trans, root);
9004 out:
9005         btrfs_put_block_group(block_group);
9006         return ret;
9007 }
9008
9009 static int find_first_block_group(struct btrfs_root *root,
9010                 struct btrfs_path *path, struct btrfs_key *key)
9011 {
9012         int ret = 0;
9013         struct btrfs_key found_key;
9014         struct extent_buffer *leaf;
9015         int slot;
9016
9017         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9018         if (ret < 0)
9019                 goto out;
9020
9021         while (1) {
9022                 slot = path->slots[0];
9023                 leaf = path->nodes[0];
9024                 if (slot >= btrfs_header_nritems(leaf)) {
9025                         ret = btrfs_next_leaf(root, path);
9026                         if (ret == 0)
9027                                 continue;
9028                         if (ret < 0)
9029                                 goto out;
9030                         break;
9031                 }
9032                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9033
9034                 if (found_key.objectid >= key->objectid &&
9035                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9036                         ret = 0;
9037                         goto out;
9038                 }
9039                 path->slots[0]++;
9040         }
9041 out:
9042         return ret;
9043 }
9044
9045 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9046 {
9047         struct btrfs_block_group_cache *block_group;
9048         u64 last = 0;
9049
9050         while (1) {
9051                 struct inode *inode;
9052
9053                 block_group = btrfs_lookup_first_block_group(info, last);
9054                 while (block_group) {
9055                         spin_lock(&block_group->lock);
9056                         if (block_group->iref)
9057                                 break;
9058                         spin_unlock(&block_group->lock);
9059                         block_group = next_block_group(info->tree_root,
9060                                                        block_group);
9061                 }
9062                 if (!block_group) {
9063                         if (last == 0)
9064                                 break;
9065                         last = 0;
9066                         continue;
9067                 }
9068
9069                 inode = block_group->inode;
9070                 block_group->iref = 0;
9071                 block_group->inode = NULL;
9072                 spin_unlock(&block_group->lock);
9073                 iput(inode);
9074                 last = block_group->key.objectid + block_group->key.offset;
9075                 btrfs_put_block_group(block_group);
9076         }
9077 }
9078
9079 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9080 {
9081         struct btrfs_block_group_cache *block_group;
9082         struct btrfs_space_info *space_info;
9083         struct btrfs_caching_control *caching_ctl;
9084         struct rb_node *n;
9085
9086         down_write(&info->commit_root_sem);
9087         while (!list_empty(&info->caching_block_groups)) {
9088                 caching_ctl = list_entry(info->caching_block_groups.next,
9089                                          struct btrfs_caching_control, list);
9090                 list_del(&caching_ctl->list);
9091                 put_caching_control(caching_ctl);
9092         }
9093         up_write(&info->commit_root_sem);
9094
9095         spin_lock(&info->unused_bgs_lock);
9096         while (!list_empty(&info->unused_bgs)) {
9097                 block_group = list_first_entry(&info->unused_bgs,
9098                                                struct btrfs_block_group_cache,
9099                                                bg_list);
9100                 list_del_init(&block_group->bg_list);
9101                 btrfs_put_block_group(block_group);
9102         }
9103         spin_unlock(&info->unused_bgs_lock);
9104
9105         spin_lock(&info->block_group_cache_lock);
9106         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9107                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9108                                        cache_node);
9109                 rb_erase(&block_group->cache_node,
9110                          &info->block_group_cache_tree);
9111                 RB_CLEAR_NODE(&block_group->cache_node);
9112                 spin_unlock(&info->block_group_cache_lock);
9113
9114                 down_write(&block_group->space_info->groups_sem);
9115                 list_del(&block_group->list);
9116                 up_write(&block_group->space_info->groups_sem);
9117
9118                 if (block_group->cached == BTRFS_CACHE_STARTED)
9119                         wait_block_group_cache_done(block_group);
9120
9121                 /*
9122                  * We haven't cached this block group, which means we could
9123                  * possibly have excluded extents on this block group.
9124                  */
9125                 if (block_group->cached == BTRFS_CACHE_NO ||
9126                     block_group->cached == BTRFS_CACHE_ERROR)
9127                         free_excluded_extents(info->extent_root, block_group);
9128
9129                 btrfs_remove_free_space_cache(block_group);
9130                 btrfs_put_block_group(block_group);
9131
9132                 spin_lock(&info->block_group_cache_lock);
9133         }
9134         spin_unlock(&info->block_group_cache_lock);
9135
9136         /* now that all the block groups are freed, go through and
9137          * free all the space_info structs.  This is only called during
9138          * the final stages of unmount, and so we know nobody is
9139          * using them.  We call synchronize_rcu() once before we start,
9140          * just to be on the safe side.
9141          */
9142         synchronize_rcu();
9143
9144         release_global_block_rsv(info);
9145
9146         while (!list_empty(&info->space_info)) {
9147                 int i;
9148
9149                 space_info = list_entry(info->space_info.next,
9150                                         struct btrfs_space_info,
9151                                         list);
9152                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9153                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9154                             space_info->bytes_reserved > 0 ||
9155                             space_info->bytes_may_use > 0)) {
9156                                 dump_space_info(space_info, 0, 0);
9157                         }
9158                 }
9159                 list_del(&space_info->list);
9160                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9161                         struct kobject *kobj;
9162                         kobj = space_info->block_group_kobjs[i];
9163                         space_info->block_group_kobjs[i] = NULL;
9164                         if (kobj) {
9165                                 kobject_del(kobj);
9166                                 kobject_put(kobj);
9167                         }
9168                 }
9169                 kobject_del(&space_info->kobj);
9170                 kobject_put(&space_info->kobj);
9171         }
9172         return 0;
9173 }
9174
9175 static void __link_block_group(struct btrfs_space_info *space_info,
9176                                struct btrfs_block_group_cache *cache)
9177 {
9178         int index = get_block_group_index(cache);
9179         bool first = false;
9180
9181         down_write(&space_info->groups_sem);
9182         if (list_empty(&space_info->block_groups[index]))
9183                 first = true;
9184         list_add_tail(&cache->list, &space_info->block_groups[index]);
9185         up_write(&space_info->groups_sem);
9186
9187         if (first) {
9188                 struct raid_kobject *rkobj;
9189                 int ret;
9190
9191                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9192                 if (!rkobj)
9193                         goto out_err;
9194                 rkobj->raid_type = index;
9195                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9196                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9197                                   "%s", get_raid_name(index));
9198                 if (ret) {
9199                         kobject_put(&rkobj->kobj);
9200                         goto out_err;
9201                 }
9202                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9203         }
9204
9205         return;
9206 out_err:
9207         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9208 }
9209
9210 static struct btrfs_block_group_cache *
9211 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9212 {
9213         struct btrfs_block_group_cache *cache;
9214
9215         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9216         if (!cache)
9217                 return NULL;
9218
9219         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9220                                         GFP_NOFS);
9221         if (!cache->free_space_ctl) {
9222                 kfree(cache);
9223                 return NULL;
9224         }
9225
9226         cache->key.objectid = start;
9227         cache->key.offset = size;
9228         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9229
9230         cache->sectorsize = root->sectorsize;
9231         cache->fs_info = root->fs_info;
9232         cache->full_stripe_len = btrfs_full_stripe_len(root,
9233                                                &root->fs_info->mapping_tree,
9234                                                start);
9235         atomic_set(&cache->count, 1);
9236         spin_lock_init(&cache->lock);
9237         init_rwsem(&cache->data_rwsem);
9238         INIT_LIST_HEAD(&cache->list);
9239         INIT_LIST_HEAD(&cache->cluster_list);
9240         INIT_LIST_HEAD(&cache->bg_list);
9241         INIT_LIST_HEAD(&cache->ro_list);
9242         INIT_LIST_HEAD(&cache->dirty_list);
9243         INIT_LIST_HEAD(&cache->io_list);
9244         btrfs_init_free_space_ctl(cache);
9245         atomic_set(&cache->trimming, 0);
9246
9247         return cache;
9248 }
9249
9250 int btrfs_read_block_groups(struct btrfs_root *root)
9251 {
9252         struct btrfs_path *path;
9253         int ret;
9254         struct btrfs_block_group_cache *cache;
9255         struct btrfs_fs_info *info = root->fs_info;
9256         struct btrfs_space_info *space_info;
9257         struct btrfs_key key;
9258         struct btrfs_key found_key;
9259         struct extent_buffer *leaf;
9260         int need_clear = 0;
9261         u64 cache_gen;
9262
9263         root = info->extent_root;
9264         key.objectid = 0;
9265         key.offset = 0;
9266         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9267         path = btrfs_alloc_path();
9268         if (!path)
9269                 return -ENOMEM;
9270         path->reada = 1;
9271
9272         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9273         if (btrfs_test_opt(root, SPACE_CACHE) &&
9274             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9275                 need_clear = 1;
9276         if (btrfs_test_opt(root, CLEAR_CACHE))
9277                 need_clear = 1;
9278
9279         while (1) {
9280                 ret = find_first_block_group(root, path, &key);
9281                 if (ret > 0)
9282                         break;
9283                 if (ret != 0)
9284                         goto error;
9285
9286                 leaf = path->nodes[0];
9287                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9288
9289                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9290                                                        found_key.offset);
9291                 if (!cache) {
9292                         ret = -ENOMEM;
9293                         goto error;
9294                 }
9295
9296                 if (need_clear) {
9297                         /*
9298                          * When we mount with old space cache, we need to
9299                          * set BTRFS_DC_CLEAR and set dirty flag.
9300                          *
9301                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9302                          *    truncate the old free space cache inode and
9303                          *    setup a new one.
9304                          * b) Setting 'dirty flag' makes sure that we flush
9305                          *    the new space cache info onto disk.
9306                          */
9307                         if (btrfs_test_opt(root, SPACE_CACHE))
9308                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9309                 }
9310
9311                 read_extent_buffer(leaf, &cache->item,
9312                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9313                                    sizeof(cache->item));
9314                 cache->flags = btrfs_block_group_flags(&cache->item);
9315
9316                 key.objectid = found_key.objectid + found_key.offset;
9317                 btrfs_release_path(path);
9318
9319                 /*
9320                  * We need to exclude the super stripes now so that the space
9321                  * info has super bytes accounted for, otherwise we'll think
9322                  * we have more space than we actually do.
9323                  */
9324                 ret = exclude_super_stripes(root, cache);
9325                 if (ret) {
9326                         /*
9327                          * We may have excluded something, so call this just in
9328                          * case.
9329                          */
9330                         free_excluded_extents(root, cache);
9331                         btrfs_put_block_group(cache);
9332                         goto error;
9333                 }
9334
9335                 /*
9336                  * check for two cases, either we are full, and therefore
9337                  * don't need to bother with the caching work since we won't
9338                  * find any space, or we are empty, and we can just add all
9339                  * the space in and be done with it.  This saves us _alot_ of
9340                  * time, particularly in the full case.
9341                  */
9342                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9343                         cache->last_byte_to_unpin = (u64)-1;
9344                         cache->cached = BTRFS_CACHE_FINISHED;
9345                         free_excluded_extents(root, cache);
9346                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9347                         cache->last_byte_to_unpin = (u64)-1;
9348                         cache->cached = BTRFS_CACHE_FINISHED;
9349                         add_new_free_space(cache, root->fs_info,
9350                                            found_key.objectid,
9351                                            found_key.objectid +
9352                                            found_key.offset);
9353                         free_excluded_extents(root, cache);
9354                 }
9355
9356                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9357                 if (ret) {
9358                         btrfs_remove_free_space_cache(cache);
9359                         btrfs_put_block_group(cache);
9360                         goto error;
9361                 }
9362
9363                 ret = update_space_info(info, cache->flags, found_key.offset,
9364                                         btrfs_block_group_used(&cache->item),
9365                                         &space_info);
9366                 if (ret) {
9367                         btrfs_remove_free_space_cache(cache);
9368                         spin_lock(&info->block_group_cache_lock);
9369                         rb_erase(&cache->cache_node,
9370                                  &info->block_group_cache_tree);
9371                         RB_CLEAR_NODE(&cache->cache_node);
9372                         spin_unlock(&info->block_group_cache_lock);
9373                         btrfs_put_block_group(cache);
9374                         goto error;
9375                 }
9376
9377                 cache->space_info = space_info;
9378                 spin_lock(&cache->space_info->lock);
9379                 cache->space_info->bytes_readonly += cache->bytes_super;
9380                 spin_unlock(&cache->space_info->lock);
9381
9382                 __link_block_group(space_info, cache);
9383
9384                 set_avail_alloc_bits(root->fs_info, cache->flags);
9385                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9386                         set_block_group_ro(cache, 1);
9387                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9388                         spin_lock(&info->unused_bgs_lock);
9389                         /* Should always be true but just in case. */
9390                         if (list_empty(&cache->bg_list)) {
9391                                 btrfs_get_block_group(cache);
9392                                 list_add_tail(&cache->bg_list,
9393                                               &info->unused_bgs);
9394                         }
9395                         spin_unlock(&info->unused_bgs_lock);
9396                 }
9397         }
9398
9399         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9400                 if (!(get_alloc_profile(root, space_info->flags) &
9401                       (BTRFS_BLOCK_GROUP_RAID10 |
9402                        BTRFS_BLOCK_GROUP_RAID1 |
9403                        BTRFS_BLOCK_GROUP_RAID5 |
9404                        BTRFS_BLOCK_GROUP_RAID6 |
9405                        BTRFS_BLOCK_GROUP_DUP)))
9406                         continue;
9407                 /*
9408                  * avoid allocating from un-mirrored block group if there are
9409                  * mirrored block groups.
9410                  */
9411                 list_for_each_entry(cache,
9412                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9413                                 list)
9414                         set_block_group_ro(cache, 1);
9415                 list_for_each_entry(cache,
9416                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9417                                 list)
9418                         set_block_group_ro(cache, 1);
9419         }
9420
9421         init_global_block_rsv(info);
9422         ret = 0;
9423 error:
9424         btrfs_free_path(path);
9425         return ret;
9426 }
9427
9428 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9429                                        struct btrfs_root *root)
9430 {
9431         struct btrfs_block_group_cache *block_group, *tmp;
9432         struct btrfs_root *extent_root = root->fs_info->extent_root;
9433         struct btrfs_block_group_item item;
9434         struct btrfs_key key;
9435         int ret = 0;
9436
9437         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9438                 if (ret)
9439                         goto next;
9440
9441                 spin_lock(&block_group->lock);
9442                 memcpy(&item, &block_group->item, sizeof(item));
9443                 memcpy(&key, &block_group->key, sizeof(key));
9444                 spin_unlock(&block_group->lock);
9445
9446                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9447                                         sizeof(item));
9448                 if (ret)
9449                         btrfs_abort_transaction(trans, extent_root, ret);
9450                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9451                                                key.objectid, key.offset);
9452                 if (ret)
9453                         btrfs_abort_transaction(trans, extent_root, ret);
9454 next:
9455                 list_del_init(&block_group->bg_list);
9456         }
9457 }
9458
9459 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9460                            struct btrfs_root *root, u64 bytes_used,
9461                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9462                            u64 size)
9463 {
9464         int ret;
9465         struct btrfs_root *extent_root;
9466         struct btrfs_block_group_cache *cache;
9467
9468         extent_root = root->fs_info->extent_root;
9469
9470         btrfs_set_log_full_commit(root->fs_info, trans);
9471
9472         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9473         if (!cache)
9474                 return -ENOMEM;
9475
9476         btrfs_set_block_group_used(&cache->item, bytes_used);
9477         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9478         btrfs_set_block_group_flags(&cache->item, type);
9479
9480         cache->flags = type;
9481         cache->last_byte_to_unpin = (u64)-1;
9482         cache->cached = BTRFS_CACHE_FINISHED;
9483         ret = exclude_super_stripes(root, cache);
9484         if (ret) {
9485                 /*
9486                  * We may have excluded something, so call this just in
9487                  * case.
9488                  */
9489                 free_excluded_extents(root, cache);
9490                 btrfs_put_block_group(cache);
9491                 return ret;
9492         }
9493
9494         add_new_free_space(cache, root->fs_info, chunk_offset,
9495                            chunk_offset + size);
9496
9497         free_excluded_extents(root, cache);
9498
9499         ret = btrfs_add_block_group_cache(root->fs_info, cache);
9500         if (ret) {
9501                 btrfs_remove_free_space_cache(cache);
9502                 btrfs_put_block_group(cache);
9503                 return ret;
9504         }
9505
9506         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
9507                                 &cache->space_info);
9508         if (ret) {
9509                 btrfs_remove_free_space_cache(cache);
9510                 spin_lock(&root->fs_info->block_group_cache_lock);
9511                 rb_erase(&cache->cache_node,
9512                          &root->fs_info->block_group_cache_tree);
9513                 RB_CLEAR_NODE(&cache->cache_node);
9514                 spin_unlock(&root->fs_info->block_group_cache_lock);
9515                 btrfs_put_block_group(cache);
9516                 return ret;
9517         }
9518         update_global_block_rsv(root->fs_info);
9519
9520         spin_lock(&cache->space_info->lock);
9521         cache->space_info->bytes_readonly += cache->bytes_super;
9522         spin_unlock(&cache->space_info->lock);
9523
9524         __link_block_group(cache->space_info, cache);
9525
9526         list_add_tail(&cache->bg_list, &trans->new_bgs);
9527
9528         set_avail_alloc_bits(extent_root->fs_info, type);
9529
9530         return 0;
9531 }
9532
9533 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9534 {
9535         u64 extra_flags = chunk_to_extended(flags) &
9536                                 BTRFS_EXTENDED_PROFILE_MASK;
9537
9538         write_seqlock(&fs_info->profiles_lock);
9539         if (flags & BTRFS_BLOCK_GROUP_DATA)
9540                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9541         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9542                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9543         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9544                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9545         write_sequnlock(&fs_info->profiles_lock);
9546 }
9547
9548 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9549                              struct btrfs_root *root, u64 group_start,
9550                              struct extent_map *em)
9551 {
9552         struct btrfs_path *path;
9553         struct btrfs_block_group_cache *block_group;
9554         struct btrfs_free_cluster *cluster;
9555         struct btrfs_root *tree_root = root->fs_info->tree_root;
9556         struct btrfs_key key;
9557         struct inode *inode;
9558         struct kobject *kobj = NULL;
9559         int ret;
9560         int index;
9561         int factor;
9562         struct btrfs_caching_control *caching_ctl = NULL;
9563         bool remove_em;
9564
9565         root = root->fs_info->extent_root;
9566
9567         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9568         BUG_ON(!block_group);
9569         BUG_ON(!block_group->ro);
9570
9571         /*
9572          * Free the reserved super bytes from this block group before
9573          * remove it.
9574          */
9575         free_excluded_extents(root, block_group);
9576
9577         memcpy(&key, &block_group->key, sizeof(key));
9578         index = get_block_group_index(block_group);
9579         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9580                                   BTRFS_BLOCK_GROUP_RAID1 |
9581                                   BTRFS_BLOCK_GROUP_RAID10))
9582                 factor = 2;
9583         else
9584                 factor = 1;
9585
9586         /* make sure this block group isn't part of an allocation cluster */
9587         cluster = &root->fs_info->data_alloc_cluster;
9588         spin_lock(&cluster->refill_lock);
9589         btrfs_return_cluster_to_free_space(block_group, cluster);
9590         spin_unlock(&cluster->refill_lock);
9591
9592         /*
9593          * make sure this block group isn't part of a metadata
9594          * allocation cluster
9595          */
9596         cluster = &root->fs_info->meta_alloc_cluster;
9597         spin_lock(&cluster->refill_lock);
9598         btrfs_return_cluster_to_free_space(block_group, cluster);
9599         spin_unlock(&cluster->refill_lock);
9600
9601         path = btrfs_alloc_path();
9602         if (!path) {
9603                 ret = -ENOMEM;
9604                 goto out;
9605         }
9606
9607         /*
9608          * get the inode first so any iput calls done for the io_list
9609          * aren't the final iput (no unlinks allowed now)
9610          */
9611         inode = lookup_free_space_inode(tree_root, block_group, path);
9612
9613         mutex_lock(&trans->transaction->cache_write_mutex);
9614         /*
9615          * make sure our free spache cache IO is done before remove the
9616          * free space inode
9617          */
9618         spin_lock(&trans->transaction->dirty_bgs_lock);
9619         if (!list_empty(&block_group->io_list)) {
9620                 list_del_init(&block_group->io_list);
9621
9622                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
9623
9624                 spin_unlock(&trans->transaction->dirty_bgs_lock);
9625                 btrfs_wait_cache_io(root, trans, block_group,
9626                                     &block_group->io_ctl, path,
9627                                     block_group->key.objectid);
9628                 btrfs_put_block_group(block_group);
9629                 spin_lock(&trans->transaction->dirty_bgs_lock);
9630         }
9631
9632         if (!list_empty(&block_group->dirty_list)) {
9633                 list_del_init(&block_group->dirty_list);
9634                 btrfs_put_block_group(block_group);
9635         }
9636         spin_unlock(&trans->transaction->dirty_bgs_lock);
9637         mutex_unlock(&trans->transaction->cache_write_mutex);
9638
9639         if (!IS_ERR(inode)) {
9640                 ret = btrfs_orphan_add(trans, inode);
9641                 if (ret) {
9642                         btrfs_add_delayed_iput(inode);
9643                         goto out;
9644                 }
9645                 clear_nlink(inode);
9646                 /* One for the block groups ref */
9647                 spin_lock(&block_group->lock);
9648                 if (block_group->iref) {
9649                         block_group->iref = 0;
9650                         block_group->inode = NULL;
9651                         spin_unlock(&block_group->lock);
9652                         iput(inode);
9653                 } else {
9654                         spin_unlock(&block_group->lock);
9655                 }
9656                 /* One for our lookup ref */
9657                 btrfs_add_delayed_iput(inode);
9658         }
9659
9660         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9661         key.offset = block_group->key.objectid;
9662         key.type = 0;
9663
9664         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9665         if (ret < 0)
9666                 goto out;
9667         if (ret > 0)
9668                 btrfs_release_path(path);
9669         if (ret == 0) {
9670                 ret = btrfs_del_item(trans, tree_root, path);
9671                 if (ret)
9672                         goto out;
9673                 btrfs_release_path(path);
9674         }
9675
9676         spin_lock(&root->fs_info->block_group_cache_lock);
9677         rb_erase(&block_group->cache_node,
9678                  &root->fs_info->block_group_cache_tree);
9679         RB_CLEAR_NODE(&block_group->cache_node);
9680
9681         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9682                 root->fs_info->first_logical_byte = (u64)-1;
9683         spin_unlock(&root->fs_info->block_group_cache_lock);
9684
9685         down_write(&block_group->space_info->groups_sem);
9686         /*
9687          * we must use list_del_init so people can check to see if they
9688          * are still on the list after taking the semaphore
9689          */
9690         list_del_init(&block_group->list);
9691         if (list_empty(&block_group->space_info->block_groups[index])) {
9692                 kobj = block_group->space_info->block_group_kobjs[index];
9693                 block_group->space_info->block_group_kobjs[index] = NULL;
9694                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9695         }
9696         up_write(&block_group->space_info->groups_sem);
9697         if (kobj) {
9698                 kobject_del(kobj);
9699                 kobject_put(kobj);
9700         }
9701
9702         if (block_group->has_caching_ctl)
9703                 caching_ctl = get_caching_control(block_group);
9704         if (block_group->cached == BTRFS_CACHE_STARTED)
9705                 wait_block_group_cache_done(block_group);
9706         if (block_group->has_caching_ctl) {
9707                 down_write(&root->fs_info->commit_root_sem);
9708                 if (!caching_ctl) {
9709                         struct btrfs_caching_control *ctl;
9710
9711                         list_for_each_entry(ctl,
9712                                     &root->fs_info->caching_block_groups, list)
9713                                 if (ctl->block_group == block_group) {
9714                                         caching_ctl = ctl;
9715                                         atomic_inc(&caching_ctl->count);
9716                                         break;
9717                                 }
9718                 }
9719                 if (caching_ctl)
9720                         list_del_init(&caching_ctl->list);
9721                 up_write(&root->fs_info->commit_root_sem);
9722                 if (caching_ctl) {
9723                         /* Once for the caching bgs list and once for us. */
9724                         put_caching_control(caching_ctl);
9725                         put_caching_control(caching_ctl);
9726                 }
9727         }
9728
9729         spin_lock(&trans->transaction->dirty_bgs_lock);
9730         if (!list_empty(&block_group->dirty_list)) {
9731                 WARN_ON(1);
9732         }
9733         if (!list_empty(&block_group->io_list)) {
9734                 WARN_ON(1);
9735         }
9736         spin_unlock(&trans->transaction->dirty_bgs_lock);
9737         btrfs_remove_free_space_cache(block_group);
9738
9739         spin_lock(&block_group->space_info->lock);
9740         list_del_init(&block_group->ro_list);
9741
9742         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
9743                 WARN_ON(block_group->space_info->total_bytes
9744                         < block_group->key.offset);
9745                 WARN_ON(block_group->space_info->bytes_readonly
9746                         < block_group->key.offset);
9747                 WARN_ON(block_group->space_info->disk_total
9748                         < block_group->key.offset * factor);
9749         }
9750         block_group->space_info->total_bytes -= block_group->key.offset;
9751         block_group->space_info->bytes_readonly -= block_group->key.offset;
9752         block_group->space_info->disk_total -= block_group->key.offset * factor;
9753
9754         spin_unlock(&block_group->space_info->lock);
9755
9756         memcpy(&key, &block_group->key, sizeof(key));
9757
9758         lock_chunks(root);
9759         if (!list_empty(&em->list)) {
9760                 /* We're in the transaction->pending_chunks list. */
9761                 free_extent_map(em);
9762         }
9763         spin_lock(&block_group->lock);
9764         block_group->removed = 1;
9765         /*
9766          * At this point trimming can't start on this block group, because we
9767          * removed the block group from the tree fs_info->block_group_cache_tree
9768          * so no one can't find it anymore and even if someone already got this
9769          * block group before we removed it from the rbtree, they have already
9770          * incremented block_group->trimming - if they didn't, they won't find
9771          * any free space entries because we already removed them all when we
9772          * called btrfs_remove_free_space_cache().
9773          *
9774          * And we must not remove the extent map from the fs_info->mapping_tree
9775          * to prevent the same logical address range and physical device space
9776          * ranges from being reused for a new block group. This is because our
9777          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
9778          * completely transactionless, so while it is trimming a range the
9779          * currently running transaction might finish and a new one start,
9780          * allowing for new block groups to be created that can reuse the same
9781          * physical device locations unless we take this special care.
9782          */
9783         remove_em = (atomic_read(&block_group->trimming) == 0);
9784         /*
9785          * Make sure a trimmer task always sees the em in the pinned_chunks list
9786          * if it sees block_group->removed == 1 (needs to lock block_group->lock
9787          * before checking block_group->removed).
9788          */
9789         if (!remove_em) {
9790                 /*
9791                  * Our em might be in trans->transaction->pending_chunks which
9792                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
9793                  * and so is the fs_info->pinned_chunks list.
9794                  *
9795                  * So at this point we must be holding the chunk_mutex to avoid
9796                  * any races with chunk allocation (more specifically at
9797                  * volumes.c:contains_pending_extent()), to ensure it always
9798                  * sees the em, either in the pending_chunks list or in the
9799                  * pinned_chunks list.
9800                  */
9801                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
9802         }
9803         spin_unlock(&block_group->lock);
9804
9805         if (remove_em) {
9806                 struct extent_map_tree *em_tree;
9807
9808                 em_tree = &root->fs_info->mapping_tree.map_tree;
9809                 write_lock(&em_tree->lock);
9810                 /*
9811                  * The em might be in the pending_chunks list, so make sure the
9812                  * chunk mutex is locked, since remove_extent_mapping() will
9813                  * delete us from that list.
9814                  */
9815                 remove_extent_mapping(em_tree, em);
9816                 write_unlock(&em_tree->lock);
9817                 /* once for the tree */
9818                 free_extent_map(em);
9819         }
9820
9821         unlock_chunks(root);
9822
9823         btrfs_put_block_group(block_group);
9824         btrfs_put_block_group(block_group);
9825
9826         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9827         if (ret > 0)
9828                 ret = -EIO;
9829         if (ret < 0)
9830                 goto out;
9831
9832         ret = btrfs_del_item(trans, root, path);
9833 out:
9834         btrfs_free_path(path);
9835         return ret;
9836 }
9837
9838 /*
9839  * Process the unused_bgs list and remove any that don't have any allocated
9840  * space inside of them.
9841  */
9842 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
9843 {
9844         struct btrfs_block_group_cache *block_group;
9845         struct btrfs_space_info *space_info;
9846         struct btrfs_root *root = fs_info->extent_root;
9847         struct btrfs_trans_handle *trans;
9848         int ret = 0;
9849
9850         if (!fs_info->open)
9851                 return;
9852
9853         spin_lock(&fs_info->unused_bgs_lock);
9854         while (!list_empty(&fs_info->unused_bgs)) {
9855                 u64 start, end;
9856
9857                 block_group = list_first_entry(&fs_info->unused_bgs,
9858                                                struct btrfs_block_group_cache,
9859                                                bg_list);
9860                 space_info = block_group->space_info;
9861                 list_del_init(&block_group->bg_list);
9862                 if (ret || btrfs_mixed_space_info(space_info)) {
9863                         btrfs_put_block_group(block_group);
9864                         continue;
9865                 }
9866                 spin_unlock(&fs_info->unused_bgs_lock);
9867
9868                 /* Don't want to race with allocators so take the groups_sem */
9869                 down_write(&space_info->groups_sem);
9870                 spin_lock(&block_group->lock);
9871                 if (block_group->reserved ||
9872                     btrfs_block_group_used(&block_group->item) ||
9873                     block_group->ro) {
9874                         /*
9875                          * We want to bail if we made new allocations or have
9876                          * outstanding allocations in this block group.  We do
9877                          * the ro check in case balance is currently acting on
9878                          * this block group.
9879                          */
9880                         spin_unlock(&block_group->lock);
9881                         up_write(&space_info->groups_sem);
9882                         goto next;
9883                 }
9884                 spin_unlock(&block_group->lock);
9885
9886                 /* We don't want to force the issue, only flip if it's ok. */
9887                 ret = set_block_group_ro(block_group, 0);
9888                 up_write(&space_info->groups_sem);
9889                 if (ret < 0) {
9890                         ret = 0;
9891                         goto next;
9892                 }
9893
9894                 /*
9895                  * Want to do this before we do anything else so we can recover
9896                  * properly if we fail to join the transaction.
9897                  */
9898                 /* 1 for btrfs_orphan_reserve_metadata() */
9899                 trans = btrfs_start_transaction(root, 1);
9900                 if (IS_ERR(trans)) {
9901                         btrfs_set_block_group_rw(root, block_group);
9902                         ret = PTR_ERR(trans);
9903                         goto next;
9904                 }
9905
9906                 /*
9907                  * We could have pending pinned extents for this block group,
9908                  * just delete them, we don't care about them anymore.
9909                  */
9910                 start = block_group->key.objectid;
9911                 end = start + block_group->key.offset - 1;
9912                 /*
9913                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
9914                  * btrfs_finish_extent_commit(). If we are at transaction N,
9915                  * another task might be running finish_extent_commit() for the
9916                  * previous transaction N - 1, and have seen a range belonging
9917                  * to the block group in freed_extents[] before we were able to
9918                  * clear the whole block group range from freed_extents[]. This
9919                  * means that task can lookup for the block group after we
9920                  * unpinned it from freed_extents[] and removed it, leading to
9921                  * a BUG_ON() at btrfs_unpin_extent_range().
9922                  */
9923                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
9924                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
9925                                   EXTENT_DIRTY, GFP_NOFS);
9926                 if (ret) {
9927                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9928                         btrfs_set_block_group_rw(root, block_group);
9929                         goto end_trans;
9930                 }
9931                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
9932                                   EXTENT_DIRTY, GFP_NOFS);
9933                 if (ret) {
9934                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9935                         btrfs_set_block_group_rw(root, block_group);
9936                         goto end_trans;
9937                 }
9938                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9939
9940                 /* Reset pinned so btrfs_put_block_group doesn't complain */
9941                 spin_lock(&space_info->lock);
9942                 spin_lock(&block_group->lock);
9943
9944                 space_info->bytes_pinned -= block_group->pinned;
9945                 space_info->bytes_readonly += block_group->pinned;
9946                 percpu_counter_add(&space_info->total_bytes_pinned,
9947                                    -block_group->pinned);
9948                 block_group->pinned = 0;
9949
9950                 spin_unlock(&block_group->lock);
9951                 spin_unlock(&space_info->lock);
9952
9953                 /*
9954                  * Btrfs_remove_chunk will abort the transaction if things go
9955                  * horribly wrong.
9956                  */
9957                 ret = btrfs_remove_chunk(trans, root,
9958                                          block_group->key.objectid);
9959 end_trans:
9960                 btrfs_end_transaction(trans, root);
9961 next:
9962                 btrfs_put_block_group(block_group);
9963                 spin_lock(&fs_info->unused_bgs_lock);
9964         }
9965         spin_unlock(&fs_info->unused_bgs_lock);
9966 }
9967
9968 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
9969 {
9970         struct btrfs_space_info *space_info;
9971         struct btrfs_super_block *disk_super;
9972         u64 features;
9973         u64 flags;
9974         int mixed = 0;
9975         int ret;
9976
9977         disk_super = fs_info->super_copy;
9978         if (!btrfs_super_root(disk_super))
9979                 return 1;
9980
9981         features = btrfs_super_incompat_flags(disk_super);
9982         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
9983                 mixed = 1;
9984
9985         flags = BTRFS_BLOCK_GROUP_SYSTEM;
9986         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9987         if (ret)
9988                 goto out;
9989
9990         if (mixed) {
9991                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
9992                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9993         } else {
9994                 flags = BTRFS_BLOCK_GROUP_METADATA;
9995                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9996                 if (ret)
9997                         goto out;
9998
9999                 flags = BTRFS_BLOCK_GROUP_DATA;
10000                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10001         }
10002 out:
10003         return ret;
10004 }
10005
10006 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10007 {
10008         return unpin_extent_range(root, start, end, false);
10009 }
10010
10011 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10012 {
10013         struct btrfs_fs_info *fs_info = root->fs_info;
10014         struct btrfs_block_group_cache *cache = NULL;
10015         u64 group_trimmed;
10016         u64 start;
10017         u64 end;
10018         u64 trimmed = 0;
10019         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10020         int ret = 0;
10021
10022         /*
10023          * try to trim all FS space, our block group may start from non-zero.
10024          */
10025         if (range->len == total_bytes)
10026                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10027         else
10028                 cache = btrfs_lookup_block_group(fs_info, range->start);
10029
10030         while (cache) {
10031                 if (cache->key.objectid >= (range->start + range->len)) {
10032                         btrfs_put_block_group(cache);
10033                         break;
10034                 }
10035
10036                 start = max(range->start, cache->key.objectid);
10037                 end = min(range->start + range->len,
10038                                 cache->key.objectid + cache->key.offset);
10039
10040                 if (end - start >= range->minlen) {
10041                         if (!block_group_cache_done(cache)) {
10042                                 ret = cache_block_group(cache, 0);
10043                                 if (ret) {
10044                                         btrfs_put_block_group(cache);
10045                                         break;
10046                                 }
10047                                 ret = wait_block_group_cache_done(cache);
10048                                 if (ret) {
10049                                         btrfs_put_block_group(cache);
10050                                         break;
10051                                 }
10052                         }
10053                         ret = btrfs_trim_block_group(cache,
10054                                                      &group_trimmed,
10055                                                      start,
10056                                                      end,
10057                                                      range->minlen);
10058
10059                         trimmed += group_trimmed;
10060                         if (ret) {
10061                                 btrfs_put_block_group(cache);
10062                                 break;
10063                         }
10064                 }
10065
10066                 cache = next_block_group(fs_info->tree_root, cache);
10067         }
10068
10069         range->len = trimmed;
10070         return ret;
10071 }
10072
10073 /*
10074  * btrfs_{start,end}_write_no_snapshoting() are similar to
10075  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10076  * data into the page cache through nocow before the subvolume is snapshoted,
10077  * but flush the data into disk after the snapshot creation, or to prevent
10078  * operations while snapshoting is ongoing and that cause the snapshot to be
10079  * inconsistent (writes followed by expanding truncates for example).
10080  */
10081 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10082 {
10083         percpu_counter_dec(&root->subv_writers->counter);
10084         /*
10085          * Make sure counter is updated before we wake up
10086          * waiters.
10087          */
10088         smp_mb();
10089         if (waitqueue_active(&root->subv_writers->wait))
10090                 wake_up(&root->subv_writers->wait);
10091 }
10092
10093 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10094 {
10095         if (atomic_read(&root->will_be_snapshoted))
10096                 return 0;
10097
10098         percpu_counter_inc(&root->subv_writers->counter);
10099         /*
10100          * Make sure counter is updated before we check for snapshot creation.
10101          */
10102         smp_mb();
10103         if (atomic_read(&root->will_be_snapshoted)) {
10104                 btrfs_end_write_no_snapshoting(root);
10105                 return 0;
10106         }
10107         return 1;
10108 }