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, 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, u64 write_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 need_commit = 2;
3862         int have_pinned_space;
3863
3864         /* make sure bytes are sectorsize aligned */
3865         bytes = ALIGN(bytes, root->sectorsize);
3866
3867         if (btrfs_is_free_space_inode(inode)) {
3868                 need_commit = 0;
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                                         have_pinned_space = 1;
3920                                         goto commit_trans;
3921                                 }
3922                         }
3923
3924                         if (!data_sinfo)
3925                                 data_sinfo = fs_info->data_sinfo;
3926
3927                         goto again;
3928                 }
3929
3930                 /*
3931                  * If we don't have enough pinned space to deal with this
3932                  * allocation, and no removed chunk in current transaction,
3933                  * don't bother committing the transaction.
3934                  */
3935                 have_pinned_space = percpu_counter_compare(
3936                         &data_sinfo->total_bytes_pinned,
3937                         used + bytes - data_sinfo->total_bytes);
3938                 spin_unlock(&data_sinfo->lock);
3939
3940                 /* commit the current transaction and try again */
3941 commit_trans:
3942                 if (need_commit &&
3943                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3944                         need_commit--;
3945
3946                         trans = btrfs_join_transaction(root);
3947                         if (IS_ERR(trans))
3948                                 return PTR_ERR(trans);
3949                         if (have_pinned_space >= 0 ||
3950                             trans->transaction->have_free_bgs ||
3951                             need_commit > 0) {
3952                                 ret = btrfs_commit_transaction(trans, root);
3953                                 if (ret)
3954                                         return ret;
3955                                 /*
3956                                  * make sure that all running delayed iput are
3957                                  * done
3958                                  */
3959                                 down_write(&root->fs_info->delayed_iput_sem);
3960                                 up_write(&root->fs_info->delayed_iput_sem);
3961                                 goto again;
3962                         } else {
3963                                 btrfs_end_transaction(trans, root);
3964                         }
3965                 }
3966
3967                 trace_btrfs_space_reservation(root->fs_info,
3968                                               "space_info:enospc",
3969                                               data_sinfo->flags, bytes, 1);
3970                 return -ENOSPC;
3971         }
3972         ret = btrfs_qgroup_reserve(root, write_bytes);
3973         if (ret)
3974                 goto out;
3975         data_sinfo->bytes_may_use += bytes;
3976         trace_btrfs_space_reservation(root->fs_info, "space_info",
3977                                       data_sinfo->flags, bytes, 1);
3978 out:
3979         spin_unlock(&data_sinfo->lock);
3980
3981         return ret;
3982 }
3983
3984 /*
3985  * Called if we need to clear a data reservation for this inode.
3986  */
3987 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3988 {
3989         struct btrfs_root *root = BTRFS_I(inode)->root;
3990         struct btrfs_space_info *data_sinfo;
3991
3992         /* make sure bytes are sectorsize aligned */
3993         bytes = ALIGN(bytes, root->sectorsize);
3994
3995         data_sinfo = root->fs_info->data_sinfo;
3996         spin_lock(&data_sinfo->lock);
3997         WARN_ON(data_sinfo->bytes_may_use < bytes);
3998         data_sinfo->bytes_may_use -= bytes;
3999         trace_btrfs_space_reservation(root->fs_info, "space_info",
4000                                       data_sinfo->flags, bytes, 0);
4001         spin_unlock(&data_sinfo->lock);
4002 }
4003
4004 static void force_metadata_allocation(struct btrfs_fs_info *info)
4005 {
4006         struct list_head *head = &info->space_info;
4007         struct btrfs_space_info *found;
4008
4009         rcu_read_lock();
4010         list_for_each_entry_rcu(found, head, list) {
4011                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4012                         found->force_alloc = CHUNK_ALLOC_FORCE;
4013         }
4014         rcu_read_unlock();
4015 }
4016
4017 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4018 {
4019         return (global->size << 1);
4020 }
4021
4022 static int should_alloc_chunk(struct btrfs_root *root,
4023                               struct btrfs_space_info *sinfo, int force)
4024 {
4025         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4026         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4027         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4028         u64 thresh;
4029
4030         if (force == CHUNK_ALLOC_FORCE)
4031                 return 1;
4032
4033         /*
4034          * We need to take into account the global rsv because for all intents
4035          * and purposes it's used space.  Don't worry about locking the
4036          * global_rsv, it doesn't change except when the transaction commits.
4037          */
4038         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4039                 num_allocated += calc_global_rsv_need_space(global_rsv);
4040
4041         /*
4042          * in limited mode, we want to have some free space up to
4043          * about 1% of the FS size.
4044          */
4045         if (force == CHUNK_ALLOC_LIMITED) {
4046                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4047                 thresh = max_t(u64, 64 * 1024 * 1024,
4048                                div_factor_fine(thresh, 1));
4049
4050                 if (num_bytes - num_allocated < thresh)
4051                         return 1;
4052         }
4053
4054         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
4055                 return 0;
4056         return 1;
4057 }
4058
4059 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
4060 {
4061         u64 num_dev;
4062
4063         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4064                     BTRFS_BLOCK_GROUP_RAID0 |
4065                     BTRFS_BLOCK_GROUP_RAID5 |
4066                     BTRFS_BLOCK_GROUP_RAID6))
4067                 num_dev = root->fs_info->fs_devices->rw_devices;
4068         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4069                 num_dev = 2;
4070         else
4071                 num_dev = 1;    /* DUP or single */
4072
4073         /* metadata for updaing devices and chunk tree */
4074         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
4075 }
4076
4077 static void check_system_chunk(struct btrfs_trans_handle *trans,
4078                                struct btrfs_root *root, u64 type)
4079 {
4080         struct btrfs_space_info *info;
4081         u64 left;
4082         u64 thresh;
4083
4084         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4085         spin_lock(&info->lock);
4086         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4087                 info->bytes_reserved - info->bytes_readonly;
4088         spin_unlock(&info->lock);
4089
4090         thresh = get_system_chunk_thresh(root, type);
4091         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4092                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4093                         left, thresh, type);
4094                 dump_space_info(info, 0, 0);
4095         }
4096
4097         if (left < thresh) {
4098                 u64 flags;
4099
4100                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4101                 btrfs_alloc_chunk(trans, root, flags);
4102         }
4103 }
4104
4105 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4106                           struct btrfs_root *extent_root, u64 flags, int force)
4107 {
4108         struct btrfs_space_info *space_info;
4109         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4110         int wait_for_alloc = 0;
4111         int ret = 0;
4112
4113         /* Don't re-enter if we're already allocating a chunk */
4114         if (trans->allocating_chunk)
4115                 return -ENOSPC;
4116
4117         space_info = __find_space_info(extent_root->fs_info, flags);
4118         if (!space_info) {
4119                 ret = update_space_info(extent_root->fs_info, flags,
4120                                         0, 0, &space_info);
4121                 BUG_ON(ret); /* -ENOMEM */
4122         }
4123         BUG_ON(!space_info); /* Logic error */
4124
4125 again:
4126         spin_lock(&space_info->lock);
4127         if (force < space_info->force_alloc)
4128                 force = space_info->force_alloc;
4129         if (space_info->full) {
4130                 if (should_alloc_chunk(extent_root, space_info, force))
4131                         ret = -ENOSPC;
4132                 else
4133                         ret = 0;
4134                 spin_unlock(&space_info->lock);
4135                 return ret;
4136         }
4137
4138         if (!should_alloc_chunk(extent_root, space_info, force)) {
4139                 spin_unlock(&space_info->lock);
4140                 return 0;
4141         } else if (space_info->chunk_alloc) {
4142                 wait_for_alloc = 1;
4143         } else {
4144                 space_info->chunk_alloc = 1;
4145         }
4146
4147         spin_unlock(&space_info->lock);
4148
4149         mutex_lock(&fs_info->chunk_mutex);
4150
4151         /*
4152          * The chunk_mutex is held throughout the entirety of a chunk
4153          * allocation, so once we've acquired the chunk_mutex we know that the
4154          * other guy is done and we need to recheck and see if we should
4155          * allocate.
4156          */
4157         if (wait_for_alloc) {
4158                 mutex_unlock(&fs_info->chunk_mutex);
4159                 wait_for_alloc = 0;
4160                 goto again;
4161         }
4162
4163         trans->allocating_chunk = true;
4164
4165         /*
4166          * If we have mixed data/metadata chunks we want to make sure we keep
4167          * allocating mixed chunks instead of individual chunks.
4168          */
4169         if (btrfs_mixed_space_info(space_info))
4170                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4171
4172         /*
4173          * if we're doing a data chunk, go ahead and make sure that
4174          * we keep a reasonable number of metadata chunks allocated in the
4175          * FS as well.
4176          */
4177         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4178                 fs_info->data_chunk_allocations++;
4179                 if (!(fs_info->data_chunk_allocations %
4180                       fs_info->metadata_ratio))
4181                         force_metadata_allocation(fs_info);
4182         }
4183
4184         /*
4185          * Check if we have enough space in SYSTEM chunk because we may need
4186          * to update devices.
4187          */
4188         check_system_chunk(trans, extent_root, flags);
4189
4190         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4191         trans->allocating_chunk = false;
4192
4193         spin_lock(&space_info->lock);
4194         if (ret < 0 && ret != -ENOSPC)
4195                 goto out;
4196         if (ret)
4197                 space_info->full = 1;
4198         else
4199                 ret = 1;
4200
4201         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4202 out:
4203         space_info->chunk_alloc = 0;
4204         spin_unlock(&space_info->lock);
4205         mutex_unlock(&fs_info->chunk_mutex);
4206         return ret;
4207 }
4208
4209 static int can_overcommit(struct btrfs_root *root,
4210                           struct btrfs_space_info *space_info, u64 bytes,
4211                           enum btrfs_reserve_flush_enum flush)
4212 {
4213         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4214         u64 profile = btrfs_get_alloc_profile(root, 0);
4215         u64 space_size;
4216         u64 avail;
4217         u64 used;
4218
4219         used = space_info->bytes_used + space_info->bytes_reserved +
4220                 space_info->bytes_pinned + space_info->bytes_readonly;
4221
4222         /*
4223          * We only want to allow over committing if we have lots of actual space
4224          * free, but if we don't have enough space to handle the global reserve
4225          * space then we could end up having a real enospc problem when trying
4226          * to allocate a chunk or some other such important allocation.
4227          */
4228         spin_lock(&global_rsv->lock);
4229         space_size = calc_global_rsv_need_space(global_rsv);
4230         spin_unlock(&global_rsv->lock);
4231         if (used + space_size >= space_info->total_bytes)
4232                 return 0;
4233
4234         used += space_info->bytes_may_use;
4235
4236         spin_lock(&root->fs_info->free_chunk_lock);
4237         avail = root->fs_info->free_chunk_space;
4238         spin_unlock(&root->fs_info->free_chunk_lock);
4239
4240         /*
4241          * If we have dup, raid1 or raid10 then only half of the free
4242          * space is actually useable.  For raid56, the space info used
4243          * doesn't include the parity drive, so we don't have to
4244          * change the math
4245          */
4246         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4247                        BTRFS_BLOCK_GROUP_RAID1 |
4248                        BTRFS_BLOCK_GROUP_RAID10))
4249                 avail >>= 1;
4250
4251         /*
4252          * If we aren't flushing all things, let us overcommit up to
4253          * 1/2th of the space. If we can flush, don't let us overcommit
4254          * too much, let it overcommit up to 1/8 of the space.
4255          */
4256         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4257                 avail >>= 3;
4258         else
4259                 avail >>= 1;
4260
4261         if (used + bytes < space_info->total_bytes + avail)
4262                 return 1;
4263         return 0;
4264 }
4265
4266 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4267                                          unsigned long nr_pages, int nr_items)
4268 {
4269         struct super_block *sb = root->fs_info->sb;
4270
4271         if (down_read_trylock(&sb->s_umount)) {
4272                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4273                 up_read(&sb->s_umount);
4274         } else {
4275                 /*
4276                  * We needn't worry the filesystem going from r/w to r/o though
4277                  * we don't acquire ->s_umount mutex, because the filesystem
4278                  * should guarantee the delalloc inodes list be empty after
4279                  * the filesystem is readonly(all dirty pages are written to
4280                  * the disk).
4281                  */
4282                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4283                 if (!current->journal_info)
4284                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4285         }
4286 }
4287
4288 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4289 {
4290         u64 bytes;
4291         int nr;
4292
4293         bytes = btrfs_calc_trans_metadata_size(root, 1);
4294         nr = (int)div64_u64(to_reclaim, bytes);
4295         if (!nr)
4296                 nr = 1;
4297         return nr;
4298 }
4299
4300 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4301
4302 /*
4303  * shrink metadata reservation for delalloc
4304  */
4305 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4306                             bool wait_ordered)
4307 {
4308         struct btrfs_block_rsv *block_rsv;
4309         struct btrfs_space_info *space_info;
4310         struct btrfs_trans_handle *trans;
4311         u64 delalloc_bytes;
4312         u64 max_reclaim;
4313         long time_left;
4314         unsigned long nr_pages;
4315         int loops;
4316         int items;
4317         enum btrfs_reserve_flush_enum flush;
4318
4319         /* Calc the number of the pages we need flush for space reservation */
4320         items = calc_reclaim_items_nr(root, to_reclaim);
4321         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4322
4323         trans = (struct btrfs_trans_handle *)current->journal_info;
4324         block_rsv = &root->fs_info->delalloc_block_rsv;
4325         space_info = block_rsv->space_info;
4326
4327         delalloc_bytes = percpu_counter_sum_positive(
4328                                                 &root->fs_info->delalloc_bytes);
4329         if (delalloc_bytes == 0) {
4330                 if (trans)
4331                         return;
4332                 if (wait_ordered)
4333                         btrfs_wait_ordered_roots(root->fs_info, items);
4334                 return;
4335         }
4336
4337         loops = 0;
4338         while (delalloc_bytes && loops < 3) {
4339                 max_reclaim = min(delalloc_bytes, to_reclaim);
4340                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4341                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4342                 /*
4343                  * We need to wait for the async pages to actually start before
4344                  * we do anything.
4345                  */
4346                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4347                 if (!max_reclaim)
4348                         goto skip_async;
4349
4350                 if (max_reclaim <= nr_pages)
4351                         max_reclaim = 0;
4352                 else
4353                         max_reclaim -= nr_pages;
4354
4355                 wait_event(root->fs_info->async_submit_wait,
4356                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4357                            (int)max_reclaim);
4358 skip_async:
4359                 if (!trans)
4360                         flush = BTRFS_RESERVE_FLUSH_ALL;
4361                 else
4362                         flush = BTRFS_RESERVE_NO_FLUSH;
4363                 spin_lock(&space_info->lock);
4364                 if (can_overcommit(root, space_info, orig, flush)) {
4365                         spin_unlock(&space_info->lock);
4366                         break;
4367                 }
4368                 spin_unlock(&space_info->lock);
4369
4370                 loops++;
4371                 if (wait_ordered && !trans) {
4372                         btrfs_wait_ordered_roots(root->fs_info, items);
4373                 } else {
4374                         time_left = schedule_timeout_killable(1);
4375                         if (time_left)
4376                                 break;
4377                 }
4378                 delalloc_bytes = percpu_counter_sum_positive(
4379                                                 &root->fs_info->delalloc_bytes);
4380         }
4381 }
4382
4383 /**
4384  * maybe_commit_transaction - possibly commit the transaction if its ok to
4385  * @root - the root we're allocating for
4386  * @bytes - the number of bytes we want to reserve
4387  * @force - force the commit
4388  *
4389  * This will check to make sure that committing the transaction will actually
4390  * get us somewhere and then commit the transaction if it does.  Otherwise it
4391  * will return -ENOSPC.
4392  */
4393 static int may_commit_transaction(struct btrfs_root *root,
4394                                   struct btrfs_space_info *space_info,
4395                                   u64 bytes, int force)
4396 {
4397         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4398         struct btrfs_trans_handle *trans;
4399
4400         trans = (struct btrfs_trans_handle *)current->journal_info;
4401         if (trans)
4402                 return -EAGAIN;
4403
4404         if (force)
4405                 goto commit;
4406
4407         /* See if there is enough pinned space to make this reservation */
4408         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4409                                    bytes) >= 0)
4410                 goto commit;
4411
4412         /*
4413          * See if there is some space in the delayed insertion reservation for
4414          * this reservation.
4415          */
4416         if (space_info != delayed_rsv->space_info)
4417                 return -ENOSPC;
4418
4419         spin_lock(&delayed_rsv->lock);
4420         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4421                                    bytes - delayed_rsv->size) >= 0) {
4422                 spin_unlock(&delayed_rsv->lock);
4423                 return -ENOSPC;
4424         }
4425         spin_unlock(&delayed_rsv->lock);
4426
4427 commit:
4428         trans = btrfs_join_transaction(root);
4429         if (IS_ERR(trans))
4430                 return -ENOSPC;
4431
4432         return btrfs_commit_transaction(trans, root);
4433 }
4434
4435 enum flush_state {
4436         FLUSH_DELAYED_ITEMS_NR  =       1,
4437         FLUSH_DELAYED_ITEMS     =       2,
4438         FLUSH_DELALLOC          =       3,
4439         FLUSH_DELALLOC_WAIT     =       4,
4440         ALLOC_CHUNK             =       5,
4441         COMMIT_TRANS            =       6,
4442 };
4443
4444 static int flush_space(struct btrfs_root *root,
4445                        struct btrfs_space_info *space_info, u64 num_bytes,
4446                        u64 orig_bytes, int state)
4447 {
4448         struct btrfs_trans_handle *trans;
4449         int nr;
4450         int ret = 0;
4451
4452         switch (state) {
4453         case FLUSH_DELAYED_ITEMS_NR:
4454         case FLUSH_DELAYED_ITEMS:
4455                 if (state == FLUSH_DELAYED_ITEMS_NR)
4456                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4457                 else
4458                         nr = -1;
4459
4460                 trans = btrfs_join_transaction(root);
4461                 if (IS_ERR(trans)) {
4462                         ret = PTR_ERR(trans);
4463                         break;
4464                 }
4465                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4466                 btrfs_end_transaction(trans, root);
4467                 break;
4468         case FLUSH_DELALLOC:
4469         case FLUSH_DELALLOC_WAIT:
4470                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4471                                 state == FLUSH_DELALLOC_WAIT);
4472                 break;
4473         case ALLOC_CHUNK:
4474                 trans = btrfs_join_transaction(root);
4475                 if (IS_ERR(trans)) {
4476                         ret = PTR_ERR(trans);
4477                         break;
4478                 }
4479                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4480                                      btrfs_get_alloc_profile(root, 0),
4481                                      CHUNK_ALLOC_NO_FORCE);
4482                 btrfs_end_transaction(trans, root);
4483                 if (ret == -ENOSPC)
4484                         ret = 0;
4485                 break;
4486         case COMMIT_TRANS:
4487                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4488                 break;
4489         default:
4490                 ret = -ENOSPC;
4491                 break;
4492         }
4493
4494         return ret;
4495 }
4496
4497 static inline u64
4498 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4499                                  struct btrfs_space_info *space_info)
4500 {
4501         u64 used;
4502         u64 expected;
4503         u64 to_reclaim;
4504
4505         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4506                                 16 * 1024 * 1024);
4507         spin_lock(&space_info->lock);
4508         if (can_overcommit(root, space_info, to_reclaim,
4509                            BTRFS_RESERVE_FLUSH_ALL)) {
4510                 to_reclaim = 0;
4511                 goto out;
4512         }
4513
4514         used = space_info->bytes_used + space_info->bytes_reserved +
4515                space_info->bytes_pinned + space_info->bytes_readonly +
4516                space_info->bytes_may_use;
4517         if (can_overcommit(root, space_info, 1024 * 1024,
4518                            BTRFS_RESERVE_FLUSH_ALL))
4519                 expected = div_factor_fine(space_info->total_bytes, 95);
4520         else
4521                 expected = div_factor_fine(space_info->total_bytes, 90);
4522
4523         if (used > expected)
4524                 to_reclaim = used - expected;
4525         else
4526                 to_reclaim = 0;
4527         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4528                                      space_info->bytes_reserved);
4529 out:
4530         spin_unlock(&space_info->lock);
4531
4532         return to_reclaim;
4533 }
4534
4535 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4536                                         struct btrfs_fs_info *fs_info, u64 used)
4537 {
4538         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4539
4540         /* If we're just plain full then async reclaim just slows us down. */
4541         if (space_info->bytes_used >= thresh)
4542                 return 0;
4543
4544         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4545                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4546 }
4547
4548 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4549                                        struct btrfs_fs_info *fs_info,
4550                                        int flush_state)
4551 {
4552         u64 used;
4553
4554         spin_lock(&space_info->lock);
4555         /*
4556          * We run out of space and have not got any free space via flush_space,
4557          * so don't bother doing async reclaim.
4558          */
4559         if (flush_state > COMMIT_TRANS && space_info->full) {
4560                 spin_unlock(&space_info->lock);
4561                 return 0;
4562         }
4563
4564         used = space_info->bytes_used + space_info->bytes_reserved +
4565                space_info->bytes_pinned + space_info->bytes_readonly +
4566                space_info->bytes_may_use;
4567         if (need_do_async_reclaim(space_info, fs_info, used)) {
4568                 spin_unlock(&space_info->lock);
4569                 return 1;
4570         }
4571         spin_unlock(&space_info->lock);
4572
4573         return 0;
4574 }
4575
4576 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4577 {
4578         struct btrfs_fs_info *fs_info;
4579         struct btrfs_space_info *space_info;
4580         u64 to_reclaim;
4581         int flush_state;
4582
4583         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4584         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4585
4586         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4587                                                       space_info);
4588         if (!to_reclaim)
4589                 return;
4590
4591         flush_state = FLUSH_DELAYED_ITEMS_NR;
4592         do {
4593                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4594                             to_reclaim, flush_state);
4595                 flush_state++;
4596                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4597                                                  flush_state))
4598                         return;
4599         } while (flush_state < COMMIT_TRANS);
4600 }
4601
4602 void btrfs_init_async_reclaim_work(struct work_struct *work)
4603 {
4604         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4605 }
4606
4607 /**
4608  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4609  * @root - the root we're allocating for
4610  * @block_rsv - the block_rsv we're allocating for
4611  * @orig_bytes - the number of bytes we want
4612  * @flush - whether or not we can flush to make our reservation
4613  *
4614  * This will reserve orgi_bytes number of bytes from the space info associated
4615  * with the block_rsv.  If there is not enough space it will make an attempt to
4616  * flush out space to make room.  It will do this by flushing delalloc if
4617  * possible or committing the transaction.  If flush is 0 then no attempts to
4618  * regain reservations will be made and this will fail if there is not enough
4619  * space already.
4620  */
4621 static int reserve_metadata_bytes(struct btrfs_root *root,
4622                                   struct btrfs_block_rsv *block_rsv,
4623                                   u64 orig_bytes,
4624                                   enum btrfs_reserve_flush_enum flush)
4625 {
4626         struct btrfs_space_info *space_info = block_rsv->space_info;
4627         u64 used;
4628         u64 num_bytes = orig_bytes;
4629         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4630         int ret = 0;
4631         bool flushing = false;
4632
4633 again:
4634         ret = 0;
4635         spin_lock(&space_info->lock);
4636         /*
4637          * We only want to wait if somebody other than us is flushing and we
4638          * are actually allowed to flush all things.
4639          */
4640         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4641                space_info->flush) {
4642                 spin_unlock(&space_info->lock);
4643                 /*
4644                  * If we have a trans handle we can't wait because the flusher
4645                  * may have to commit the transaction, which would mean we would
4646                  * deadlock since we are waiting for the flusher to finish, but
4647                  * hold the current transaction open.
4648                  */
4649                 if (current->journal_info)
4650                         return -EAGAIN;
4651                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4652                 /* Must have been killed, return */
4653                 if (ret)
4654                         return -EINTR;
4655
4656                 spin_lock(&space_info->lock);
4657         }
4658
4659         ret = -ENOSPC;
4660         used = space_info->bytes_used + space_info->bytes_reserved +
4661                 space_info->bytes_pinned + space_info->bytes_readonly +
4662                 space_info->bytes_may_use;
4663
4664         /*
4665          * The idea here is that we've not already over-reserved the block group
4666          * then we can go ahead and save our reservation first and then start
4667          * flushing if we need to.  Otherwise if we've already overcommitted
4668          * lets start flushing stuff first and then come back and try to make
4669          * our reservation.
4670          */
4671         if (used <= space_info->total_bytes) {
4672                 if (used + orig_bytes <= space_info->total_bytes) {
4673                         space_info->bytes_may_use += orig_bytes;
4674                         trace_btrfs_space_reservation(root->fs_info,
4675                                 "space_info", space_info->flags, orig_bytes, 1);
4676                         ret = 0;
4677                 } else {
4678                         /*
4679                          * Ok set num_bytes to orig_bytes since we aren't
4680                          * overocmmitted, this way we only try and reclaim what
4681                          * we need.
4682                          */
4683                         num_bytes = orig_bytes;
4684                 }
4685         } else {
4686                 /*
4687                  * Ok we're over committed, set num_bytes to the overcommitted
4688                  * amount plus the amount of bytes that we need for this
4689                  * reservation.
4690                  */
4691                 num_bytes = used - space_info->total_bytes +
4692                         (orig_bytes * 2);
4693         }
4694
4695         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4696                 space_info->bytes_may_use += orig_bytes;
4697                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4698                                               space_info->flags, orig_bytes,
4699                                               1);
4700                 ret = 0;
4701         }
4702
4703         /*
4704          * Couldn't make our reservation, save our place so while we're trying
4705          * to reclaim space we can actually use it instead of somebody else
4706          * stealing it from us.
4707          *
4708          * We make the other tasks wait for the flush only when we can flush
4709          * all things.
4710          */
4711         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4712                 flushing = true;
4713                 space_info->flush = 1;
4714         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4715                 used += orig_bytes;
4716                 /*
4717                  * We will do the space reservation dance during log replay,
4718                  * which means we won't have fs_info->fs_root set, so don't do
4719                  * the async reclaim as we will panic.
4720                  */
4721                 if (!root->fs_info->log_root_recovering &&
4722                     need_do_async_reclaim(space_info, root->fs_info, used) &&
4723                     !work_busy(&root->fs_info->async_reclaim_work))
4724                         queue_work(system_unbound_wq,
4725                                    &root->fs_info->async_reclaim_work);
4726         }
4727         spin_unlock(&space_info->lock);
4728
4729         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4730                 goto out;
4731
4732         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4733                           flush_state);
4734         flush_state++;
4735
4736         /*
4737          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4738          * would happen. So skip delalloc flush.
4739          */
4740         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4741             (flush_state == FLUSH_DELALLOC ||
4742              flush_state == FLUSH_DELALLOC_WAIT))
4743                 flush_state = ALLOC_CHUNK;
4744
4745         if (!ret)
4746                 goto again;
4747         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4748                  flush_state < COMMIT_TRANS)
4749                 goto again;
4750         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4751                  flush_state <= COMMIT_TRANS)
4752                 goto again;
4753
4754 out:
4755         if (ret == -ENOSPC &&
4756             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4757                 struct btrfs_block_rsv *global_rsv =
4758                         &root->fs_info->global_block_rsv;
4759
4760                 if (block_rsv != global_rsv &&
4761                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4762                         ret = 0;
4763         }
4764         if (ret == -ENOSPC)
4765                 trace_btrfs_space_reservation(root->fs_info,
4766                                               "space_info:enospc",
4767                                               space_info->flags, orig_bytes, 1);
4768         if (flushing) {
4769                 spin_lock(&space_info->lock);
4770                 space_info->flush = 0;
4771                 wake_up_all(&space_info->wait);
4772                 spin_unlock(&space_info->lock);
4773         }
4774         return ret;
4775 }
4776
4777 static struct btrfs_block_rsv *get_block_rsv(
4778                                         const struct btrfs_trans_handle *trans,
4779                                         const struct btrfs_root *root)
4780 {
4781         struct btrfs_block_rsv *block_rsv = NULL;
4782
4783         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4784                 block_rsv = trans->block_rsv;
4785
4786         if (root == root->fs_info->csum_root && trans->adding_csums)
4787                 block_rsv = trans->block_rsv;
4788
4789         if (root == root->fs_info->uuid_root)
4790                 block_rsv = trans->block_rsv;
4791
4792         if (!block_rsv)
4793                 block_rsv = root->block_rsv;
4794
4795         if (!block_rsv)
4796                 block_rsv = &root->fs_info->empty_block_rsv;
4797
4798         return block_rsv;
4799 }
4800
4801 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4802                                u64 num_bytes)
4803 {
4804         int ret = -ENOSPC;
4805         spin_lock(&block_rsv->lock);
4806         if (block_rsv->reserved >= num_bytes) {
4807                 block_rsv->reserved -= num_bytes;
4808                 if (block_rsv->reserved < block_rsv->size)
4809                         block_rsv->full = 0;
4810                 ret = 0;
4811         }
4812         spin_unlock(&block_rsv->lock);
4813         return ret;
4814 }
4815
4816 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4817                                 u64 num_bytes, int update_size)
4818 {
4819         spin_lock(&block_rsv->lock);
4820         block_rsv->reserved += num_bytes;
4821         if (update_size)
4822                 block_rsv->size += num_bytes;
4823         else if (block_rsv->reserved >= block_rsv->size)
4824                 block_rsv->full = 1;
4825         spin_unlock(&block_rsv->lock);
4826 }
4827
4828 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4829                              struct btrfs_block_rsv *dest, u64 num_bytes,
4830                              int min_factor)
4831 {
4832         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4833         u64 min_bytes;
4834
4835         if (global_rsv->space_info != dest->space_info)
4836                 return -ENOSPC;
4837
4838         spin_lock(&global_rsv->lock);
4839         min_bytes = div_factor(global_rsv->size, min_factor);
4840         if (global_rsv->reserved < min_bytes + num_bytes) {
4841                 spin_unlock(&global_rsv->lock);
4842                 return -ENOSPC;
4843         }
4844         global_rsv->reserved -= num_bytes;
4845         if (global_rsv->reserved < global_rsv->size)
4846                 global_rsv->full = 0;
4847         spin_unlock(&global_rsv->lock);
4848
4849         block_rsv_add_bytes(dest, num_bytes, 1);
4850         return 0;
4851 }
4852
4853 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4854                                     struct btrfs_block_rsv *block_rsv,
4855                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4856 {
4857         struct btrfs_space_info *space_info = block_rsv->space_info;
4858
4859         spin_lock(&block_rsv->lock);
4860         if (num_bytes == (u64)-1)
4861                 num_bytes = block_rsv->size;
4862         block_rsv->size -= num_bytes;
4863         if (block_rsv->reserved >= block_rsv->size) {
4864                 num_bytes = block_rsv->reserved - block_rsv->size;
4865                 block_rsv->reserved = block_rsv->size;
4866                 block_rsv->full = 1;
4867         } else {
4868                 num_bytes = 0;
4869         }
4870         spin_unlock(&block_rsv->lock);
4871
4872         if (num_bytes > 0) {
4873                 if (dest) {
4874                         spin_lock(&dest->lock);
4875                         if (!dest->full) {
4876                                 u64 bytes_to_add;
4877
4878                                 bytes_to_add = dest->size - dest->reserved;
4879                                 bytes_to_add = min(num_bytes, bytes_to_add);
4880                                 dest->reserved += bytes_to_add;
4881                                 if (dest->reserved >= dest->size)
4882                                         dest->full = 1;
4883                                 num_bytes -= bytes_to_add;
4884                         }
4885                         spin_unlock(&dest->lock);
4886                 }
4887                 if (num_bytes) {
4888                         spin_lock(&space_info->lock);
4889                         space_info->bytes_may_use -= num_bytes;
4890                         trace_btrfs_space_reservation(fs_info, "space_info",
4891                                         space_info->flags, num_bytes, 0);
4892                         spin_unlock(&space_info->lock);
4893                 }
4894         }
4895 }
4896
4897 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4898                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4899 {
4900         int ret;
4901
4902         ret = block_rsv_use_bytes(src, num_bytes);
4903         if (ret)
4904                 return ret;
4905
4906         block_rsv_add_bytes(dst, num_bytes, 1);
4907         return 0;
4908 }
4909
4910 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4911 {
4912         memset(rsv, 0, sizeof(*rsv));
4913         spin_lock_init(&rsv->lock);
4914         rsv->type = type;
4915 }
4916
4917 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4918                                               unsigned short type)
4919 {
4920         struct btrfs_block_rsv *block_rsv;
4921         struct btrfs_fs_info *fs_info = root->fs_info;
4922
4923         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4924         if (!block_rsv)
4925                 return NULL;
4926
4927         btrfs_init_block_rsv(block_rsv, type);
4928         block_rsv->space_info = __find_space_info(fs_info,
4929                                                   BTRFS_BLOCK_GROUP_METADATA);
4930         return block_rsv;
4931 }
4932
4933 void btrfs_free_block_rsv(struct btrfs_root *root,
4934                           struct btrfs_block_rsv *rsv)
4935 {
4936         if (!rsv)
4937                 return;
4938         btrfs_block_rsv_release(root, rsv, (u64)-1);
4939         kfree(rsv);
4940 }
4941
4942 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
4943 {
4944         kfree(rsv);
4945 }
4946
4947 int btrfs_block_rsv_add(struct btrfs_root *root,
4948                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4949                         enum btrfs_reserve_flush_enum flush)
4950 {
4951         int ret;
4952
4953         if (num_bytes == 0)
4954                 return 0;
4955
4956         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4957         if (!ret) {
4958                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4959                 return 0;
4960         }
4961
4962         return ret;
4963 }
4964
4965 int btrfs_block_rsv_check(struct btrfs_root *root,
4966                           struct btrfs_block_rsv *block_rsv, int min_factor)
4967 {
4968         u64 num_bytes = 0;
4969         int ret = -ENOSPC;
4970
4971         if (!block_rsv)
4972                 return 0;
4973
4974         spin_lock(&block_rsv->lock);
4975         num_bytes = div_factor(block_rsv->size, min_factor);
4976         if (block_rsv->reserved >= num_bytes)
4977                 ret = 0;
4978         spin_unlock(&block_rsv->lock);
4979
4980         return ret;
4981 }
4982
4983 int btrfs_block_rsv_refill(struct btrfs_root *root,
4984                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4985                            enum btrfs_reserve_flush_enum flush)
4986 {
4987         u64 num_bytes = 0;
4988         int ret = -ENOSPC;
4989
4990         if (!block_rsv)
4991                 return 0;
4992
4993         spin_lock(&block_rsv->lock);
4994         num_bytes = min_reserved;
4995         if (block_rsv->reserved >= num_bytes)
4996                 ret = 0;
4997         else
4998                 num_bytes -= block_rsv->reserved;
4999         spin_unlock(&block_rsv->lock);
5000
5001         if (!ret)
5002                 return 0;
5003
5004         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5005         if (!ret) {
5006                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5007                 return 0;
5008         }
5009
5010         return ret;
5011 }
5012
5013 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5014                             struct btrfs_block_rsv *dst_rsv,
5015                             u64 num_bytes)
5016 {
5017         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5018 }
5019
5020 void btrfs_block_rsv_release(struct btrfs_root *root,
5021                              struct btrfs_block_rsv *block_rsv,
5022                              u64 num_bytes)
5023 {
5024         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5025         if (global_rsv == block_rsv ||
5026             block_rsv->space_info != global_rsv->space_info)
5027                 global_rsv = NULL;
5028         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5029                                 num_bytes);
5030 }
5031
5032 /*
5033  * helper to calculate size of global block reservation.
5034  * the desired value is sum of space used by extent tree,
5035  * checksum tree and root tree
5036  */
5037 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5038 {
5039         struct btrfs_space_info *sinfo;
5040         u64 num_bytes;
5041         u64 meta_used;
5042         u64 data_used;
5043         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5044
5045         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5046         spin_lock(&sinfo->lock);
5047         data_used = sinfo->bytes_used;
5048         spin_unlock(&sinfo->lock);
5049
5050         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5051         spin_lock(&sinfo->lock);
5052         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5053                 data_used = 0;
5054         meta_used = sinfo->bytes_used;
5055         spin_unlock(&sinfo->lock);
5056
5057         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5058                     csum_size * 2;
5059         num_bytes += div_u64(data_used + meta_used, 50);
5060
5061         if (num_bytes * 3 > meta_used)
5062                 num_bytes = div_u64(meta_used, 3);
5063
5064         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5065 }
5066
5067 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5068 {
5069         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5070         struct btrfs_space_info *sinfo = block_rsv->space_info;
5071         u64 num_bytes;
5072
5073         num_bytes = calc_global_metadata_size(fs_info);
5074
5075         spin_lock(&sinfo->lock);
5076         spin_lock(&block_rsv->lock);
5077
5078         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
5079
5080         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5081                     sinfo->bytes_reserved + sinfo->bytes_readonly +
5082                     sinfo->bytes_may_use;
5083
5084         if (sinfo->total_bytes > num_bytes) {
5085                 num_bytes = sinfo->total_bytes - num_bytes;
5086                 block_rsv->reserved += num_bytes;
5087                 sinfo->bytes_may_use += num_bytes;
5088                 trace_btrfs_space_reservation(fs_info, "space_info",
5089                                       sinfo->flags, num_bytes, 1);
5090         }
5091
5092         if (block_rsv->reserved >= block_rsv->size) {
5093                 num_bytes = block_rsv->reserved - block_rsv->size;
5094                 sinfo->bytes_may_use -= num_bytes;
5095                 trace_btrfs_space_reservation(fs_info, "space_info",
5096                                       sinfo->flags, num_bytes, 0);
5097                 block_rsv->reserved = block_rsv->size;
5098                 block_rsv->full = 1;
5099         }
5100
5101         spin_unlock(&block_rsv->lock);
5102         spin_unlock(&sinfo->lock);
5103 }
5104
5105 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5106 {
5107         struct btrfs_space_info *space_info;
5108
5109         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5110         fs_info->chunk_block_rsv.space_info = space_info;
5111
5112         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5113         fs_info->global_block_rsv.space_info = space_info;
5114         fs_info->delalloc_block_rsv.space_info = space_info;
5115         fs_info->trans_block_rsv.space_info = space_info;
5116         fs_info->empty_block_rsv.space_info = space_info;
5117         fs_info->delayed_block_rsv.space_info = space_info;
5118
5119         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5120         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5121         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5122         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5123         if (fs_info->quota_root)
5124                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5125         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5126
5127         update_global_block_rsv(fs_info);
5128 }
5129
5130 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5131 {
5132         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5133                                 (u64)-1);
5134         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5135         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5136         WARN_ON(fs_info->trans_block_rsv.size > 0);
5137         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5138         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5139         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5140         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5141         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5142 }
5143
5144 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5145                                   struct btrfs_root *root)
5146 {
5147         if (!trans->block_rsv)
5148                 return;
5149
5150         if (!trans->bytes_reserved)
5151                 return;
5152
5153         trace_btrfs_space_reservation(root->fs_info, "transaction",
5154                                       trans->transid, trans->bytes_reserved, 0);
5155         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5156         trans->bytes_reserved = 0;
5157 }
5158
5159 /* Can only return 0 or -ENOSPC */
5160 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5161                                   struct inode *inode)
5162 {
5163         struct btrfs_root *root = BTRFS_I(inode)->root;
5164         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5165         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5166
5167         /*
5168          * We need to hold space in order to delete our orphan item once we've
5169          * added it, so this takes the reservation so we can release it later
5170          * when we are truly done with the orphan item.
5171          */
5172         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5173         trace_btrfs_space_reservation(root->fs_info, "orphan",
5174                                       btrfs_ino(inode), num_bytes, 1);
5175         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5176 }
5177
5178 void btrfs_orphan_release_metadata(struct inode *inode)
5179 {
5180         struct btrfs_root *root = BTRFS_I(inode)->root;
5181         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5182         trace_btrfs_space_reservation(root->fs_info, "orphan",
5183                                       btrfs_ino(inode), num_bytes, 0);
5184         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5185 }
5186
5187 /*
5188  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5189  * root: the root of the parent directory
5190  * rsv: block reservation
5191  * items: the number of items that we need do reservation
5192  * qgroup_reserved: used to return the reserved size in qgroup
5193  *
5194  * This function is used to reserve the space for snapshot/subvolume
5195  * creation and deletion. Those operations are different with the
5196  * common file/directory operations, they change two fs/file trees
5197  * and root tree, the number of items that the qgroup reserves is
5198  * different with the free space reservation. So we can not use
5199  * the space reseravtion mechanism in start_transaction().
5200  */
5201 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5202                                      struct btrfs_block_rsv *rsv,
5203                                      int items,
5204                                      u64 *qgroup_reserved,
5205                                      bool use_global_rsv)
5206 {
5207         u64 num_bytes;
5208         int ret;
5209         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5210
5211         if (root->fs_info->quota_enabled) {
5212                 /* One for parent inode, two for dir entries */
5213                 num_bytes = 3 * root->nodesize;
5214                 ret = btrfs_qgroup_reserve(root, num_bytes);
5215                 if (ret)
5216                         return ret;
5217         } else {
5218                 num_bytes = 0;
5219         }
5220
5221         *qgroup_reserved = num_bytes;
5222
5223         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5224         rsv->space_info = __find_space_info(root->fs_info,
5225                                             BTRFS_BLOCK_GROUP_METADATA);
5226         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5227                                   BTRFS_RESERVE_FLUSH_ALL);
5228
5229         if (ret == -ENOSPC && use_global_rsv)
5230                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5231
5232         if (ret) {
5233                 if (*qgroup_reserved)
5234                         btrfs_qgroup_free(root, *qgroup_reserved);
5235         }
5236
5237         return ret;
5238 }
5239
5240 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5241                                       struct btrfs_block_rsv *rsv,
5242                                       u64 qgroup_reserved)
5243 {
5244         btrfs_block_rsv_release(root, rsv, (u64)-1);
5245 }
5246
5247 /**
5248  * drop_outstanding_extent - drop an outstanding extent
5249  * @inode: the inode we're dropping the extent for
5250  * @num_bytes: the number of bytes we're relaseing.
5251  *
5252  * This is called when we are freeing up an outstanding extent, either called
5253  * after an error or after an extent is written.  This will return the number of
5254  * reserved extents that need to be freed.  This must be called with
5255  * BTRFS_I(inode)->lock held.
5256  */
5257 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5258 {
5259         unsigned drop_inode_space = 0;
5260         unsigned dropped_extents = 0;
5261         unsigned num_extents = 0;
5262
5263         num_extents = (unsigned)div64_u64(num_bytes +
5264                                           BTRFS_MAX_EXTENT_SIZE - 1,
5265                                           BTRFS_MAX_EXTENT_SIZE);
5266         ASSERT(num_extents);
5267         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5268         BTRFS_I(inode)->outstanding_extents -= num_extents;
5269
5270         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5271             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5272                                &BTRFS_I(inode)->runtime_flags))
5273                 drop_inode_space = 1;
5274
5275         /*
5276          * If we have more or the same amount of outsanding extents than we have
5277          * reserved then we need to leave the reserved extents count alone.
5278          */
5279         if (BTRFS_I(inode)->outstanding_extents >=
5280             BTRFS_I(inode)->reserved_extents)
5281                 return drop_inode_space;
5282
5283         dropped_extents = BTRFS_I(inode)->reserved_extents -
5284                 BTRFS_I(inode)->outstanding_extents;
5285         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5286         return dropped_extents + drop_inode_space;
5287 }
5288
5289 /**
5290  * calc_csum_metadata_size - return the amount of metada space that must be
5291  *      reserved/free'd for the given bytes.
5292  * @inode: the inode we're manipulating
5293  * @num_bytes: the number of bytes in question
5294  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5295  *
5296  * This adjusts the number of csum_bytes in the inode and then returns the
5297  * correct amount of metadata that must either be reserved or freed.  We
5298  * calculate how many checksums we can fit into one leaf and then divide the
5299  * number of bytes that will need to be checksumed by this value to figure out
5300  * how many checksums will be required.  If we are adding bytes then the number
5301  * may go up and we will return the number of additional bytes that must be
5302  * reserved.  If it is going down we will return the number of bytes that must
5303  * be freed.
5304  *
5305  * This must be called with BTRFS_I(inode)->lock held.
5306  */
5307 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5308                                    int reserve)
5309 {
5310         struct btrfs_root *root = BTRFS_I(inode)->root;
5311         u64 old_csums, num_csums;
5312
5313         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5314             BTRFS_I(inode)->csum_bytes == 0)
5315                 return 0;
5316
5317         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5318         if (reserve)
5319                 BTRFS_I(inode)->csum_bytes += num_bytes;
5320         else
5321                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5322         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5323
5324         /* No change, no need to reserve more */
5325         if (old_csums == num_csums)
5326                 return 0;
5327
5328         if (reserve)
5329                 return btrfs_calc_trans_metadata_size(root,
5330                                                       num_csums - old_csums);
5331
5332         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5333 }
5334
5335 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5336 {
5337         struct btrfs_root *root = BTRFS_I(inode)->root;
5338         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5339         u64 to_reserve = 0;
5340         u64 csum_bytes;
5341         unsigned nr_extents = 0;
5342         int extra_reserve = 0;
5343         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5344         int ret = 0;
5345         bool delalloc_lock = true;
5346         u64 to_free = 0;
5347         unsigned dropped;
5348
5349         /* If we are a free space inode we need to not flush since we will be in
5350          * the middle of a transaction commit.  We also don't need the delalloc
5351          * mutex since we won't race with anybody.  We need this mostly to make
5352          * lockdep shut its filthy mouth.
5353          */
5354         if (btrfs_is_free_space_inode(inode)) {
5355                 flush = BTRFS_RESERVE_NO_FLUSH;
5356                 delalloc_lock = false;
5357         }
5358
5359         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5360             btrfs_transaction_in_commit(root->fs_info))
5361                 schedule_timeout(1);
5362
5363         if (delalloc_lock)
5364                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5365
5366         num_bytes = ALIGN(num_bytes, root->sectorsize);
5367
5368         spin_lock(&BTRFS_I(inode)->lock);
5369         nr_extents = (unsigned)div64_u64(num_bytes +
5370                                          BTRFS_MAX_EXTENT_SIZE - 1,
5371                                          BTRFS_MAX_EXTENT_SIZE);
5372         BTRFS_I(inode)->outstanding_extents += nr_extents;
5373         nr_extents = 0;
5374
5375         if (BTRFS_I(inode)->outstanding_extents >
5376             BTRFS_I(inode)->reserved_extents)
5377                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5378                         BTRFS_I(inode)->reserved_extents;
5379
5380         /*
5381          * Add an item to reserve for updating the inode when we complete the
5382          * delalloc io.
5383          */
5384         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5385                       &BTRFS_I(inode)->runtime_flags)) {
5386                 nr_extents++;
5387                 extra_reserve = 1;
5388         }
5389
5390         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5391         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5392         csum_bytes = BTRFS_I(inode)->csum_bytes;
5393         spin_unlock(&BTRFS_I(inode)->lock);
5394
5395         if (root->fs_info->quota_enabled) {
5396                 ret = btrfs_qgroup_reserve(root, nr_extents * root->nodesize);
5397                 if (ret)
5398                         goto out_fail;
5399         }
5400
5401         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5402         if (unlikely(ret)) {
5403                 if (root->fs_info->quota_enabled)
5404                         btrfs_qgroup_free(root, nr_extents * root->nodesize);
5405                 goto out_fail;
5406         }
5407
5408         spin_lock(&BTRFS_I(inode)->lock);
5409         if (extra_reserve) {
5410                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5411                         &BTRFS_I(inode)->runtime_flags);
5412                 nr_extents--;
5413         }
5414         BTRFS_I(inode)->reserved_extents += nr_extents;
5415         spin_unlock(&BTRFS_I(inode)->lock);
5416
5417         if (delalloc_lock)
5418                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5419
5420         if (to_reserve)
5421                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5422                                               btrfs_ino(inode), to_reserve, 1);
5423         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5424
5425         return 0;
5426
5427 out_fail:
5428         spin_lock(&BTRFS_I(inode)->lock);
5429         dropped = drop_outstanding_extent(inode, num_bytes);
5430         /*
5431          * If the inodes csum_bytes is the same as the original
5432          * csum_bytes then we know we haven't raced with any free()ers
5433          * so we can just reduce our inodes csum bytes and carry on.
5434          */
5435         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5436                 calc_csum_metadata_size(inode, num_bytes, 0);
5437         } else {
5438                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5439                 u64 bytes;
5440
5441                 /*
5442                  * This is tricky, but first we need to figure out how much we
5443                  * free'd from any free-ers that occured during this
5444                  * reservation, so we reset ->csum_bytes to the csum_bytes
5445                  * before we dropped our lock, and then call the free for the
5446                  * number of bytes that were freed while we were trying our
5447                  * reservation.
5448                  */
5449                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5450                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5451                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5452
5453
5454                 /*
5455                  * Now we need to see how much we would have freed had we not
5456                  * been making this reservation and our ->csum_bytes were not
5457                  * artificially inflated.
5458                  */
5459                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5460                 bytes = csum_bytes - orig_csum_bytes;
5461                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5462
5463                 /*
5464                  * Now reset ->csum_bytes to what it should be.  If bytes is
5465                  * more than to_free then we would have free'd more space had we
5466                  * not had an artificially high ->csum_bytes, so we need to free
5467                  * the remainder.  If bytes is the same or less then we don't
5468                  * need to do anything, the other free-ers did the correct
5469                  * thing.
5470                  */
5471                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5472                 if (bytes > to_free)
5473                         to_free = bytes - to_free;
5474                 else
5475                         to_free = 0;
5476         }
5477         spin_unlock(&BTRFS_I(inode)->lock);
5478         if (dropped)
5479                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5480
5481         if (to_free) {
5482                 btrfs_block_rsv_release(root, block_rsv, to_free);
5483                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5484                                               btrfs_ino(inode), to_free, 0);
5485         }
5486         if (delalloc_lock)
5487                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5488         return ret;
5489 }
5490
5491 /**
5492  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5493  * @inode: the inode to release the reservation for
5494  * @num_bytes: the number of bytes we're releasing
5495  *
5496  * This will release the metadata reservation for an inode.  This can be called
5497  * once we complete IO for a given set of bytes to release their metadata
5498  * reservations.
5499  */
5500 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5501 {
5502         struct btrfs_root *root = BTRFS_I(inode)->root;
5503         u64 to_free = 0;
5504         unsigned dropped;
5505
5506         num_bytes = ALIGN(num_bytes, root->sectorsize);
5507         spin_lock(&BTRFS_I(inode)->lock);
5508         dropped = drop_outstanding_extent(inode, num_bytes);
5509
5510         if (num_bytes)
5511                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5512         spin_unlock(&BTRFS_I(inode)->lock);
5513         if (dropped > 0)
5514                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5515
5516         if (btrfs_test_is_dummy_root(root))
5517                 return;
5518
5519         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5520                                       btrfs_ino(inode), to_free, 0);
5521
5522         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5523                                 to_free);
5524 }
5525
5526 /**
5527  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5528  * @inode: inode we're writing to
5529  * @num_bytes: the number of bytes we want to allocate
5530  *
5531  * This will do the following things
5532  *
5533  * o reserve space in the data space info for num_bytes
5534  * o reserve space in the metadata space info based on number of outstanding
5535  *   extents and how much csums will be needed
5536  * o add to the inodes ->delalloc_bytes
5537  * o add it to the fs_info's delalloc inodes list.
5538  *
5539  * This will return 0 for success and -ENOSPC if there is no space left.
5540  */
5541 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5542 {
5543         int ret;
5544
5545         ret = btrfs_check_data_free_space(inode, num_bytes, num_bytes);
5546         if (ret)
5547                 return ret;
5548
5549         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5550         if (ret) {
5551                 btrfs_free_reserved_data_space(inode, num_bytes);
5552                 return ret;
5553         }
5554
5555         return 0;
5556 }
5557
5558 /**
5559  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5560  * @inode: inode we're releasing space for
5561  * @num_bytes: the number of bytes we want to free up
5562  *
5563  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5564  * called in the case that we don't need the metadata AND data reservations
5565  * anymore.  So if there is an error or we insert an inline extent.
5566  *
5567  * This function will release the metadata space that was not used and will
5568  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5569  * list if there are no delalloc bytes left.
5570  */
5571 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5572 {
5573         btrfs_delalloc_release_metadata(inode, num_bytes);
5574         btrfs_free_reserved_data_space(inode, num_bytes);
5575 }
5576
5577 static int update_block_group(struct btrfs_trans_handle *trans,
5578                               struct btrfs_root *root, u64 bytenr,
5579                               u64 num_bytes, int alloc)
5580 {
5581         struct btrfs_block_group_cache *cache = NULL;
5582         struct btrfs_fs_info *info = root->fs_info;
5583         u64 total = num_bytes;
5584         u64 old_val;
5585         u64 byte_in_group;
5586         int factor;
5587
5588         /* block accounting for super block */
5589         spin_lock(&info->delalloc_root_lock);
5590         old_val = btrfs_super_bytes_used(info->super_copy);
5591         if (alloc)
5592                 old_val += num_bytes;
5593         else
5594                 old_val -= num_bytes;
5595         btrfs_set_super_bytes_used(info->super_copy, old_val);
5596         spin_unlock(&info->delalloc_root_lock);
5597
5598         while (total) {
5599                 cache = btrfs_lookup_block_group(info, bytenr);
5600                 if (!cache)
5601                         return -ENOENT;
5602                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5603                                     BTRFS_BLOCK_GROUP_RAID1 |
5604                                     BTRFS_BLOCK_GROUP_RAID10))
5605                         factor = 2;
5606                 else
5607                         factor = 1;
5608                 /*
5609                  * If this block group has free space cache written out, we
5610                  * need to make sure to load it if we are removing space.  This
5611                  * is because we need the unpinning stage to actually add the
5612                  * space back to the block group, otherwise we will leak space.
5613                  */
5614                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5615                         cache_block_group(cache, 1);
5616
5617                 byte_in_group = bytenr - cache->key.objectid;
5618                 WARN_ON(byte_in_group > cache->key.offset);
5619
5620                 spin_lock(&cache->space_info->lock);
5621                 spin_lock(&cache->lock);
5622
5623                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5624                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5625                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5626
5627                 old_val = btrfs_block_group_used(&cache->item);
5628                 num_bytes = min(total, cache->key.offset - byte_in_group);
5629                 if (alloc) {
5630                         old_val += num_bytes;
5631                         btrfs_set_block_group_used(&cache->item, old_val);
5632                         cache->reserved -= num_bytes;
5633                         cache->space_info->bytes_reserved -= num_bytes;
5634                         cache->space_info->bytes_used += num_bytes;
5635                         cache->space_info->disk_used += num_bytes * factor;
5636                         spin_unlock(&cache->lock);
5637                         spin_unlock(&cache->space_info->lock);
5638                 } else {
5639                         old_val -= num_bytes;
5640                         btrfs_set_block_group_used(&cache->item, old_val);
5641                         cache->pinned += num_bytes;
5642                         cache->space_info->bytes_pinned += num_bytes;
5643                         cache->space_info->bytes_used -= num_bytes;
5644                         cache->space_info->disk_used -= num_bytes * factor;
5645                         spin_unlock(&cache->lock);
5646                         spin_unlock(&cache->space_info->lock);
5647
5648                         set_extent_dirty(info->pinned_extents,
5649                                          bytenr, bytenr + num_bytes - 1,
5650                                          GFP_NOFS | __GFP_NOFAIL);
5651                         /*
5652                          * No longer have used bytes in this block group, queue
5653                          * it for deletion.
5654                          */
5655                         if (old_val == 0) {
5656                                 spin_lock(&info->unused_bgs_lock);
5657                                 if (list_empty(&cache->bg_list)) {
5658                                         btrfs_get_block_group(cache);
5659                                         list_add_tail(&cache->bg_list,
5660                                                       &info->unused_bgs);
5661                                 }
5662                                 spin_unlock(&info->unused_bgs_lock);
5663                         }
5664                 }
5665
5666                 spin_lock(&trans->transaction->dirty_bgs_lock);
5667                 if (list_empty(&cache->dirty_list)) {
5668                         list_add_tail(&cache->dirty_list,
5669                                       &trans->transaction->dirty_bgs);
5670                                 trans->transaction->num_dirty_bgs++;
5671                         btrfs_get_block_group(cache);
5672                 }
5673                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5674
5675                 btrfs_put_block_group(cache);
5676                 total -= num_bytes;
5677                 bytenr += num_bytes;
5678         }
5679         return 0;
5680 }
5681
5682 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5683 {
5684         struct btrfs_block_group_cache *cache;
5685         u64 bytenr;
5686
5687         spin_lock(&root->fs_info->block_group_cache_lock);
5688         bytenr = root->fs_info->first_logical_byte;
5689         spin_unlock(&root->fs_info->block_group_cache_lock);
5690
5691         if (bytenr < (u64)-1)
5692                 return bytenr;
5693
5694         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5695         if (!cache)
5696                 return 0;
5697
5698         bytenr = cache->key.objectid;
5699         btrfs_put_block_group(cache);
5700
5701         return bytenr;
5702 }
5703
5704 static int pin_down_extent(struct btrfs_root *root,
5705                            struct btrfs_block_group_cache *cache,
5706                            u64 bytenr, u64 num_bytes, int reserved)
5707 {
5708         spin_lock(&cache->space_info->lock);
5709         spin_lock(&cache->lock);
5710         cache->pinned += num_bytes;
5711         cache->space_info->bytes_pinned += num_bytes;
5712         if (reserved) {
5713                 cache->reserved -= num_bytes;
5714                 cache->space_info->bytes_reserved -= num_bytes;
5715         }
5716         spin_unlock(&cache->lock);
5717         spin_unlock(&cache->space_info->lock);
5718
5719         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5720                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5721         if (reserved)
5722                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5723         return 0;
5724 }
5725
5726 /*
5727  * this function must be called within transaction
5728  */
5729 int btrfs_pin_extent(struct btrfs_root *root,
5730                      u64 bytenr, u64 num_bytes, int reserved)
5731 {
5732         struct btrfs_block_group_cache *cache;
5733
5734         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5735         BUG_ON(!cache); /* Logic error */
5736
5737         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5738
5739         btrfs_put_block_group(cache);
5740         return 0;
5741 }
5742
5743 /*
5744  * this function must be called within transaction
5745  */
5746 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5747                                     u64 bytenr, u64 num_bytes)
5748 {
5749         struct btrfs_block_group_cache *cache;
5750         int ret;
5751
5752         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5753         if (!cache)
5754                 return -EINVAL;
5755
5756         /*
5757          * pull in the free space cache (if any) so that our pin
5758          * removes the free space from the cache.  We have load_only set
5759          * to one because the slow code to read in the free extents does check
5760          * the pinned extents.
5761          */
5762         cache_block_group(cache, 1);
5763
5764         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5765
5766         /* remove us from the free space cache (if we're there at all) */
5767         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5768         btrfs_put_block_group(cache);
5769         return ret;
5770 }
5771
5772 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5773 {
5774         int ret;
5775         struct btrfs_block_group_cache *block_group;
5776         struct btrfs_caching_control *caching_ctl;
5777
5778         block_group = btrfs_lookup_block_group(root->fs_info, start);
5779         if (!block_group)
5780                 return -EINVAL;
5781
5782         cache_block_group(block_group, 0);
5783         caching_ctl = get_caching_control(block_group);
5784
5785         if (!caching_ctl) {
5786                 /* Logic error */
5787                 BUG_ON(!block_group_cache_done(block_group));
5788                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5789         } else {
5790                 mutex_lock(&caching_ctl->mutex);
5791
5792                 if (start >= caching_ctl->progress) {
5793                         ret = add_excluded_extent(root, start, num_bytes);
5794                 } else if (start + num_bytes <= caching_ctl->progress) {
5795                         ret = btrfs_remove_free_space(block_group,
5796                                                       start, num_bytes);
5797                 } else {
5798                         num_bytes = caching_ctl->progress - start;
5799                         ret = btrfs_remove_free_space(block_group,
5800                                                       start, num_bytes);
5801                         if (ret)
5802                                 goto out_lock;
5803
5804                         num_bytes = (start + num_bytes) -
5805                                 caching_ctl->progress;
5806                         start = caching_ctl->progress;
5807                         ret = add_excluded_extent(root, start, num_bytes);
5808                 }
5809 out_lock:
5810                 mutex_unlock(&caching_ctl->mutex);
5811                 put_caching_control(caching_ctl);
5812         }
5813         btrfs_put_block_group(block_group);
5814         return ret;
5815 }
5816
5817 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5818                                  struct extent_buffer *eb)
5819 {
5820         struct btrfs_file_extent_item *item;
5821         struct btrfs_key key;
5822         int found_type;
5823         int i;
5824
5825         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5826                 return 0;
5827
5828         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5829                 btrfs_item_key_to_cpu(eb, &key, i);
5830                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5831                         continue;
5832                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5833                 found_type = btrfs_file_extent_type(eb, item);
5834                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5835                         continue;
5836                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5837                         continue;
5838                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5839                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5840                 __exclude_logged_extent(log, key.objectid, key.offset);
5841         }
5842
5843         return 0;
5844 }
5845
5846 /**
5847  * btrfs_update_reserved_bytes - update the block_group and space info counters
5848  * @cache:      The cache we are manipulating
5849  * @num_bytes:  The number of bytes in question
5850  * @reserve:    One of the reservation enums
5851  * @delalloc:   The blocks are allocated for the delalloc write
5852  *
5853  * This is called by the allocator when it reserves space, or by somebody who is
5854  * freeing space that was never actually used on disk.  For example if you
5855  * reserve some space for a new leaf in transaction A and before transaction A
5856  * commits you free that leaf, you call this with reserve set to 0 in order to
5857  * clear the reservation.
5858  *
5859  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5860  * ENOSPC accounting.  For data we handle the reservation through clearing the
5861  * delalloc bits in the io_tree.  We have to do this since we could end up
5862  * allocating less disk space for the amount of data we have reserved in the
5863  * case of compression.
5864  *
5865  * If this is a reservation and the block group has become read only we cannot
5866  * make the reservation and return -EAGAIN, otherwise this function always
5867  * succeeds.
5868  */
5869 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5870                                        u64 num_bytes, int reserve, int delalloc)
5871 {
5872         struct btrfs_space_info *space_info = cache->space_info;
5873         int ret = 0;
5874
5875         spin_lock(&space_info->lock);
5876         spin_lock(&cache->lock);
5877         if (reserve != RESERVE_FREE) {
5878                 if (cache->ro) {
5879                         ret = -EAGAIN;
5880                 } else {
5881                         cache->reserved += num_bytes;
5882                         space_info->bytes_reserved += num_bytes;
5883                         if (reserve == RESERVE_ALLOC) {
5884                                 trace_btrfs_space_reservation(cache->fs_info,
5885                                                 "space_info", space_info->flags,
5886                                                 num_bytes, 0);
5887                                 space_info->bytes_may_use -= num_bytes;
5888                         }
5889
5890                         if (delalloc)
5891                                 cache->delalloc_bytes += num_bytes;
5892                 }
5893         } else {
5894                 if (cache->ro)
5895                         space_info->bytes_readonly += num_bytes;
5896                 cache->reserved -= num_bytes;
5897                 space_info->bytes_reserved -= num_bytes;
5898
5899                 if (delalloc)
5900                         cache->delalloc_bytes -= num_bytes;
5901         }
5902         spin_unlock(&cache->lock);
5903         spin_unlock(&space_info->lock);
5904         return ret;
5905 }
5906
5907 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5908                                 struct btrfs_root *root)
5909 {
5910         struct btrfs_fs_info *fs_info = root->fs_info;
5911         struct btrfs_caching_control *next;
5912         struct btrfs_caching_control *caching_ctl;
5913         struct btrfs_block_group_cache *cache;
5914
5915         down_write(&fs_info->commit_root_sem);
5916
5917         list_for_each_entry_safe(caching_ctl, next,
5918                                  &fs_info->caching_block_groups, list) {
5919                 cache = caching_ctl->block_group;
5920                 if (block_group_cache_done(cache)) {
5921                         cache->last_byte_to_unpin = (u64)-1;
5922                         list_del_init(&caching_ctl->list);
5923                         put_caching_control(caching_ctl);
5924                 } else {
5925                         cache->last_byte_to_unpin = caching_ctl->progress;
5926                 }
5927         }
5928
5929         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5930                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5931         else
5932                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5933
5934         up_write(&fs_info->commit_root_sem);
5935
5936         update_global_block_rsv(fs_info);
5937 }
5938
5939 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
5940                               const bool return_free_space)
5941 {
5942         struct btrfs_fs_info *fs_info = root->fs_info;
5943         struct btrfs_block_group_cache *cache = NULL;
5944         struct btrfs_space_info *space_info;
5945         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5946         u64 len;
5947         bool readonly;
5948
5949         while (start <= end) {
5950                 readonly = false;
5951                 if (!cache ||
5952                     start >= cache->key.objectid + cache->key.offset) {
5953                         if (cache)
5954                                 btrfs_put_block_group(cache);
5955                         cache = btrfs_lookup_block_group(fs_info, start);
5956                         BUG_ON(!cache); /* Logic error */
5957                 }
5958
5959                 len = cache->key.objectid + cache->key.offset - start;
5960                 len = min(len, end + 1 - start);
5961
5962                 if (start < cache->last_byte_to_unpin) {
5963                         len = min(len, cache->last_byte_to_unpin - start);
5964                         if (return_free_space)
5965                                 btrfs_add_free_space(cache, start, len);
5966                 }
5967
5968                 start += len;
5969                 space_info = cache->space_info;
5970
5971                 spin_lock(&space_info->lock);
5972                 spin_lock(&cache->lock);
5973                 cache->pinned -= len;
5974                 space_info->bytes_pinned -= len;
5975                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
5976                 if (cache->ro) {
5977                         space_info->bytes_readonly += len;
5978                         readonly = true;
5979                 }
5980                 spin_unlock(&cache->lock);
5981                 if (!readonly && global_rsv->space_info == space_info) {
5982                         spin_lock(&global_rsv->lock);
5983                         if (!global_rsv->full) {
5984                                 len = min(len, global_rsv->size -
5985                                           global_rsv->reserved);
5986                                 global_rsv->reserved += len;
5987                                 space_info->bytes_may_use += len;
5988                                 if (global_rsv->reserved >= global_rsv->size)
5989                                         global_rsv->full = 1;
5990                         }
5991                         spin_unlock(&global_rsv->lock);
5992                 }
5993                 spin_unlock(&space_info->lock);
5994         }
5995
5996         if (cache)
5997                 btrfs_put_block_group(cache);
5998         return 0;
5999 }
6000
6001 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6002                                struct btrfs_root *root)
6003 {
6004         struct btrfs_fs_info *fs_info = root->fs_info;
6005         struct extent_io_tree *unpin;
6006         u64 start;
6007         u64 end;
6008         int ret;
6009
6010         if (trans->aborted)
6011                 return 0;
6012
6013         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6014                 unpin = &fs_info->freed_extents[1];
6015         else
6016                 unpin = &fs_info->freed_extents[0];
6017
6018         while (1) {
6019                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6020                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6021                                             EXTENT_DIRTY, NULL);
6022                 if (ret) {
6023                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6024                         break;
6025                 }
6026
6027                 if (btrfs_test_opt(root, DISCARD))
6028                         ret = btrfs_discard_extent(root, start,
6029                                                    end + 1 - start, NULL);
6030
6031                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
6032                 unpin_extent_range(root, start, end, true);
6033                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6034                 cond_resched();
6035         }
6036
6037         return 0;
6038 }
6039
6040 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6041                              u64 owner, u64 root_objectid)
6042 {
6043         struct btrfs_space_info *space_info;
6044         u64 flags;
6045
6046         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6047                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6048                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6049                 else
6050                         flags = BTRFS_BLOCK_GROUP_METADATA;
6051         } else {
6052                 flags = BTRFS_BLOCK_GROUP_DATA;
6053         }
6054
6055         space_info = __find_space_info(fs_info, flags);
6056         BUG_ON(!space_info); /* Logic bug */
6057         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6058 }
6059
6060
6061 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6062                                 struct btrfs_root *root,
6063                                 u64 bytenr, u64 num_bytes, u64 parent,
6064                                 u64 root_objectid, u64 owner_objectid,
6065                                 u64 owner_offset, int refs_to_drop,
6066                                 struct btrfs_delayed_extent_op *extent_op,
6067                                 int no_quota)
6068 {
6069         struct btrfs_key key;
6070         struct btrfs_path *path;
6071         struct btrfs_fs_info *info = root->fs_info;
6072         struct btrfs_root *extent_root = info->extent_root;
6073         struct extent_buffer *leaf;
6074         struct btrfs_extent_item *ei;
6075         struct btrfs_extent_inline_ref *iref;
6076         int ret;
6077         int is_data;
6078         int extent_slot = 0;
6079         int found_extent = 0;
6080         int num_to_del = 1;
6081         u32 item_size;
6082         u64 refs;
6083         int last_ref = 0;
6084         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_SUB_EXCL;
6085         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6086                                                  SKINNY_METADATA);
6087
6088         if (!info->quota_enabled || !is_fstree(root_objectid))
6089                 no_quota = 1;
6090
6091         path = btrfs_alloc_path();
6092         if (!path)
6093                 return -ENOMEM;
6094
6095         path->reada = 1;
6096         path->leave_spinning = 1;
6097
6098         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6099         BUG_ON(!is_data && refs_to_drop != 1);
6100
6101         if (is_data)
6102                 skinny_metadata = 0;
6103
6104         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6105                                     bytenr, num_bytes, parent,
6106                                     root_objectid, owner_objectid,
6107                                     owner_offset);
6108         if (ret == 0) {
6109                 extent_slot = path->slots[0];
6110                 while (extent_slot >= 0) {
6111                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6112                                               extent_slot);
6113                         if (key.objectid != bytenr)
6114                                 break;
6115                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6116                             key.offset == num_bytes) {
6117                                 found_extent = 1;
6118                                 break;
6119                         }
6120                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6121                             key.offset == owner_objectid) {
6122                                 found_extent = 1;
6123                                 break;
6124                         }
6125                         if (path->slots[0] - extent_slot > 5)
6126                                 break;
6127                         extent_slot--;
6128                 }
6129 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6130                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6131                 if (found_extent && item_size < sizeof(*ei))
6132                         found_extent = 0;
6133 #endif
6134                 if (!found_extent) {
6135                         BUG_ON(iref);
6136                         ret = remove_extent_backref(trans, extent_root, path,
6137                                                     NULL, refs_to_drop,
6138                                                     is_data, &last_ref);
6139                         if (ret) {
6140                                 btrfs_abort_transaction(trans, extent_root, ret);
6141                                 goto out;
6142                         }
6143                         btrfs_release_path(path);
6144                         path->leave_spinning = 1;
6145
6146                         key.objectid = bytenr;
6147                         key.type = BTRFS_EXTENT_ITEM_KEY;
6148                         key.offset = num_bytes;
6149
6150                         if (!is_data && skinny_metadata) {
6151                                 key.type = BTRFS_METADATA_ITEM_KEY;
6152                                 key.offset = owner_objectid;
6153                         }
6154
6155                         ret = btrfs_search_slot(trans, extent_root,
6156                                                 &key, path, -1, 1);
6157                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6158                                 /*
6159                                  * Couldn't find our skinny metadata item,
6160                                  * see if we have ye olde extent item.
6161                                  */
6162                                 path->slots[0]--;
6163                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6164                                                       path->slots[0]);
6165                                 if (key.objectid == bytenr &&
6166                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6167                                     key.offset == num_bytes)
6168                                         ret = 0;
6169                         }
6170
6171                         if (ret > 0 && skinny_metadata) {
6172                                 skinny_metadata = false;
6173                                 key.objectid = bytenr;
6174                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6175                                 key.offset = num_bytes;
6176                                 btrfs_release_path(path);
6177                                 ret = btrfs_search_slot(trans, extent_root,
6178                                                         &key, path, -1, 1);
6179                         }
6180
6181                         if (ret) {
6182                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6183                                         ret, bytenr);
6184                                 if (ret > 0)
6185                                         btrfs_print_leaf(extent_root,
6186                                                          path->nodes[0]);
6187                         }
6188                         if (ret < 0) {
6189                                 btrfs_abort_transaction(trans, extent_root, ret);
6190                                 goto out;
6191                         }
6192                         extent_slot = path->slots[0];
6193                 }
6194         } else if (WARN_ON(ret == -ENOENT)) {
6195                 btrfs_print_leaf(extent_root, path->nodes[0]);
6196                 btrfs_err(info,
6197                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6198                         bytenr, parent, root_objectid, owner_objectid,
6199                         owner_offset);
6200                 btrfs_abort_transaction(trans, extent_root, ret);
6201                 goto out;
6202         } else {
6203                 btrfs_abort_transaction(trans, extent_root, ret);
6204                 goto out;
6205         }
6206
6207         leaf = path->nodes[0];
6208         item_size = btrfs_item_size_nr(leaf, extent_slot);
6209 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6210         if (item_size < sizeof(*ei)) {
6211                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6212                 ret = convert_extent_item_v0(trans, extent_root, path,
6213                                              owner_objectid, 0);
6214                 if (ret < 0) {
6215                         btrfs_abort_transaction(trans, extent_root, ret);
6216                         goto out;
6217                 }
6218
6219                 btrfs_release_path(path);
6220                 path->leave_spinning = 1;
6221
6222                 key.objectid = bytenr;
6223                 key.type = BTRFS_EXTENT_ITEM_KEY;
6224                 key.offset = num_bytes;
6225
6226                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6227                                         -1, 1);
6228                 if (ret) {
6229                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6230                                 ret, bytenr);
6231                         btrfs_print_leaf(extent_root, path->nodes[0]);
6232                 }
6233                 if (ret < 0) {
6234                         btrfs_abort_transaction(trans, extent_root, ret);
6235                         goto out;
6236                 }
6237
6238                 extent_slot = path->slots[0];
6239                 leaf = path->nodes[0];
6240                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6241         }
6242 #endif
6243         BUG_ON(item_size < sizeof(*ei));
6244         ei = btrfs_item_ptr(leaf, extent_slot,
6245                             struct btrfs_extent_item);
6246         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6247             key.type == BTRFS_EXTENT_ITEM_KEY) {
6248                 struct btrfs_tree_block_info *bi;
6249                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6250                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6251                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6252         }
6253
6254         refs = btrfs_extent_refs(leaf, ei);
6255         if (refs < refs_to_drop) {
6256                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6257                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6258                 ret = -EINVAL;
6259                 btrfs_abort_transaction(trans, extent_root, ret);
6260                 goto out;
6261         }
6262         refs -= refs_to_drop;
6263
6264         if (refs > 0) {
6265                 type = BTRFS_QGROUP_OPER_SUB_SHARED;
6266                 if (extent_op)
6267                         __run_delayed_extent_op(extent_op, leaf, ei);
6268                 /*
6269                  * In the case of inline back ref, reference count will
6270                  * be updated by remove_extent_backref
6271                  */
6272                 if (iref) {
6273                         BUG_ON(!found_extent);
6274                 } else {
6275                         btrfs_set_extent_refs(leaf, ei, refs);
6276                         btrfs_mark_buffer_dirty(leaf);
6277                 }
6278                 if (found_extent) {
6279                         ret = remove_extent_backref(trans, extent_root, path,
6280                                                     iref, refs_to_drop,
6281                                                     is_data, &last_ref);
6282                         if (ret) {
6283                                 btrfs_abort_transaction(trans, extent_root, ret);
6284                                 goto out;
6285                         }
6286                 }
6287                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6288                                  root_objectid);
6289         } else {
6290                 if (found_extent) {
6291                         BUG_ON(is_data && refs_to_drop !=
6292                                extent_data_ref_count(root, path, iref));
6293                         if (iref) {
6294                                 BUG_ON(path->slots[0] != extent_slot);
6295                         } else {
6296                                 BUG_ON(path->slots[0] != extent_slot + 1);
6297                                 path->slots[0] = extent_slot;
6298                                 num_to_del = 2;
6299                         }
6300                 }
6301
6302                 last_ref = 1;
6303                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6304                                       num_to_del);
6305                 if (ret) {
6306                         btrfs_abort_transaction(trans, extent_root, ret);
6307                         goto out;
6308                 }
6309                 btrfs_release_path(path);
6310
6311                 if (is_data) {
6312                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6313                         if (ret) {
6314                                 btrfs_abort_transaction(trans, extent_root, ret);
6315                                 goto out;
6316                         }
6317                 }
6318
6319                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6320                 if (ret) {
6321                         btrfs_abort_transaction(trans, extent_root, ret);
6322                         goto out;
6323                 }
6324         }
6325         btrfs_release_path(path);
6326
6327         /* Deal with the quota accounting */
6328         if (!ret && last_ref && !no_quota) {
6329                 int mod_seq = 0;
6330
6331                 if (owner_objectid >= BTRFS_FIRST_FREE_OBJECTID &&
6332                     type == BTRFS_QGROUP_OPER_SUB_SHARED)
6333                         mod_seq = 1;
6334
6335                 ret = btrfs_qgroup_record_ref(trans, info, root_objectid,
6336                                               bytenr, num_bytes, type,
6337                                               mod_seq);
6338         }
6339 out:
6340         btrfs_free_path(path);
6341         return ret;
6342 }
6343
6344 /*
6345  * when we free an block, it is possible (and likely) that we free the last
6346  * delayed ref for that extent as well.  This searches the delayed ref tree for
6347  * a given extent, and if there are no other delayed refs to be processed, it
6348  * removes it from the tree.
6349  */
6350 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6351                                       struct btrfs_root *root, u64 bytenr)
6352 {
6353         struct btrfs_delayed_ref_head *head;
6354         struct btrfs_delayed_ref_root *delayed_refs;
6355         int ret = 0;
6356
6357         delayed_refs = &trans->transaction->delayed_refs;
6358         spin_lock(&delayed_refs->lock);
6359         head = btrfs_find_delayed_ref_head(trans, bytenr);
6360         if (!head)
6361                 goto out_delayed_unlock;
6362
6363         spin_lock(&head->lock);
6364         if (rb_first(&head->ref_root))
6365                 goto out;
6366
6367         if (head->extent_op) {
6368                 if (!head->must_insert_reserved)
6369                         goto out;
6370                 btrfs_free_delayed_extent_op(head->extent_op);
6371                 head->extent_op = NULL;
6372         }
6373
6374         /*
6375          * waiting for the lock here would deadlock.  If someone else has it
6376          * locked they are already in the process of dropping it anyway
6377          */
6378         if (!mutex_trylock(&head->mutex))
6379                 goto out;
6380
6381         /*
6382          * at this point we have a head with no other entries.  Go
6383          * ahead and process it.
6384          */
6385         head->node.in_tree = 0;
6386         rb_erase(&head->href_node, &delayed_refs->href_root);
6387
6388         atomic_dec(&delayed_refs->num_entries);
6389
6390         /*
6391          * we don't take a ref on the node because we're removing it from the
6392          * tree, so we just steal the ref the tree was holding.
6393          */
6394         delayed_refs->num_heads--;
6395         if (head->processing == 0)
6396                 delayed_refs->num_heads_ready--;
6397         head->processing = 0;
6398         spin_unlock(&head->lock);
6399         spin_unlock(&delayed_refs->lock);
6400
6401         BUG_ON(head->extent_op);
6402         if (head->must_insert_reserved)
6403                 ret = 1;
6404
6405         mutex_unlock(&head->mutex);
6406         btrfs_put_delayed_ref(&head->node);
6407         return ret;
6408 out:
6409         spin_unlock(&head->lock);
6410
6411 out_delayed_unlock:
6412         spin_unlock(&delayed_refs->lock);
6413         return 0;
6414 }
6415
6416 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6417                            struct btrfs_root *root,
6418                            struct extent_buffer *buf,
6419                            u64 parent, int last_ref)
6420 {
6421         int pin = 1;
6422         int ret;
6423
6424         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6425                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6426                                         buf->start, buf->len,
6427                                         parent, root->root_key.objectid,
6428                                         btrfs_header_level(buf),
6429                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6430                 BUG_ON(ret); /* -ENOMEM */
6431         }
6432
6433         if (!last_ref)
6434                 return;
6435
6436         if (btrfs_header_generation(buf) == trans->transid) {
6437                 struct btrfs_block_group_cache *cache;
6438
6439                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6440                         ret = check_ref_cleanup(trans, root, buf->start);
6441                         if (!ret)
6442                                 goto out;
6443                 }
6444
6445                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6446
6447                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6448                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6449                         btrfs_put_block_group(cache);
6450                         goto out;
6451                 }
6452
6453                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6454
6455                 btrfs_add_free_space(cache, buf->start, buf->len);
6456                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6457                 btrfs_put_block_group(cache);
6458                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6459                 pin = 0;
6460         }
6461 out:
6462         if (pin)
6463                 add_pinned_bytes(root->fs_info, buf->len,
6464                                  btrfs_header_level(buf),
6465                                  root->root_key.objectid);
6466
6467         /*
6468          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6469          * anymore.
6470          */
6471         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6472 }
6473
6474 /* Can return -ENOMEM */
6475 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6476                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6477                       u64 owner, u64 offset, int no_quota)
6478 {
6479         int ret;
6480         struct btrfs_fs_info *fs_info = root->fs_info;
6481
6482         if (btrfs_test_is_dummy_root(root))
6483                 return 0;
6484
6485         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6486
6487         /*
6488          * tree log blocks never actually go into the extent allocation
6489          * tree, just update pinning info and exit early.
6490          */
6491         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6492                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6493                 /* unlocks the pinned mutex */
6494                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6495                 ret = 0;
6496         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6497                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6498                                         num_bytes,
6499                                         parent, root_objectid, (int)owner,
6500                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6501         } else {
6502                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6503                                                 num_bytes,
6504                                                 parent, root_objectid, owner,
6505                                                 offset, BTRFS_DROP_DELAYED_REF,
6506                                                 NULL, no_quota);
6507         }
6508         return ret;
6509 }
6510
6511 /*
6512  * when we wait for progress in the block group caching, its because
6513  * our allocation attempt failed at least once.  So, we must sleep
6514  * and let some progress happen before we try again.
6515  *
6516  * This function will sleep at least once waiting for new free space to
6517  * show up, and then it will check the block group free space numbers
6518  * for our min num_bytes.  Another option is to have it go ahead
6519  * and look in the rbtree for a free extent of a given size, but this
6520  * is a good start.
6521  *
6522  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6523  * any of the information in this block group.
6524  */
6525 static noinline void
6526 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6527                                 u64 num_bytes)
6528 {
6529         struct btrfs_caching_control *caching_ctl;
6530
6531         caching_ctl = get_caching_control(cache);
6532         if (!caching_ctl)
6533                 return;
6534
6535         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6536                    (cache->free_space_ctl->free_space >= num_bytes));
6537
6538         put_caching_control(caching_ctl);
6539 }
6540
6541 static noinline int
6542 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6543 {
6544         struct btrfs_caching_control *caching_ctl;
6545         int ret = 0;
6546
6547         caching_ctl = get_caching_control(cache);
6548         if (!caching_ctl)
6549                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6550
6551         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6552         if (cache->cached == BTRFS_CACHE_ERROR)
6553                 ret = -EIO;
6554         put_caching_control(caching_ctl);
6555         return ret;
6556 }
6557
6558 int __get_raid_index(u64 flags)
6559 {
6560         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6561                 return BTRFS_RAID_RAID10;
6562         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6563                 return BTRFS_RAID_RAID1;
6564         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6565                 return BTRFS_RAID_DUP;
6566         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6567                 return BTRFS_RAID_RAID0;
6568         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6569                 return BTRFS_RAID_RAID5;
6570         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6571                 return BTRFS_RAID_RAID6;
6572
6573         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6574 }
6575
6576 int get_block_group_index(struct btrfs_block_group_cache *cache)
6577 {
6578         return __get_raid_index(cache->flags);
6579 }
6580
6581 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6582         [BTRFS_RAID_RAID10]     = "raid10",
6583         [BTRFS_RAID_RAID1]      = "raid1",
6584         [BTRFS_RAID_DUP]        = "dup",
6585         [BTRFS_RAID_RAID0]      = "raid0",
6586         [BTRFS_RAID_SINGLE]     = "single",
6587         [BTRFS_RAID_RAID5]      = "raid5",
6588         [BTRFS_RAID_RAID6]      = "raid6",
6589 };
6590
6591 static const char *get_raid_name(enum btrfs_raid_types type)
6592 {
6593         if (type >= BTRFS_NR_RAID_TYPES)
6594                 return NULL;
6595
6596         return btrfs_raid_type_names[type];
6597 }
6598
6599 enum btrfs_loop_type {
6600         LOOP_CACHING_NOWAIT = 0,
6601         LOOP_CACHING_WAIT = 1,
6602         LOOP_ALLOC_CHUNK = 2,
6603         LOOP_NO_EMPTY_SIZE = 3,
6604 };
6605
6606 static inline void
6607 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6608                        int delalloc)
6609 {
6610         if (delalloc)
6611                 down_read(&cache->data_rwsem);
6612 }
6613
6614 static inline void
6615 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6616                        int delalloc)
6617 {
6618         btrfs_get_block_group(cache);
6619         if (delalloc)
6620                 down_read(&cache->data_rwsem);
6621 }
6622
6623 static struct btrfs_block_group_cache *
6624 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6625                    struct btrfs_free_cluster *cluster,
6626                    int delalloc)
6627 {
6628         struct btrfs_block_group_cache *used_bg;
6629         bool locked = false;
6630 again:
6631         spin_lock(&cluster->refill_lock);
6632         if (locked) {
6633                 if (used_bg == cluster->block_group)
6634                         return used_bg;
6635
6636                 up_read(&used_bg->data_rwsem);
6637                 btrfs_put_block_group(used_bg);
6638         }
6639
6640         used_bg = cluster->block_group;
6641         if (!used_bg)
6642                 return NULL;
6643
6644         if (used_bg == block_group)
6645                 return used_bg;
6646
6647         btrfs_get_block_group(used_bg);
6648
6649         if (!delalloc)
6650                 return used_bg;
6651
6652         if (down_read_trylock(&used_bg->data_rwsem))
6653                 return used_bg;
6654
6655         spin_unlock(&cluster->refill_lock);
6656         down_read(&used_bg->data_rwsem);
6657         locked = true;
6658         goto again;
6659 }
6660
6661 static inline void
6662 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6663                          int delalloc)
6664 {
6665         if (delalloc)
6666                 up_read(&cache->data_rwsem);
6667         btrfs_put_block_group(cache);
6668 }
6669
6670 /*
6671  * walks the btree of allocated extents and find a hole of a given size.
6672  * The key ins is changed to record the hole:
6673  * ins->objectid == start position
6674  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6675  * ins->offset == the size of the hole.
6676  * Any available blocks before search_start are skipped.
6677  *
6678  * If there is no suitable free space, we will record the max size of
6679  * the free space extent currently.
6680  */
6681 static noinline int find_free_extent(struct btrfs_root *orig_root,
6682                                      u64 num_bytes, u64 empty_size,
6683                                      u64 hint_byte, struct btrfs_key *ins,
6684                                      u64 flags, int delalloc)
6685 {
6686         int ret = 0;
6687         struct btrfs_root *root = orig_root->fs_info->extent_root;
6688         struct btrfs_free_cluster *last_ptr = NULL;
6689         struct btrfs_block_group_cache *block_group = NULL;
6690         u64 search_start = 0;
6691         u64 max_extent_size = 0;
6692         int empty_cluster = 2 * 1024 * 1024;
6693         struct btrfs_space_info *space_info;
6694         int loop = 0;
6695         int index = __get_raid_index(flags);
6696         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6697                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6698         bool failed_cluster_refill = false;
6699         bool failed_alloc = false;
6700         bool use_cluster = true;
6701         bool have_caching_bg = false;
6702
6703         WARN_ON(num_bytes < root->sectorsize);
6704         ins->type = BTRFS_EXTENT_ITEM_KEY;
6705         ins->objectid = 0;
6706         ins->offset = 0;
6707
6708         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6709
6710         space_info = __find_space_info(root->fs_info, flags);
6711         if (!space_info) {
6712                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6713                 return -ENOSPC;
6714         }
6715
6716         /*
6717          * If the space info is for both data and metadata it means we have a
6718          * small filesystem and we can't use the clustering stuff.
6719          */
6720         if (btrfs_mixed_space_info(space_info))
6721                 use_cluster = false;
6722
6723         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6724                 last_ptr = &root->fs_info->meta_alloc_cluster;
6725                 if (!btrfs_test_opt(root, SSD))
6726                         empty_cluster = 64 * 1024;
6727         }
6728
6729         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6730             btrfs_test_opt(root, SSD)) {
6731                 last_ptr = &root->fs_info->data_alloc_cluster;
6732         }
6733
6734         if (last_ptr) {
6735                 spin_lock(&last_ptr->lock);
6736                 if (last_ptr->block_group)
6737                         hint_byte = last_ptr->window_start;
6738                 spin_unlock(&last_ptr->lock);
6739         }
6740
6741         search_start = max(search_start, first_logical_byte(root, 0));
6742         search_start = max(search_start, hint_byte);
6743
6744         if (!last_ptr)
6745                 empty_cluster = 0;
6746
6747         if (search_start == hint_byte) {
6748                 block_group = btrfs_lookup_block_group(root->fs_info,
6749                                                        search_start);
6750                 /*
6751                  * we don't want to use the block group if it doesn't match our
6752                  * allocation bits, or if its not cached.
6753                  *
6754                  * However if we are re-searching with an ideal block group
6755                  * picked out then we don't care that the block group is cached.
6756                  */
6757                 if (block_group && block_group_bits(block_group, flags) &&
6758                     block_group->cached != BTRFS_CACHE_NO) {
6759                         down_read(&space_info->groups_sem);
6760                         if (list_empty(&block_group->list) ||
6761                             block_group->ro) {
6762                                 /*
6763                                  * someone is removing this block group,
6764                                  * we can't jump into the have_block_group
6765                                  * target because our list pointers are not
6766                                  * valid
6767                                  */
6768                                 btrfs_put_block_group(block_group);
6769                                 up_read(&space_info->groups_sem);
6770                         } else {
6771                                 index = get_block_group_index(block_group);
6772                                 btrfs_lock_block_group(block_group, delalloc);
6773                                 goto have_block_group;
6774                         }
6775                 } else if (block_group) {
6776                         btrfs_put_block_group(block_group);
6777                 }
6778         }
6779 search:
6780         have_caching_bg = false;
6781         down_read(&space_info->groups_sem);
6782         list_for_each_entry(block_group, &space_info->block_groups[index],
6783                             list) {
6784                 u64 offset;
6785                 int cached;
6786
6787                 btrfs_grab_block_group(block_group, delalloc);
6788                 search_start = block_group->key.objectid;
6789
6790                 /*
6791                  * this can happen if we end up cycling through all the
6792                  * raid types, but we want to make sure we only allocate
6793                  * for the proper type.
6794                  */
6795                 if (!block_group_bits(block_group, flags)) {
6796                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6797                                 BTRFS_BLOCK_GROUP_RAID1 |
6798                                 BTRFS_BLOCK_GROUP_RAID5 |
6799                                 BTRFS_BLOCK_GROUP_RAID6 |
6800                                 BTRFS_BLOCK_GROUP_RAID10;
6801
6802                         /*
6803                          * if they asked for extra copies and this block group
6804                          * doesn't provide them, bail.  This does allow us to
6805                          * fill raid0 from raid1.
6806                          */
6807                         if ((flags & extra) && !(block_group->flags & extra))
6808                                 goto loop;
6809                 }
6810
6811 have_block_group:
6812                 cached = block_group_cache_done(block_group);
6813                 if (unlikely(!cached)) {
6814                         ret = cache_block_group(block_group, 0);
6815                         BUG_ON(ret < 0);
6816                         ret = 0;
6817                 }
6818
6819                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6820                         goto loop;
6821                 if (unlikely(block_group->ro))
6822                         goto loop;
6823
6824                 /*
6825                  * Ok we want to try and use the cluster allocator, so
6826                  * lets look there
6827                  */
6828                 if (last_ptr) {
6829                         struct btrfs_block_group_cache *used_block_group;
6830                         unsigned long aligned_cluster;
6831                         /*
6832                          * the refill lock keeps out other
6833                          * people trying to start a new cluster
6834                          */
6835                         used_block_group = btrfs_lock_cluster(block_group,
6836                                                               last_ptr,
6837                                                               delalloc);
6838                         if (!used_block_group)
6839                                 goto refill_cluster;
6840
6841                         if (used_block_group != block_group &&
6842                             (used_block_group->ro ||
6843                              !block_group_bits(used_block_group, flags)))
6844                                 goto release_cluster;
6845
6846                         offset = btrfs_alloc_from_cluster(used_block_group,
6847                                                 last_ptr,
6848                                                 num_bytes,
6849                                                 used_block_group->key.objectid,
6850                                                 &max_extent_size);
6851                         if (offset) {
6852                                 /* we have a block, we're done */
6853                                 spin_unlock(&last_ptr->refill_lock);
6854                                 trace_btrfs_reserve_extent_cluster(root,
6855                                                 used_block_group,
6856                                                 search_start, num_bytes);
6857                                 if (used_block_group != block_group) {
6858                                         btrfs_release_block_group(block_group,
6859                                                                   delalloc);
6860                                         block_group = used_block_group;
6861                                 }
6862                                 goto checks;
6863                         }
6864
6865                         WARN_ON(last_ptr->block_group != used_block_group);
6866 release_cluster:
6867                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6868                          * set up a new clusters, so lets just skip it
6869                          * and let the allocator find whatever block
6870                          * it can find.  If we reach this point, we
6871                          * will have tried the cluster allocator
6872                          * plenty of times and not have found
6873                          * anything, so we are likely way too
6874                          * fragmented for the clustering stuff to find
6875                          * anything.
6876                          *
6877                          * However, if the cluster is taken from the
6878                          * current block group, release the cluster
6879                          * first, so that we stand a better chance of
6880                          * succeeding in the unclustered
6881                          * allocation.  */
6882                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6883                             used_block_group != block_group) {
6884                                 spin_unlock(&last_ptr->refill_lock);
6885                                 btrfs_release_block_group(used_block_group,
6886                                                           delalloc);
6887                                 goto unclustered_alloc;
6888                         }
6889
6890                         /*
6891                          * this cluster didn't work out, free it and
6892                          * start over
6893                          */
6894                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6895
6896                         if (used_block_group != block_group)
6897                                 btrfs_release_block_group(used_block_group,
6898                                                           delalloc);
6899 refill_cluster:
6900                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6901                                 spin_unlock(&last_ptr->refill_lock);
6902                                 goto unclustered_alloc;
6903                         }
6904
6905                         aligned_cluster = max_t(unsigned long,
6906                                                 empty_cluster + empty_size,
6907                                               block_group->full_stripe_len);
6908
6909                         /* allocate a cluster in this block group */
6910                         ret = btrfs_find_space_cluster(root, block_group,
6911                                                        last_ptr, search_start,
6912                                                        num_bytes,
6913                                                        aligned_cluster);
6914                         if (ret == 0) {
6915                                 /*
6916                                  * now pull our allocation out of this
6917                                  * cluster
6918                                  */
6919                                 offset = btrfs_alloc_from_cluster(block_group,
6920                                                         last_ptr,
6921                                                         num_bytes,
6922                                                         search_start,
6923                                                         &max_extent_size);
6924                                 if (offset) {
6925                                         /* we found one, proceed */
6926                                         spin_unlock(&last_ptr->refill_lock);
6927                                         trace_btrfs_reserve_extent_cluster(root,
6928                                                 block_group, search_start,
6929                                                 num_bytes);
6930                                         goto checks;
6931                                 }
6932                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6933                                    && !failed_cluster_refill) {
6934                                 spin_unlock(&last_ptr->refill_lock);
6935
6936                                 failed_cluster_refill = true;
6937                                 wait_block_group_cache_progress(block_group,
6938                                        num_bytes + empty_cluster + empty_size);
6939                                 goto have_block_group;
6940                         }
6941
6942                         /*
6943                          * at this point we either didn't find a cluster
6944                          * or we weren't able to allocate a block from our
6945                          * cluster.  Free the cluster we've been trying
6946                          * to use, and go to the next block group
6947                          */
6948                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6949                         spin_unlock(&last_ptr->refill_lock);
6950                         goto loop;
6951                 }
6952
6953 unclustered_alloc:
6954                 spin_lock(&block_group->free_space_ctl->tree_lock);
6955                 if (cached &&
6956                     block_group->free_space_ctl->free_space <
6957                     num_bytes + empty_cluster + empty_size) {
6958                         if (block_group->free_space_ctl->free_space >
6959                             max_extent_size)
6960                                 max_extent_size =
6961                                         block_group->free_space_ctl->free_space;
6962                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6963                         goto loop;
6964                 }
6965                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6966
6967                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6968                                                     num_bytes, empty_size,
6969                                                     &max_extent_size);
6970                 /*
6971                  * If we didn't find a chunk, and we haven't failed on this
6972                  * block group before, and this block group is in the middle of
6973                  * caching and we are ok with waiting, then go ahead and wait
6974                  * for progress to be made, and set failed_alloc to true.
6975                  *
6976                  * If failed_alloc is true then we've already waited on this
6977                  * block group once and should move on to the next block group.
6978                  */
6979                 if (!offset && !failed_alloc && !cached &&
6980                     loop > LOOP_CACHING_NOWAIT) {
6981                         wait_block_group_cache_progress(block_group,
6982                                                 num_bytes + empty_size);
6983                         failed_alloc = true;
6984                         goto have_block_group;
6985                 } else if (!offset) {
6986                         if (!cached)
6987                                 have_caching_bg = true;
6988                         goto loop;
6989                 }
6990 checks:
6991                 search_start = ALIGN(offset, root->stripesize);
6992
6993                 /* move on to the next group */
6994                 if (search_start + num_bytes >
6995                     block_group->key.objectid + block_group->key.offset) {
6996                         btrfs_add_free_space(block_group, offset, num_bytes);
6997                         goto loop;
6998                 }
6999
7000                 if (offset < search_start)
7001                         btrfs_add_free_space(block_group, offset,
7002                                              search_start - offset);
7003                 BUG_ON(offset > search_start);
7004
7005                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7006                                                   alloc_type, delalloc);
7007                 if (ret == -EAGAIN) {
7008                         btrfs_add_free_space(block_group, offset, num_bytes);
7009                         goto loop;
7010                 }
7011
7012                 /* we are all good, lets return */
7013                 ins->objectid = search_start;
7014                 ins->offset = num_bytes;
7015
7016                 trace_btrfs_reserve_extent(orig_root, block_group,
7017                                            search_start, num_bytes);
7018                 btrfs_release_block_group(block_group, delalloc);
7019                 break;
7020 loop:
7021                 failed_cluster_refill = false;
7022                 failed_alloc = false;
7023                 BUG_ON(index != get_block_group_index(block_group));
7024                 btrfs_release_block_group(block_group, delalloc);
7025         }
7026         up_read(&space_info->groups_sem);
7027
7028         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7029                 goto search;
7030
7031         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7032                 goto search;
7033
7034         /*
7035          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7036          *                      caching kthreads as we move along
7037          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7038          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7039          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7040          *                      again
7041          */
7042         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7043                 index = 0;
7044                 loop++;
7045                 if (loop == LOOP_ALLOC_CHUNK) {
7046                         struct btrfs_trans_handle *trans;
7047                         int exist = 0;
7048
7049                         trans = current->journal_info;
7050                         if (trans)
7051                                 exist = 1;
7052                         else
7053                                 trans = btrfs_join_transaction(root);
7054
7055                         if (IS_ERR(trans)) {
7056                                 ret = PTR_ERR(trans);
7057                                 goto out;
7058                         }
7059
7060                         ret = do_chunk_alloc(trans, root, flags,
7061                                              CHUNK_ALLOC_FORCE);
7062                         /*
7063                          * Do not bail out on ENOSPC since we
7064                          * can do more things.
7065                          */
7066                         if (ret < 0 && ret != -ENOSPC)
7067                                 btrfs_abort_transaction(trans,
7068                                                         root, ret);
7069                         else
7070                                 ret = 0;
7071                         if (!exist)
7072                                 btrfs_end_transaction(trans, root);
7073                         if (ret)
7074                                 goto out;
7075                 }
7076
7077                 if (loop == LOOP_NO_EMPTY_SIZE) {
7078                         empty_size = 0;
7079                         empty_cluster = 0;
7080                 }
7081
7082                 goto search;
7083         } else if (!ins->objectid) {
7084                 ret = -ENOSPC;
7085         } else if (ins->objectid) {
7086                 ret = 0;
7087         }
7088 out:
7089         if (ret == -ENOSPC)
7090                 ins->offset = max_extent_size;
7091         return ret;
7092 }
7093
7094 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7095                             int dump_block_groups)
7096 {
7097         struct btrfs_block_group_cache *cache;
7098         int index = 0;
7099
7100         spin_lock(&info->lock);
7101         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7102                info->flags,
7103                info->total_bytes - info->bytes_used - info->bytes_pinned -
7104                info->bytes_reserved - info->bytes_readonly,
7105                (info->full) ? "" : "not ");
7106         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7107                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7108                info->total_bytes, info->bytes_used, info->bytes_pinned,
7109                info->bytes_reserved, info->bytes_may_use,
7110                info->bytes_readonly);
7111         spin_unlock(&info->lock);
7112
7113         if (!dump_block_groups)
7114                 return;
7115
7116         down_read(&info->groups_sem);
7117 again:
7118         list_for_each_entry(cache, &info->block_groups[index], list) {
7119                 spin_lock(&cache->lock);
7120                 printk(KERN_INFO "BTRFS: "
7121                            "block group %llu has %llu bytes, "
7122                            "%llu used %llu pinned %llu reserved %s\n",
7123                        cache->key.objectid, cache->key.offset,
7124                        btrfs_block_group_used(&cache->item), cache->pinned,
7125                        cache->reserved, cache->ro ? "[readonly]" : "");
7126                 btrfs_dump_free_space(cache, bytes);
7127                 spin_unlock(&cache->lock);
7128         }
7129         if (++index < BTRFS_NR_RAID_TYPES)
7130                 goto again;
7131         up_read(&info->groups_sem);
7132 }
7133
7134 int btrfs_reserve_extent(struct btrfs_root *root,
7135                          u64 num_bytes, u64 min_alloc_size,
7136                          u64 empty_size, u64 hint_byte,
7137                          struct btrfs_key *ins, int is_data, int delalloc)
7138 {
7139         bool final_tried = false;
7140         u64 flags;
7141         int ret;
7142
7143         flags = btrfs_get_alloc_profile(root, is_data);
7144 again:
7145         WARN_ON(num_bytes < root->sectorsize);
7146         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7147                                flags, delalloc);
7148
7149         if (ret == -ENOSPC) {
7150                 if (!final_tried && ins->offset) {
7151                         num_bytes = min(num_bytes >> 1, ins->offset);
7152                         num_bytes = round_down(num_bytes, root->sectorsize);
7153                         num_bytes = max(num_bytes, min_alloc_size);
7154                         if (num_bytes == min_alloc_size)
7155                                 final_tried = true;
7156                         goto again;
7157                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7158                         struct btrfs_space_info *sinfo;
7159
7160                         sinfo = __find_space_info(root->fs_info, flags);
7161                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7162                                 flags, num_bytes);
7163                         if (sinfo)
7164                                 dump_space_info(sinfo, num_bytes, 1);
7165                 }
7166         }
7167
7168         return ret;
7169 }
7170
7171 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7172                                         u64 start, u64 len,
7173                                         int pin, int delalloc)
7174 {
7175         struct btrfs_block_group_cache *cache;
7176         int ret = 0;
7177
7178         cache = btrfs_lookup_block_group(root->fs_info, start);
7179         if (!cache) {
7180                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7181                         start);
7182                 return -ENOSPC;
7183         }
7184
7185         if (pin)
7186                 pin_down_extent(root, cache, start, len, 1);
7187         else {
7188                 if (btrfs_test_opt(root, DISCARD))
7189                         ret = btrfs_discard_extent(root, start, len, NULL);
7190                 btrfs_add_free_space(cache, start, len);
7191                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7192         }
7193
7194         btrfs_put_block_group(cache);
7195
7196         trace_btrfs_reserved_extent_free(root, start, len);
7197
7198         return ret;
7199 }
7200
7201 int btrfs_free_reserved_extent(struct btrfs_root *root,
7202                                u64 start, u64 len, int delalloc)
7203 {
7204         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7205 }
7206
7207 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7208                                        u64 start, u64 len)
7209 {
7210         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7211 }
7212
7213 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7214                                       struct btrfs_root *root,
7215                                       u64 parent, u64 root_objectid,
7216                                       u64 flags, u64 owner, u64 offset,
7217                                       struct btrfs_key *ins, int ref_mod)
7218 {
7219         int ret;
7220         struct btrfs_fs_info *fs_info = root->fs_info;
7221         struct btrfs_extent_item *extent_item;
7222         struct btrfs_extent_inline_ref *iref;
7223         struct btrfs_path *path;
7224         struct extent_buffer *leaf;
7225         int type;
7226         u32 size;
7227
7228         if (parent > 0)
7229                 type = BTRFS_SHARED_DATA_REF_KEY;
7230         else
7231                 type = BTRFS_EXTENT_DATA_REF_KEY;
7232
7233         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7234
7235         path = btrfs_alloc_path();
7236         if (!path)
7237                 return -ENOMEM;
7238
7239         path->leave_spinning = 1;
7240         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7241                                       ins, size);
7242         if (ret) {
7243                 btrfs_free_path(path);
7244                 return ret;
7245         }
7246
7247         leaf = path->nodes[0];
7248         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7249                                      struct btrfs_extent_item);
7250         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7251         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7252         btrfs_set_extent_flags(leaf, extent_item,
7253                                flags | BTRFS_EXTENT_FLAG_DATA);
7254
7255         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7256         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7257         if (parent > 0) {
7258                 struct btrfs_shared_data_ref *ref;
7259                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7260                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7261                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7262         } else {
7263                 struct btrfs_extent_data_ref *ref;
7264                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7265                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7266                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7267                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7268                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7269         }
7270
7271         btrfs_mark_buffer_dirty(path->nodes[0]);
7272         btrfs_free_path(path);
7273
7274         /* Always set parent to 0 here since its exclusive anyway. */
7275         ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7276                                       ins->objectid, ins->offset,
7277                                       BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7278         if (ret)
7279                 return ret;
7280
7281         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7282         if (ret) { /* -ENOENT, logic error */
7283                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7284                         ins->objectid, ins->offset);
7285                 BUG();
7286         }
7287         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7288         return ret;
7289 }
7290
7291 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7292                                      struct btrfs_root *root,
7293                                      u64 parent, u64 root_objectid,
7294                                      u64 flags, struct btrfs_disk_key *key,
7295                                      int level, struct btrfs_key *ins,
7296                                      int no_quota)
7297 {
7298         int ret;
7299         struct btrfs_fs_info *fs_info = root->fs_info;
7300         struct btrfs_extent_item *extent_item;
7301         struct btrfs_tree_block_info *block_info;
7302         struct btrfs_extent_inline_ref *iref;
7303         struct btrfs_path *path;
7304         struct extent_buffer *leaf;
7305         u32 size = sizeof(*extent_item) + sizeof(*iref);
7306         u64 num_bytes = ins->offset;
7307         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7308                                                  SKINNY_METADATA);
7309
7310         if (!skinny_metadata)
7311                 size += sizeof(*block_info);
7312
7313         path = btrfs_alloc_path();
7314         if (!path) {
7315                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7316                                                    root->nodesize);
7317                 return -ENOMEM;
7318         }
7319
7320         path->leave_spinning = 1;
7321         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7322                                       ins, size);
7323         if (ret) {
7324                 btrfs_free_path(path);
7325                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7326                                                    root->nodesize);
7327                 return ret;
7328         }
7329
7330         leaf = path->nodes[0];
7331         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7332                                      struct btrfs_extent_item);
7333         btrfs_set_extent_refs(leaf, extent_item, 1);
7334         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7335         btrfs_set_extent_flags(leaf, extent_item,
7336                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7337
7338         if (skinny_metadata) {
7339                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7340                 num_bytes = root->nodesize;
7341         } else {
7342                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7343                 btrfs_set_tree_block_key(leaf, block_info, key);
7344                 btrfs_set_tree_block_level(leaf, block_info, level);
7345                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7346         }
7347
7348         if (parent > 0) {
7349                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7350                 btrfs_set_extent_inline_ref_type(leaf, iref,
7351                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7352                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7353         } else {
7354                 btrfs_set_extent_inline_ref_type(leaf, iref,
7355                                                  BTRFS_TREE_BLOCK_REF_KEY);
7356                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7357         }
7358
7359         btrfs_mark_buffer_dirty(leaf);
7360         btrfs_free_path(path);
7361
7362         if (!no_quota) {
7363                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7364                                               ins->objectid, num_bytes,
7365                                               BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7366                 if (ret)
7367                         return ret;
7368         }
7369
7370         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7371                                  1);
7372         if (ret) { /* -ENOENT, logic error */
7373                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7374                         ins->objectid, ins->offset);
7375                 BUG();
7376         }
7377
7378         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7379         return ret;
7380 }
7381
7382 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7383                                      struct btrfs_root *root,
7384                                      u64 root_objectid, u64 owner,
7385                                      u64 offset, struct btrfs_key *ins)
7386 {
7387         int ret;
7388
7389         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7390
7391         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7392                                          ins->offset, 0,
7393                                          root_objectid, owner, offset,
7394                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7395         return ret;
7396 }
7397
7398 /*
7399  * this is used by the tree logging recovery code.  It records that
7400  * an extent has been allocated and makes sure to clear the free
7401  * space cache bits as well
7402  */
7403 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7404                                    struct btrfs_root *root,
7405                                    u64 root_objectid, u64 owner, u64 offset,
7406                                    struct btrfs_key *ins)
7407 {
7408         int ret;
7409         struct btrfs_block_group_cache *block_group;
7410
7411         /*
7412          * Mixed block groups will exclude before processing the log so we only
7413          * need to do the exlude dance if this fs isn't mixed.
7414          */
7415         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7416                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7417                 if (ret)
7418                         return ret;
7419         }
7420
7421         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7422         if (!block_group)
7423                 return -EINVAL;
7424
7425         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7426                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7427         BUG_ON(ret); /* logic error */
7428         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7429                                          0, owner, offset, ins, 1);
7430         btrfs_put_block_group(block_group);
7431         return ret;
7432 }
7433
7434 static struct extent_buffer *
7435 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7436                       u64 bytenr, int level)
7437 {
7438         struct extent_buffer *buf;
7439
7440         buf = btrfs_find_create_tree_block(root, bytenr);
7441         if (!buf)
7442                 return ERR_PTR(-ENOMEM);
7443         btrfs_set_header_generation(buf, trans->transid);
7444         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7445         btrfs_tree_lock(buf);
7446         clean_tree_block(trans, root->fs_info, buf);
7447         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7448
7449         btrfs_set_lock_blocking(buf);
7450         btrfs_set_buffer_uptodate(buf);
7451
7452         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7453                 buf->log_index = root->log_transid % 2;
7454                 /*
7455                  * we allow two log transactions at a time, use different
7456                  * EXENT bit to differentiate dirty pages.
7457                  */
7458                 if (buf->log_index == 0)
7459                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7460                                         buf->start + buf->len - 1, GFP_NOFS);
7461                 else
7462                         set_extent_new(&root->dirty_log_pages, buf->start,
7463                                         buf->start + buf->len - 1, GFP_NOFS);
7464         } else {
7465                 buf->log_index = -1;
7466                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7467                          buf->start + buf->len - 1, GFP_NOFS);
7468         }
7469         trans->blocks_used++;
7470         /* this returns a buffer locked for blocking */
7471         return buf;
7472 }
7473
7474 static struct btrfs_block_rsv *
7475 use_block_rsv(struct btrfs_trans_handle *trans,
7476               struct btrfs_root *root, u32 blocksize)
7477 {
7478         struct btrfs_block_rsv *block_rsv;
7479         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7480         int ret;
7481         bool global_updated = false;
7482
7483         block_rsv = get_block_rsv(trans, root);
7484
7485         if (unlikely(block_rsv->size == 0))
7486                 goto try_reserve;
7487 again:
7488         ret = block_rsv_use_bytes(block_rsv, blocksize);
7489         if (!ret)
7490                 return block_rsv;
7491
7492         if (block_rsv->failfast)
7493                 return ERR_PTR(ret);
7494
7495         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7496                 global_updated = true;
7497                 update_global_block_rsv(root->fs_info);
7498                 goto again;
7499         }
7500
7501         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7502                 static DEFINE_RATELIMIT_STATE(_rs,
7503                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7504                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7505                 if (__ratelimit(&_rs))
7506                         WARN(1, KERN_DEBUG
7507                                 "BTRFS: block rsv returned %d\n", ret);
7508         }
7509 try_reserve:
7510         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7511                                      BTRFS_RESERVE_NO_FLUSH);
7512         if (!ret)
7513                 return block_rsv;
7514         /*
7515          * If we couldn't reserve metadata bytes try and use some from
7516          * the global reserve if its space type is the same as the global
7517          * reservation.
7518          */
7519         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7520             block_rsv->space_info == global_rsv->space_info) {
7521                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7522                 if (!ret)
7523                         return global_rsv;
7524         }
7525         return ERR_PTR(ret);
7526 }
7527
7528 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7529                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7530 {
7531         block_rsv_add_bytes(block_rsv, blocksize, 0);
7532         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7533 }
7534
7535 /*
7536  * finds a free extent and does all the dirty work required for allocation
7537  * returns the key for the extent through ins, and a tree buffer for
7538  * the first block of the extent through buf.
7539  *
7540  * returns the tree buffer or NULL.
7541  */
7542 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7543                                         struct btrfs_root *root,
7544                                         u64 parent, u64 root_objectid,
7545                                         struct btrfs_disk_key *key, int level,
7546                                         u64 hint, u64 empty_size)
7547 {
7548         struct btrfs_key ins;
7549         struct btrfs_block_rsv *block_rsv;
7550         struct extent_buffer *buf;
7551         u64 flags = 0;
7552         int ret;
7553         u32 blocksize = root->nodesize;
7554         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7555                                                  SKINNY_METADATA);
7556
7557         if (btrfs_test_is_dummy_root(root)) {
7558                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7559                                             level);
7560                 if (!IS_ERR(buf))
7561                         root->alloc_bytenr += blocksize;
7562                 return buf;
7563         }
7564
7565         block_rsv = use_block_rsv(trans, root, blocksize);
7566         if (IS_ERR(block_rsv))
7567                 return ERR_CAST(block_rsv);
7568
7569         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7570                                    empty_size, hint, &ins, 0, 0);
7571         if (ret) {
7572                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7573                 return ERR_PTR(ret);
7574         }
7575
7576         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
7577         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
7578
7579         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7580                 if (parent == 0)
7581                         parent = ins.objectid;
7582                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7583         } else
7584                 BUG_ON(parent > 0);
7585
7586         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7587                 struct btrfs_delayed_extent_op *extent_op;
7588                 extent_op = btrfs_alloc_delayed_extent_op();
7589                 BUG_ON(!extent_op); /* -ENOMEM */
7590                 if (key)
7591                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7592                 else
7593                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7594                 extent_op->flags_to_set = flags;
7595                 if (skinny_metadata)
7596                         extent_op->update_key = 0;
7597                 else
7598                         extent_op->update_key = 1;
7599                 extent_op->update_flags = 1;
7600                 extent_op->is_data = 0;
7601                 extent_op->level = level;
7602
7603                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7604                                         ins.objectid,
7605                                         ins.offset, parent, root_objectid,
7606                                         level, BTRFS_ADD_DELAYED_EXTENT,
7607                                         extent_op, 0);
7608                 BUG_ON(ret); /* -ENOMEM */
7609         }
7610         return buf;
7611 }
7612
7613 struct walk_control {
7614         u64 refs[BTRFS_MAX_LEVEL];
7615         u64 flags[BTRFS_MAX_LEVEL];
7616         struct btrfs_key update_progress;
7617         int stage;
7618         int level;
7619         int shared_level;
7620         int update_ref;
7621         int keep_locks;
7622         int reada_slot;
7623         int reada_count;
7624         int for_reloc;
7625 };
7626
7627 #define DROP_REFERENCE  1
7628 #define UPDATE_BACKREF  2
7629
7630 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7631                                      struct btrfs_root *root,
7632                                      struct walk_control *wc,
7633                                      struct btrfs_path *path)
7634 {
7635         u64 bytenr;
7636         u64 generation;
7637         u64 refs;
7638         u64 flags;
7639         u32 nritems;
7640         u32 blocksize;
7641         struct btrfs_key key;
7642         struct extent_buffer *eb;
7643         int ret;
7644         int slot;
7645         int nread = 0;
7646
7647         if (path->slots[wc->level] < wc->reada_slot) {
7648                 wc->reada_count = wc->reada_count * 2 / 3;
7649                 wc->reada_count = max(wc->reada_count, 2);
7650         } else {
7651                 wc->reada_count = wc->reada_count * 3 / 2;
7652                 wc->reada_count = min_t(int, wc->reada_count,
7653                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7654         }
7655
7656         eb = path->nodes[wc->level];
7657         nritems = btrfs_header_nritems(eb);
7658         blocksize = root->nodesize;
7659
7660         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7661                 if (nread >= wc->reada_count)
7662                         break;
7663
7664                 cond_resched();
7665                 bytenr = btrfs_node_blockptr(eb, slot);
7666                 generation = btrfs_node_ptr_generation(eb, slot);
7667
7668                 if (slot == path->slots[wc->level])
7669                         goto reada;
7670
7671                 if (wc->stage == UPDATE_BACKREF &&
7672                     generation <= root->root_key.offset)
7673                         continue;
7674
7675                 /* We don't lock the tree block, it's OK to be racy here */
7676                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7677                                                wc->level - 1, 1, &refs,
7678                                                &flags);
7679                 /* We don't care about errors in readahead. */
7680                 if (ret < 0)
7681                         continue;
7682                 BUG_ON(refs == 0);
7683
7684                 if (wc->stage == DROP_REFERENCE) {
7685                         if (refs == 1)
7686                                 goto reada;
7687
7688                         if (wc->level == 1 &&
7689                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7690                                 continue;
7691                         if (!wc->update_ref ||
7692                             generation <= root->root_key.offset)
7693                                 continue;
7694                         btrfs_node_key_to_cpu(eb, &key, slot);
7695                         ret = btrfs_comp_cpu_keys(&key,
7696                                                   &wc->update_progress);
7697                         if (ret < 0)
7698                                 continue;
7699                 } else {
7700                         if (wc->level == 1 &&
7701                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7702                                 continue;
7703                 }
7704 reada:
7705                 readahead_tree_block(root, bytenr);
7706                 nread++;
7707         }
7708         wc->reada_slot = slot;
7709 }
7710
7711 static int account_leaf_items(struct btrfs_trans_handle *trans,
7712                               struct btrfs_root *root,
7713                               struct extent_buffer *eb)
7714 {
7715         int nr = btrfs_header_nritems(eb);
7716         int i, extent_type, ret;
7717         struct btrfs_key key;
7718         struct btrfs_file_extent_item *fi;
7719         u64 bytenr, num_bytes;
7720
7721         for (i = 0; i < nr; i++) {
7722                 btrfs_item_key_to_cpu(eb, &key, i);
7723
7724                 if (key.type != BTRFS_EXTENT_DATA_KEY)
7725                         continue;
7726
7727                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
7728                 /* filter out non qgroup-accountable extents  */
7729                 extent_type = btrfs_file_extent_type(eb, fi);
7730
7731                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
7732                         continue;
7733
7734                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
7735                 if (!bytenr)
7736                         continue;
7737
7738                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
7739
7740                 ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7741                                               root->objectid,
7742                                               bytenr, num_bytes,
7743                                               BTRFS_QGROUP_OPER_SUB_SUBTREE, 0);
7744                 if (ret)
7745                         return ret;
7746         }
7747         return 0;
7748 }
7749
7750 /*
7751  * Walk up the tree from the bottom, freeing leaves and any interior
7752  * nodes which have had all slots visited. If a node (leaf or
7753  * interior) is freed, the node above it will have it's slot
7754  * incremented. The root node will never be freed.
7755  *
7756  * At the end of this function, we should have a path which has all
7757  * slots incremented to the next position for a search. If we need to
7758  * read a new node it will be NULL and the node above it will have the
7759  * correct slot selected for a later read.
7760  *
7761  * If we increment the root nodes slot counter past the number of
7762  * elements, 1 is returned to signal completion of the search.
7763  */
7764 static int adjust_slots_upwards(struct btrfs_root *root,
7765                                 struct btrfs_path *path, int root_level)
7766 {
7767         int level = 0;
7768         int nr, slot;
7769         struct extent_buffer *eb;
7770
7771         if (root_level == 0)
7772                 return 1;
7773
7774         while (level <= root_level) {
7775                 eb = path->nodes[level];
7776                 nr = btrfs_header_nritems(eb);
7777                 path->slots[level]++;
7778                 slot = path->slots[level];
7779                 if (slot >= nr || level == 0) {
7780                         /*
7781                          * Don't free the root -  we will detect this
7782                          * condition after our loop and return a
7783                          * positive value for caller to stop walking the tree.
7784                          */
7785                         if (level != root_level) {
7786                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7787                                 path->locks[level] = 0;
7788
7789                                 free_extent_buffer(eb);
7790                                 path->nodes[level] = NULL;
7791                                 path->slots[level] = 0;
7792                         }
7793                 } else {
7794                         /*
7795                          * We have a valid slot to walk back down
7796                          * from. Stop here so caller can process these
7797                          * new nodes.
7798                          */
7799                         break;
7800                 }
7801
7802                 level++;
7803         }
7804
7805         eb = path->nodes[root_level];
7806         if (path->slots[root_level] >= btrfs_header_nritems(eb))
7807                 return 1;
7808
7809         return 0;
7810 }
7811
7812 /*
7813  * root_eb is the subtree root and is locked before this function is called.
7814  */
7815 static int account_shared_subtree(struct btrfs_trans_handle *trans,
7816                                   struct btrfs_root *root,
7817                                   struct extent_buffer *root_eb,
7818                                   u64 root_gen,
7819                                   int root_level)
7820 {
7821         int ret = 0;
7822         int level;
7823         struct extent_buffer *eb = root_eb;
7824         struct btrfs_path *path = NULL;
7825
7826         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
7827         BUG_ON(root_eb == NULL);
7828
7829         if (!root->fs_info->quota_enabled)
7830                 return 0;
7831
7832         if (!extent_buffer_uptodate(root_eb)) {
7833                 ret = btrfs_read_buffer(root_eb, root_gen);
7834                 if (ret)
7835                         goto out;
7836         }
7837
7838         if (root_level == 0) {
7839                 ret = account_leaf_items(trans, root, root_eb);
7840                 goto out;
7841         }
7842
7843         path = btrfs_alloc_path();
7844         if (!path)
7845                 return -ENOMEM;
7846
7847         /*
7848          * Walk down the tree.  Missing extent blocks are filled in as
7849          * we go. Metadata is accounted every time we read a new
7850          * extent block.
7851          *
7852          * When we reach a leaf, we account for file extent items in it,
7853          * walk back up the tree (adjusting slot pointers as we go)
7854          * and restart the search process.
7855          */
7856         extent_buffer_get(root_eb); /* For path */
7857         path->nodes[root_level] = root_eb;
7858         path->slots[root_level] = 0;
7859         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
7860 walk_down:
7861         level = root_level;
7862         while (level >= 0) {
7863                 if (path->nodes[level] == NULL) {
7864                         int parent_slot;
7865                         u64 child_gen;
7866                         u64 child_bytenr;
7867
7868                         /* We need to get child blockptr/gen from
7869                          * parent before we can read it. */
7870                         eb = path->nodes[level + 1];
7871                         parent_slot = path->slots[level + 1];
7872                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
7873                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
7874
7875                         eb = read_tree_block(root, child_bytenr, child_gen);
7876                         if (!eb || !extent_buffer_uptodate(eb)) {
7877                                 ret = -EIO;
7878                                 goto out;
7879                         }
7880
7881                         path->nodes[level] = eb;
7882                         path->slots[level] = 0;
7883
7884                         btrfs_tree_read_lock(eb);
7885                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
7886                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
7887
7888                         ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7889                                                 root->objectid,
7890                                                 child_bytenr,
7891                                                 root->nodesize,
7892                                                 BTRFS_QGROUP_OPER_SUB_SUBTREE,
7893                                                 0);
7894                         if (ret)
7895                                 goto out;
7896
7897                 }
7898
7899                 if (level == 0) {
7900                         ret = account_leaf_items(trans, root, path->nodes[level]);
7901                         if (ret)
7902                                 goto out;
7903
7904                         /* Nonzero return here means we completed our search */
7905                         ret = adjust_slots_upwards(root, path, root_level);
7906                         if (ret)
7907                                 break;
7908
7909                         /* Restart search with new slots */
7910                         goto walk_down;
7911                 }
7912
7913                 level--;
7914         }
7915
7916         ret = 0;
7917 out:
7918         btrfs_free_path(path);
7919
7920         return ret;
7921 }
7922
7923 /*
7924  * helper to process tree block while walking down the tree.
7925  *
7926  * when wc->stage == UPDATE_BACKREF, this function updates
7927  * back refs for pointers in the block.
7928  *
7929  * NOTE: return value 1 means we should stop walking down.
7930  */
7931 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7932                                    struct btrfs_root *root,
7933                                    struct btrfs_path *path,
7934                                    struct walk_control *wc, int lookup_info)
7935 {
7936         int level = wc->level;
7937         struct extent_buffer *eb = path->nodes[level];
7938         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7939         int ret;
7940
7941         if (wc->stage == UPDATE_BACKREF &&
7942             btrfs_header_owner(eb) != root->root_key.objectid)
7943                 return 1;
7944
7945         /*
7946          * when reference count of tree block is 1, it won't increase
7947          * again. once full backref flag is set, we never clear it.
7948          */
7949         if (lookup_info &&
7950             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7951              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7952                 BUG_ON(!path->locks[level]);
7953                 ret = btrfs_lookup_extent_info(trans, root,
7954                                                eb->start, level, 1,
7955                                                &wc->refs[level],
7956                                                &wc->flags[level]);
7957                 BUG_ON(ret == -ENOMEM);
7958                 if (ret)
7959                         return ret;
7960                 BUG_ON(wc->refs[level] == 0);
7961         }
7962
7963         if (wc->stage == DROP_REFERENCE) {
7964                 if (wc->refs[level] > 1)
7965                         return 1;
7966
7967                 if (path->locks[level] && !wc->keep_locks) {
7968                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7969                         path->locks[level] = 0;
7970                 }
7971                 return 0;
7972         }
7973
7974         /* wc->stage == UPDATE_BACKREF */
7975         if (!(wc->flags[level] & flag)) {
7976                 BUG_ON(!path->locks[level]);
7977                 ret = btrfs_inc_ref(trans, root, eb, 1);
7978                 BUG_ON(ret); /* -ENOMEM */
7979                 ret = btrfs_dec_ref(trans, root, eb, 0);
7980                 BUG_ON(ret); /* -ENOMEM */
7981                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7982                                                   eb->len, flag,
7983                                                   btrfs_header_level(eb), 0);
7984                 BUG_ON(ret); /* -ENOMEM */
7985                 wc->flags[level] |= flag;
7986         }
7987
7988         /*
7989          * the block is shared by multiple trees, so it's not good to
7990          * keep the tree lock
7991          */
7992         if (path->locks[level] && level > 0) {
7993                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7994                 path->locks[level] = 0;
7995         }
7996         return 0;
7997 }
7998
7999 /*
8000  * helper to process tree block pointer.
8001  *
8002  * when wc->stage == DROP_REFERENCE, this function checks
8003  * reference count of the block pointed to. if the block
8004  * is shared and we need update back refs for the subtree
8005  * rooted at the block, this function changes wc->stage to
8006  * UPDATE_BACKREF. if the block is shared and there is no
8007  * need to update back, this function drops the reference
8008  * to the block.
8009  *
8010  * NOTE: return value 1 means we should stop walking down.
8011  */
8012 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8013                                  struct btrfs_root *root,
8014                                  struct btrfs_path *path,
8015                                  struct walk_control *wc, int *lookup_info)
8016 {
8017         u64 bytenr;
8018         u64 generation;
8019         u64 parent;
8020         u32 blocksize;
8021         struct btrfs_key key;
8022         struct extent_buffer *next;
8023         int level = wc->level;
8024         int reada = 0;
8025         int ret = 0;
8026         bool need_account = false;
8027
8028         generation = btrfs_node_ptr_generation(path->nodes[level],
8029                                                path->slots[level]);
8030         /*
8031          * if the lower level block was created before the snapshot
8032          * was created, we know there is no need to update back refs
8033          * for the subtree
8034          */
8035         if (wc->stage == UPDATE_BACKREF &&
8036             generation <= root->root_key.offset) {
8037                 *lookup_info = 1;
8038                 return 1;
8039         }
8040
8041         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8042         blocksize = root->nodesize;
8043
8044         next = btrfs_find_tree_block(root->fs_info, bytenr);
8045         if (!next) {
8046                 next = btrfs_find_create_tree_block(root, bytenr);
8047                 if (!next)
8048                         return -ENOMEM;
8049                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8050                                                level - 1);
8051                 reada = 1;
8052         }
8053         btrfs_tree_lock(next);
8054         btrfs_set_lock_blocking(next);
8055
8056         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8057                                        &wc->refs[level - 1],
8058                                        &wc->flags[level - 1]);
8059         if (ret < 0) {
8060                 btrfs_tree_unlock(next);
8061                 return ret;
8062         }
8063
8064         if (unlikely(wc->refs[level - 1] == 0)) {
8065                 btrfs_err(root->fs_info, "Missing references.");
8066                 BUG();
8067         }
8068         *lookup_info = 0;
8069
8070         if (wc->stage == DROP_REFERENCE) {
8071                 if (wc->refs[level - 1] > 1) {
8072                         need_account = true;
8073                         if (level == 1 &&
8074                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8075                                 goto skip;
8076
8077                         if (!wc->update_ref ||
8078                             generation <= root->root_key.offset)
8079                                 goto skip;
8080
8081                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8082                                               path->slots[level]);
8083                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8084                         if (ret < 0)
8085                                 goto skip;
8086
8087                         wc->stage = UPDATE_BACKREF;
8088                         wc->shared_level = level - 1;
8089                 }
8090         } else {
8091                 if (level == 1 &&
8092                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8093                         goto skip;
8094         }
8095
8096         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8097                 btrfs_tree_unlock(next);
8098                 free_extent_buffer(next);
8099                 next = NULL;
8100                 *lookup_info = 1;
8101         }
8102
8103         if (!next) {
8104                 if (reada && level == 1)
8105                         reada_walk_down(trans, root, wc, path);
8106                 next = read_tree_block(root, bytenr, generation);
8107                 if (!next || !extent_buffer_uptodate(next)) {
8108                         free_extent_buffer(next);
8109                         return -EIO;
8110                 }
8111                 btrfs_tree_lock(next);
8112                 btrfs_set_lock_blocking(next);
8113         }
8114
8115         level--;
8116         BUG_ON(level != btrfs_header_level(next));
8117         path->nodes[level] = next;
8118         path->slots[level] = 0;
8119         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8120         wc->level = level;
8121         if (wc->level == 1)
8122                 wc->reada_slot = 0;
8123         return 0;
8124 skip:
8125         wc->refs[level - 1] = 0;
8126         wc->flags[level - 1] = 0;
8127         if (wc->stage == DROP_REFERENCE) {
8128                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8129                         parent = path->nodes[level]->start;
8130                 } else {
8131                         BUG_ON(root->root_key.objectid !=
8132                                btrfs_header_owner(path->nodes[level]));
8133                         parent = 0;
8134                 }
8135
8136                 if (need_account) {
8137                         ret = account_shared_subtree(trans, root, next,
8138                                                      generation, level - 1);
8139                         if (ret) {
8140                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8141                                         "%d accounting shared subtree. Quota "
8142                                         "is out of sync, rescan required.\n",
8143                                         root->fs_info->sb->s_id, ret);
8144                         }
8145                 }
8146                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8147                                 root->root_key.objectid, level - 1, 0, 0);
8148                 BUG_ON(ret); /* -ENOMEM */
8149         }
8150         btrfs_tree_unlock(next);
8151         free_extent_buffer(next);
8152         *lookup_info = 1;
8153         return 1;
8154 }
8155
8156 /*
8157  * helper to process tree block while walking up the tree.
8158  *
8159  * when wc->stage == DROP_REFERENCE, this function drops
8160  * reference count on the block.
8161  *
8162  * when wc->stage == UPDATE_BACKREF, this function changes
8163  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8164  * to UPDATE_BACKREF previously while processing the block.
8165  *
8166  * NOTE: return value 1 means we should stop walking up.
8167  */
8168 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8169                                  struct btrfs_root *root,
8170                                  struct btrfs_path *path,
8171                                  struct walk_control *wc)
8172 {
8173         int ret;
8174         int level = wc->level;
8175         struct extent_buffer *eb = path->nodes[level];
8176         u64 parent = 0;
8177
8178         if (wc->stage == UPDATE_BACKREF) {
8179                 BUG_ON(wc->shared_level < level);
8180                 if (level < wc->shared_level)
8181                         goto out;
8182
8183                 ret = find_next_key(path, level + 1, &wc->update_progress);
8184                 if (ret > 0)
8185                         wc->update_ref = 0;
8186
8187                 wc->stage = DROP_REFERENCE;
8188                 wc->shared_level = -1;
8189                 path->slots[level] = 0;
8190
8191                 /*
8192                  * check reference count again if the block isn't locked.
8193                  * we should start walking down the tree again if reference
8194                  * count is one.
8195                  */
8196                 if (!path->locks[level]) {
8197                         BUG_ON(level == 0);
8198                         btrfs_tree_lock(eb);
8199                         btrfs_set_lock_blocking(eb);
8200                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8201
8202                         ret = btrfs_lookup_extent_info(trans, root,
8203                                                        eb->start, level, 1,
8204                                                        &wc->refs[level],
8205                                                        &wc->flags[level]);
8206                         if (ret < 0) {
8207                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8208                                 path->locks[level] = 0;
8209                                 return ret;
8210                         }
8211                         BUG_ON(wc->refs[level] == 0);
8212                         if (wc->refs[level] == 1) {
8213                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8214                                 path->locks[level] = 0;
8215                                 return 1;
8216                         }
8217                 }
8218         }
8219
8220         /* wc->stage == DROP_REFERENCE */
8221         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8222
8223         if (wc->refs[level] == 1) {
8224                 if (level == 0) {
8225                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8226                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8227                         else
8228                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8229                         BUG_ON(ret); /* -ENOMEM */
8230                         ret = account_leaf_items(trans, root, eb);
8231                         if (ret) {
8232                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8233                                         "%d accounting leaf items. Quota "
8234                                         "is out of sync, rescan required.\n",
8235                                         root->fs_info->sb->s_id, ret);
8236                         }
8237                 }
8238                 /* make block locked assertion in clean_tree_block happy */
8239                 if (!path->locks[level] &&
8240                     btrfs_header_generation(eb) == trans->transid) {
8241                         btrfs_tree_lock(eb);
8242                         btrfs_set_lock_blocking(eb);
8243                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8244                 }
8245                 clean_tree_block(trans, root->fs_info, eb);
8246         }
8247
8248         if (eb == root->node) {
8249                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8250                         parent = eb->start;
8251                 else
8252                         BUG_ON(root->root_key.objectid !=
8253                                btrfs_header_owner(eb));
8254         } else {
8255                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8256                         parent = path->nodes[level + 1]->start;
8257                 else
8258                         BUG_ON(root->root_key.objectid !=
8259                                btrfs_header_owner(path->nodes[level + 1]));
8260         }
8261
8262         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8263 out:
8264         wc->refs[level] = 0;
8265         wc->flags[level] = 0;
8266         return 0;
8267 }
8268
8269 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8270                                    struct btrfs_root *root,
8271                                    struct btrfs_path *path,
8272                                    struct walk_control *wc)
8273 {
8274         int level = wc->level;
8275         int lookup_info = 1;
8276         int ret;
8277
8278         while (level >= 0) {
8279                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8280                 if (ret > 0)
8281                         break;
8282
8283                 if (level == 0)
8284                         break;
8285
8286                 if (path->slots[level] >=
8287                     btrfs_header_nritems(path->nodes[level]))
8288                         break;
8289
8290                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8291                 if (ret > 0) {
8292                         path->slots[level]++;
8293                         continue;
8294                 } else if (ret < 0)
8295                         return ret;
8296                 level = wc->level;
8297         }
8298         return 0;
8299 }
8300
8301 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8302                                  struct btrfs_root *root,
8303                                  struct btrfs_path *path,
8304                                  struct walk_control *wc, int max_level)
8305 {
8306         int level = wc->level;
8307         int ret;
8308
8309         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8310         while (level < max_level && path->nodes[level]) {
8311                 wc->level = level;
8312                 if (path->slots[level] + 1 <
8313                     btrfs_header_nritems(path->nodes[level])) {
8314                         path->slots[level]++;
8315                         return 0;
8316                 } else {
8317                         ret = walk_up_proc(trans, root, path, wc);
8318                         if (ret > 0)
8319                                 return 0;
8320
8321                         if (path->locks[level]) {
8322                                 btrfs_tree_unlock_rw(path->nodes[level],
8323                                                      path->locks[level]);
8324                                 path->locks[level] = 0;
8325                         }
8326                         free_extent_buffer(path->nodes[level]);
8327                         path->nodes[level] = NULL;
8328                         level++;
8329                 }
8330         }
8331         return 1;
8332 }
8333
8334 /*
8335  * drop a subvolume tree.
8336  *
8337  * this function traverses the tree freeing any blocks that only
8338  * referenced by the tree.
8339  *
8340  * when a shared tree block is found. this function decreases its
8341  * reference count by one. if update_ref is true, this function
8342  * also make sure backrefs for the shared block and all lower level
8343  * blocks are properly updated.
8344  *
8345  * If called with for_reloc == 0, may exit early with -EAGAIN
8346  */
8347 int btrfs_drop_snapshot(struct btrfs_root *root,
8348                          struct btrfs_block_rsv *block_rsv, int update_ref,
8349                          int for_reloc)
8350 {
8351         struct btrfs_path *path;
8352         struct btrfs_trans_handle *trans;
8353         struct btrfs_root *tree_root = root->fs_info->tree_root;
8354         struct btrfs_root_item *root_item = &root->root_item;
8355         struct walk_control *wc;
8356         struct btrfs_key key;
8357         int err = 0;
8358         int ret;
8359         int level;
8360         bool root_dropped = false;
8361
8362         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8363
8364         path = btrfs_alloc_path();
8365         if (!path) {
8366                 err = -ENOMEM;
8367                 goto out;
8368         }
8369
8370         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8371         if (!wc) {
8372                 btrfs_free_path(path);
8373                 err = -ENOMEM;
8374                 goto out;
8375         }
8376
8377         trans = btrfs_start_transaction(tree_root, 0);
8378         if (IS_ERR(trans)) {
8379                 err = PTR_ERR(trans);
8380                 goto out_free;
8381         }
8382
8383         if (block_rsv)
8384                 trans->block_rsv = block_rsv;
8385
8386         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8387                 level = btrfs_header_level(root->node);
8388                 path->nodes[level] = btrfs_lock_root_node(root);
8389                 btrfs_set_lock_blocking(path->nodes[level]);
8390                 path->slots[level] = 0;
8391                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8392                 memset(&wc->update_progress, 0,
8393                        sizeof(wc->update_progress));
8394         } else {
8395                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8396                 memcpy(&wc->update_progress, &key,
8397                        sizeof(wc->update_progress));
8398
8399                 level = root_item->drop_level;
8400                 BUG_ON(level == 0);
8401                 path->lowest_level = level;
8402                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8403                 path->lowest_level = 0;
8404                 if (ret < 0) {
8405                         err = ret;
8406                         goto out_end_trans;
8407                 }
8408                 WARN_ON(ret > 0);
8409
8410                 /*
8411                  * unlock our path, this is safe because only this
8412                  * function is allowed to delete this snapshot
8413                  */
8414                 btrfs_unlock_up_safe(path, 0);
8415
8416                 level = btrfs_header_level(root->node);
8417                 while (1) {
8418                         btrfs_tree_lock(path->nodes[level]);
8419                         btrfs_set_lock_blocking(path->nodes[level]);
8420                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8421
8422                         ret = btrfs_lookup_extent_info(trans, root,
8423                                                 path->nodes[level]->start,
8424                                                 level, 1, &wc->refs[level],
8425                                                 &wc->flags[level]);
8426                         if (ret < 0) {
8427                                 err = ret;
8428                                 goto out_end_trans;
8429                         }
8430                         BUG_ON(wc->refs[level] == 0);
8431
8432                         if (level == root_item->drop_level)
8433                                 break;
8434
8435                         btrfs_tree_unlock(path->nodes[level]);
8436                         path->locks[level] = 0;
8437                         WARN_ON(wc->refs[level] != 1);
8438                         level--;
8439                 }
8440         }
8441
8442         wc->level = level;
8443         wc->shared_level = -1;
8444         wc->stage = DROP_REFERENCE;
8445         wc->update_ref = update_ref;
8446         wc->keep_locks = 0;
8447         wc->for_reloc = for_reloc;
8448         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8449
8450         while (1) {
8451
8452                 ret = walk_down_tree(trans, root, path, wc);
8453                 if (ret < 0) {
8454                         err = ret;
8455                         break;
8456                 }
8457
8458                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8459                 if (ret < 0) {
8460                         err = ret;
8461                         break;
8462                 }
8463
8464                 if (ret > 0) {
8465                         BUG_ON(wc->stage != DROP_REFERENCE);
8466                         break;
8467                 }
8468
8469                 if (wc->stage == DROP_REFERENCE) {
8470                         level = wc->level;
8471                         btrfs_node_key(path->nodes[level],
8472                                        &root_item->drop_progress,
8473                                        path->slots[level]);
8474                         root_item->drop_level = level;
8475                 }
8476
8477                 BUG_ON(wc->level == 0);
8478                 if (btrfs_should_end_transaction(trans, tree_root) ||
8479                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8480                         ret = btrfs_update_root(trans, tree_root,
8481                                                 &root->root_key,
8482                                                 root_item);
8483                         if (ret) {
8484                                 btrfs_abort_transaction(trans, tree_root, ret);
8485                                 err = ret;
8486                                 goto out_end_trans;
8487                         }
8488
8489                         /*
8490                          * Qgroup update accounting is run from
8491                          * delayed ref handling. This usually works
8492                          * out because delayed refs are normally the
8493                          * only way qgroup updates are added. However,
8494                          * we may have added updates during our tree
8495                          * walk so run qgroups here to make sure we
8496                          * don't lose any updates.
8497                          */
8498                         ret = btrfs_delayed_qgroup_accounting(trans,
8499                                                               root->fs_info);
8500                         if (ret)
8501                                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8502                                                    "running qgroup updates "
8503                                                    "during snapshot delete. "
8504                                                    "Quota is out of sync, "
8505                                                    "rescan required.\n", ret);
8506
8507                         btrfs_end_transaction_throttle(trans, tree_root);
8508                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8509                                 pr_debug("BTRFS: drop snapshot early exit\n");
8510                                 err = -EAGAIN;
8511                                 goto out_free;
8512                         }
8513
8514                         trans = btrfs_start_transaction(tree_root, 0);
8515                         if (IS_ERR(trans)) {
8516                                 err = PTR_ERR(trans);
8517                                 goto out_free;
8518                         }
8519                         if (block_rsv)
8520                                 trans->block_rsv = block_rsv;
8521                 }
8522         }
8523         btrfs_release_path(path);
8524         if (err)
8525                 goto out_end_trans;
8526
8527         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8528         if (ret) {
8529                 btrfs_abort_transaction(trans, tree_root, ret);
8530                 goto out_end_trans;
8531         }
8532
8533         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8534                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8535                                       NULL, NULL);
8536                 if (ret < 0) {
8537                         btrfs_abort_transaction(trans, tree_root, ret);
8538                         err = ret;
8539                         goto out_end_trans;
8540                 } else if (ret > 0) {
8541                         /* if we fail to delete the orphan item this time
8542                          * around, it'll get picked up the next time.
8543                          *
8544                          * The most common failure here is just -ENOENT.
8545                          */
8546                         btrfs_del_orphan_item(trans, tree_root,
8547                                               root->root_key.objectid);
8548                 }
8549         }
8550
8551         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8552                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
8553         } else {
8554                 free_extent_buffer(root->node);
8555                 free_extent_buffer(root->commit_root);
8556                 btrfs_put_fs_root(root);
8557         }
8558         root_dropped = true;
8559 out_end_trans:
8560         ret = btrfs_delayed_qgroup_accounting(trans, tree_root->fs_info);
8561         if (ret)
8562                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8563                                    "running qgroup updates "
8564                                    "during snapshot delete. "
8565                                    "Quota is out of sync, "
8566                                    "rescan required.\n", ret);
8567
8568         btrfs_end_transaction_throttle(trans, tree_root);
8569 out_free:
8570         kfree(wc);
8571         btrfs_free_path(path);
8572 out:
8573         /*
8574          * So if we need to stop dropping the snapshot for whatever reason we
8575          * need to make sure to add it back to the dead root list so that we
8576          * keep trying to do the work later.  This also cleans up roots if we
8577          * don't have it in the radix (like when we recover after a power fail
8578          * or unmount) so we don't leak memory.
8579          */
8580         if (!for_reloc && root_dropped == false)
8581                 btrfs_add_dead_root(root);
8582         if (err && err != -EAGAIN)
8583                 btrfs_std_error(root->fs_info, err);
8584         return err;
8585 }
8586
8587 /*
8588  * drop subtree rooted at tree block 'node'.
8589  *
8590  * NOTE: this function will unlock and release tree block 'node'
8591  * only used by relocation code
8592  */
8593 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8594                         struct btrfs_root *root,
8595                         struct extent_buffer *node,
8596                         struct extent_buffer *parent)
8597 {
8598         struct btrfs_path *path;
8599         struct walk_control *wc;
8600         int level;
8601         int parent_level;
8602         int ret = 0;
8603         int wret;
8604
8605         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8606
8607         path = btrfs_alloc_path();
8608         if (!path)
8609                 return -ENOMEM;
8610
8611         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8612         if (!wc) {
8613                 btrfs_free_path(path);
8614                 return -ENOMEM;
8615         }
8616
8617         btrfs_assert_tree_locked(parent);
8618         parent_level = btrfs_header_level(parent);
8619         extent_buffer_get(parent);
8620         path->nodes[parent_level] = parent;
8621         path->slots[parent_level] = btrfs_header_nritems(parent);
8622
8623         btrfs_assert_tree_locked(node);
8624         level = btrfs_header_level(node);
8625         path->nodes[level] = node;
8626         path->slots[level] = 0;
8627         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8628
8629         wc->refs[parent_level] = 1;
8630         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8631         wc->level = level;
8632         wc->shared_level = -1;
8633         wc->stage = DROP_REFERENCE;
8634         wc->update_ref = 0;
8635         wc->keep_locks = 1;
8636         wc->for_reloc = 1;
8637         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8638
8639         while (1) {
8640                 wret = walk_down_tree(trans, root, path, wc);
8641                 if (wret < 0) {
8642                         ret = wret;
8643                         break;
8644                 }
8645
8646                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8647                 if (wret < 0)
8648                         ret = wret;
8649                 if (wret != 0)
8650                         break;
8651         }
8652
8653         kfree(wc);
8654         btrfs_free_path(path);
8655         return ret;
8656 }
8657
8658 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8659 {
8660         u64 num_devices;
8661         u64 stripped;
8662
8663         /*
8664          * if restripe for this chunk_type is on pick target profile and
8665          * return, otherwise do the usual balance
8666          */
8667         stripped = get_restripe_target(root->fs_info, flags);
8668         if (stripped)
8669                 return extended_to_chunk(stripped);
8670
8671         num_devices = root->fs_info->fs_devices->rw_devices;
8672
8673         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8674                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8675                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8676
8677         if (num_devices == 1) {
8678                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8679                 stripped = flags & ~stripped;
8680
8681                 /* turn raid0 into single device chunks */
8682                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8683                         return stripped;
8684
8685                 /* turn mirroring into duplication */
8686                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8687                              BTRFS_BLOCK_GROUP_RAID10))
8688                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8689         } else {
8690                 /* they already had raid on here, just return */
8691                 if (flags & stripped)
8692                         return flags;
8693
8694                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8695                 stripped = flags & ~stripped;
8696
8697                 /* switch duplicated blocks with raid1 */
8698                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8699                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8700
8701                 /* this is drive concat, leave it alone */
8702         }
8703
8704         return flags;
8705 }
8706
8707 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8708 {
8709         struct btrfs_space_info *sinfo = cache->space_info;
8710         u64 num_bytes;
8711         u64 min_allocable_bytes;
8712         int ret = -ENOSPC;
8713
8714
8715         /*
8716          * We need some metadata space and system metadata space for
8717          * allocating chunks in some corner cases until we force to set
8718          * it to be readonly.
8719          */
8720         if ((sinfo->flags &
8721              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8722             !force)
8723                 min_allocable_bytes = 1 * 1024 * 1024;
8724         else
8725                 min_allocable_bytes = 0;
8726
8727         spin_lock(&sinfo->lock);
8728         spin_lock(&cache->lock);
8729
8730         if (cache->ro) {
8731                 ret = 0;
8732                 goto out;
8733         }
8734
8735         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8736                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8737
8738         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8739             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8740             min_allocable_bytes <= sinfo->total_bytes) {
8741                 sinfo->bytes_readonly += num_bytes;
8742                 cache->ro = 1;
8743                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
8744                 ret = 0;
8745         }
8746 out:
8747         spin_unlock(&cache->lock);
8748         spin_unlock(&sinfo->lock);
8749         return ret;
8750 }
8751
8752 int btrfs_set_block_group_ro(struct btrfs_root *root,
8753                              struct btrfs_block_group_cache *cache)
8754
8755 {
8756         struct btrfs_trans_handle *trans;
8757         u64 alloc_flags;
8758         int ret;
8759
8760         BUG_ON(cache->ro);
8761
8762 again:
8763         trans = btrfs_join_transaction(root);
8764         if (IS_ERR(trans))
8765                 return PTR_ERR(trans);
8766
8767         /*
8768          * we're not allowed to set block groups readonly after the dirty
8769          * block groups cache has started writing.  If it already started,
8770          * back off and let this transaction commit
8771          */
8772         mutex_lock(&root->fs_info->ro_block_group_mutex);
8773         if (trans->transaction->dirty_bg_run) {
8774                 u64 transid = trans->transid;
8775
8776                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
8777                 btrfs_end_transaction(trans, root);
8778
8779                 ret = btrfs_wait_for_commit(root, transid);
8780                 if (ret)
8781                         return ret;
8782                 goto again;
8783         }
8784
8785
8786         ret = set_block_group_ro(cache, 0);
8787         if (!ret)
8788                 goto out;
8789         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8790         ret = do_chunk_alloc(trans, root, alloc_flags,
8791                              CHUNK_ALLOC_FORCE);
8792         if (ret < 0)
8793                 goto out;
8794         ret = set_block_group_ro(cache, 0);
8795 out:
8796         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
8797                 alloc_flags = update_block_group_flags(root, cache->flags);
8798                 check_system_chunk(trans, root, alloc_flags);
8799         }
8800         mutex_unlock(&root->fs_info->ro_block_group_mutex);
8801
8802         btrfs_end_transaction(trans, root);
8803         return ret;
8804 }
8805
8806 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8807                             struct btrfs_root *root, u64 type)
8808 {
8809         u64 alloc_flags = get_alloc_profile(root, type);
8810         return do_chunk_alloc(trans, root, alloc_flags,
8811                               CHUNK_ALLOC_FORCE);
8812 }
8813
8814 /*
8815  * helper to account the unused space of all the readonly block group in the
8816  * space_info. takes mirrors into account.
8817  */
8818 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8819 {
8820         struct btrfs_block_group_cache *block_group;
8821         u64 free_bytes = 0;
8822         int factor;
8823
8824         /* It's df, we don't care if it's racey */
8825         if (list_empty(&sinfo->ro_bgs))
8826                 return 0;
8827
8828         spin_lock(&sinfo->lock);
8829         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
8830                 spin_lock(&block_group->lock);
8831
8832                 if (!block_group->ro) {
8833                         spin_unlock(&block_group->lock);
8834                         continue;
8835                 }
8836
8837                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8838                                           BTRFS_BLOCK_GROUP_RAID10 |
8839                                           BTRFS_BLOCK_GROUP_DUP))
8840                         factor = 2;
8841                 else
8842                         factor = 1;
8843
8844                 free_bytes += (block_group->key.offset -
8845                                btrfs_block_group_used(&block_group->item)) *
8846                                factor;
8847
8848                 spin_unlock(&block_group->lock);
8849         }
8850         spin_unlock(&sinfo->lock);
8851
8852         return free_bytes;
8853 }
8854
8855 void btrfs_set_block_group_rw(struct btrfs_root *root,
8856                               struct btrfs_block_group_cache *cache)
8857 {
8858         struct btrfs_space_info *sinfo = cache->space_info;
8859         u64 num_bytes;
8860
8861         BUG_ON(!cache->ro);
8862
8863         spin_lock(&sinfo->lock);
8864         spin_lock(&cache->lock);
8865         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8866                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8867         sinfo->bytes_readonly -= num_bytes;
8868         cache->ro = 0;
8869         list_del_init(&cache->ro_list);
8870         spin_unlock(&cache->lock);
8871         spin_unlock(&sinfo->lock);
8872 }
8873
8874 /*
8875  * checks to see if its even possible to relocate this block group.
8876  *
8877  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8878  * ok to go ahead and try.
8879  */
8880 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8881 {
8882         struct btrfs_block_group_cache *block_group;
8883         struct btrfs_space_info *space_info;
8884         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8885         struct btrfs_device *device;
8886         struct btrfs_trans_handle *trans;
8887         u64 min_free;
8888         u64 dev_min = 1;
8889         u64 dev_nr = 0;
8890         u64 target;
8891         int index;
8892         int full = 0;
8893         int ret = 0;
8894
8895         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8896
8897         /* odd, couldn't find the block group, leave it alone */
8898         if (!block_group)
8899                 return -1;
8900
8901         min_free = btrfs_block_group_used(&block_group->item);
8902
8903         /* no bytes used, we're good */
8904         if (!min_free)
8905                 goto out;
8906
8907         space_info = block_group->space_info;
8908         spin_lock(&space_info->lock);
8909
8910         full = space_info->full;
8911
8912         /*
8913          * if this is the last block group we have in this space, we can't
8914          * relocate it unless we're able to allocate a new chunk below.
8915          *
8916          * Otherwise, we need to make sure we have room in the space to handle
8917          * all of the extents from this block group.  If we can, we're good
8918          */
8919         if ((space_info->total_bytes != block_group->key.offset) &&
8920             (space_info->bytes_used + space_info->bytes_reserved +
8921              space_info->bytes_pinned + space_info->bytes_readonly +
8922              min_free < space_info->total_bytes)) {
8923                 spin_unlock(&space_info->lock);
8924                 goto out;
8925         }
8926         spin_unlock(&space_info->lock);
8927
8928         /*
8929          * ok we don't have enough space, but maybe we have free space on our
8930          * devices to allocate new chunks for relocation, so loop through our
8931          * alloc devices and guess if we have enough space.  if this block
8932          * group is going to be restriped, run checks against the target
8933          * profile instead of the current one.
8934          */
8935         ret = -1;
8936
8937         /*
8938          * index:
8939          *      0: raid10
8940          *      1: raid1
8941          *      2: dup
8942          *      3: raid0
8943          *      4: single
8944          */
8945         target = get_restripe_target(root->fs_info, block_group->flags);
8946         if (target) {
8947                 index = __get_raid_index(extended_to_chunk(target));
8948         } else {
8949                 /*
8950                  * this is just a balance, so if we were marked as full
8951                  * we know there is no space for a new chunk
8952                  */
8953                 if (full)
8954                         goto out;
8955
8956                 index = get_block_group_index(block_group);
8957         }
8958
8959         if (index == BTRFS_RAID_RAID10) {
8960                 dev_min = 4;
8961                 /* Divide by 2 */
8962                 min_free >>= 1;
8963         } else if (index == BTRFS_RAID_RAID1) {
8964                 dev_min = 2;
8965         } else if (index == BTRFS_RAID_DUP) {
8966                 /* Multiply by 2 */
8967                 min_free <<= 1;
8968         } else if (index == BTRFS_RAID_RAID0) {
8969                 dev_min = fs_devices->rw_devices;
8970                 min_free = div64_u64(min_free, dev_min);
8971         }
8972
8973         /* We need to do this so that we can look at pending chunks */
8974         trans = btrfs_join_transaction(root);
8975         if (IS_ERR(trans)) {
8976                 ret = PTR_ERR(trans);
8977                 goto out;
8978         }
8979
8980         mutex_lock(&root->fs_info->chunk_mutex);
8981         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8982                 u64 dev_offset;
8983
8984                 /*
8985                  * check to make sure we can actually find a chunk with enough
8986                  * space to fit our block group in.
8987                  */
8988                 if (device->total_bytes > device->bytes_used + min_free &&
8989                     !device->is_tgtdev_for_dev_replace) {
8990                         ret = find_free_dev_extent(trans, device, min_free,
8991                                                    &dev_offset, NULL);
8992                         if (!ret)
8993                                 dev_nr++;
8994
8995                         if (dev_nr >= dev_min)
8996                                 break;
8997
8998                         ret = -1;
8999                 }
9000         }
9001         mutex_unlock(&root->fs_info->chunk_mutex);
9002         btrfs_end_transaction(trans, root);
9003 out:
9004         btrfs_put_block_group(block_group);
9005         return ret;
9006 }
9007
9008 static int find_first_block_group(struct btrfs_root *root,
9009                 struct btrfs_path *path, struct btrfs_key *key)
9010 {
9011         int ret = 0;
9012         struct btrfs_key found_key;
9013         struct extent_buffer *leaf;
9014         int slot;
9015
9016         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9017         if (ret < 0)
9018                 goto out;
9019
9020         while (1) {
9021                 slot = path->slots[0];
9022                 leaf = path->nodes[0];
9023                 if (slot >= btrfs_header_nritems(leaf)) {
9024                         ret = btrfs_next_leaf(root, path);
9025                         if (ret == 0)
9026                                 continue;
9027                         if (ret < 0)
9028                                 goto out;
9029                         break;
9030                 }
9031                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9032
9033                 if (found_key.objectid >= key->objectid &&
9034                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9035                         ret = 0;
9036                         goto out;
9037                 }
9038                 path->slots[0]++;
9039         }
9040 out:
9041         return ret;
9042 }
9043
9044 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9045 {
9046         struct btrfs_block_group_cache *block_group;
9047         u64 last = 0;
9048
9049         while (1) {
9050                 struct inode *inode;
9051
9052                 block_group = btrfs_lookup_first_block_group(info, last);
9053                 while (block_group) {
9054                         spin_lock(&block_group->lock);
9055                         if (block_group->iref)
9056                                 break;
9057                         spin_unlock(&block_group->lock);
9058                         block_group = next_block_group(info->tree_root,
9059                                                        block_group);
9060                 }
9061                 if (!block_group) {
9062                         if (last == 0)
9063                                 break;
9064                         last = 0;
9065                         continue;
9066                 }
9067
9068                 inode = block_group->inode;
9069                 block_group->iref = 0;
9070                 block_group->inode = NULL;
9071                 spin_unlock(&block_group->lock);
9072                 iput(inode);
9073                 last = block_group->key.objectid + block_group->key.offset;
9074                 btrfs_put_block_group(block_group);
9075         }
9076 }
9077
9078 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9079 {
9080         struct btrfs_block_group_cache *block_group;
9081         struct btrfs_space_info *space_info;
9082         struct btrfs_caching_control *caching_ctl;
9083         struct rb_node *n;
9084
9085         down_write(&info->commit_root_sem);
9086         while (!list_empty(&info->caching_block_groups)) {
9087                 caching_ctl = list_entry(info->caching_block_groups.next,
9088                                          struct btrfs_caching_control, list);
9089                 list_del(&caching_ctl->list);
9090                 put_caching_control(caching_ctl);
9091         }
9092         up_write(&info->commit_root_sem);
9093
9094         spin_lock(&info->unused_bgs_lock);
9095         while (!list_empty(&info->unused_bgs)) {
9096                 block_group = list_first_entry(&info->unused_bgs,
9097                                                struct btrfs_block_group_cache,
9098                                                bg_list);
9099                 list_del_init(&block_group->bg_list);
9100                 btrfs_put_block_group(block_group);
9101         }
9102         spin_unlock(&info->unused_bgs_lock);
9103
9104         spin_lock(&info->block_group_cache_lock);
9105         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9106                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9107                                        cache_node);
9108                 rb_erase(&block_group->cache_node,
9109                          &info->block_group_cache_tree);
9110                 RB_CLEAR_NODE(&block_group->cache_node);
9111                 spin_unlock(&info->block_group_cache_lock);
9112
9113                 down_write(&block_group->space_info->groups_sem);
9114                 list_del(&block_group->list);
9115                 up_write(&block_group->space_info->groups_sem);
9116
9117                 if (block_group->cached == BTRFS_CACHE_STARTED)
9118                         wait_block_group_cache_done(block_group);
9119
9120                 /*
9121                  * We haven't cached this block group, which means we could
9122                  * possibly have excluded extents on this block group.
9123                  */
9124                 if (block_group->cached == BTRFS_CACHE_NO ||
9125                     block_group->cached == BTRFS_CACHE_ERROR)
9126                         free_excluded_extents(info->extent_root, block_group);
9127
9128                 btrfs_remove_free_space_cache(block_group);
9129                 btrfs_put_block_group(block_group);
9130
9131                 spin_lock(&info->block_group_cache_lock);
9132         }
9133         spin_unlock(&info->block_group_cache_lock);
9134
9135         /* now that all the block groups are freed, go through and
9136          * free all the space_info structs.  This is only called during
9137          * the final stages of unmount, and so we know nobody is
9138          * using them.  We call synchronize_rcu() once before we start,
9139          * just to be on the safe side.
9140          */
9141         synchronize_rcu();
9142
9143         release_global_block_rsv(info);
9144
9145         while (!list_empty(&info->space_info)) {
9146                 int i;
9147
9148                 space_info = list_entry(info->space_info.next,
9149                                         struct btrfs_space_info,
9150                                         list);
9151                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9152                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9153                             space_info->bytes_reserved > 0 ||
9154                             space_info->bytes_may_use > 0)) {
9155                                 dump_space_info(space_info, 0, 0);
9156                         }
9157                 }
9158                 list_del(&space_info->list);
9159                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9160                         struct kobject *kobj;
9161                         kobj = space_info->block_group_kobjs[i];
9162                         space_info->block_group_kobjs[i] = NULL;
9163                         if (kobj) {
9164                                 kobject_del(kobj);
9165                                 kobject_put(kobj);
9166                         }
9167                 }
9168                 kobject_del(&space_info->kobj);
9169                 kobject_put(&space_info->kobj);
9170         }
9171         return 0;
9172 }
9173
9174 static void __link_block_group(struct btrfs_space_info *space_info,
9175                                struct btrfs_block_group_cache *cache)
9176 {
9177         int index = get_block_group_index(cache);
9178         bool first = false;
9179
9180         down_write(&space_info->groups_sem);
9181         if (list_empty(&space_info->block_groups[index]))
9182                 first = true;
9183         list_add_tail(&cache->list, &space_info->block_groups[index]);
9184         up_write(&space_info->groups_sem);
9185
9186         if (first) {
9187                 struct raid_kobject *rkobj;
9188                 int ret;
9189
9190                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9191                 if (!rkobj)
9192                         goto out_err;
9193                 rkobj->raid_type = index;
9194                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9195                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9196                                   "%s", get_raid_name(index));
9197                 if (ret) {
9198                         kobject_put(&rkobj->kobj);
9199                         goto out_err;
9200                 }
9201                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9202         }
9203
9204         return;
9205 out_err:
9206         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9207 }
9208
9209 static struct btrfs_block_group_cache *
9210 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9211 {
9212         struct btrfs_block_group_cache *cache;
9213
9214         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9215         if (!cache)
9216                 return NULL;
9217
9218         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9219                                         GFP_NOFS);
9220         if (!cache->free_space_ctl) {
9221                 kfree(cache);
9222                 return NULL;
9223         }
9224
9225         cache->key.objectid = start;
9226         cache->key.offset = size;
9227         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9228
9229         cache->sectorsize = root->sectorsize;
9230         cache->fs_info = root->fs_info;
9231         cache->full_stripe_len = btrfs_full_stripe_len(root,
9232                                                &root->fs_info->mapping_tree,
9233                                                start);
9234         atomic_set(&cache->count, 1);
9235         spin_lock_init(&cache->lock);
9236         init_rwsem(&cache->data_rwsem);
9237         INIT_LIST_HEAD(&cache->list);
9238         INIT_LIST_HEAD(&cache->cluster_list);
9239         INIT_LIST_HEAD(&cache->bg_list);
9240         INIT_LIST_HEAD(&cache->ro_list);
9241         INIT_LIST_HEAD(&cache->dirty_list);
9242         INIT_LIST_HEAD(&cache->io_list);
9243         btrfs_init_free_space_ctl(cache);
9244         atomic_set(&cache->trimming, 0);
9245
9246         return cache;
9247 }
9248
9249 int btrfs_read_block_groups(struct btrfs_root *root)
9250 {
9251         struct btrfs_path *path;
9252         int ret;
9253         struct btrfs_block_group_cache *cache;
9254         struct btrfs_fs_info *info = root->fs_info;
9255         struct btrfs_space_info *space_info;
9256         struct btrfs_key key;
9257         struct btrfs_key found_key;
9258         struct extent_buffer *leaf;
9259         int need_clear = 0;
9260         u64 cache_gen;
9261
9262         root = info->extent_root;
9263         key.objectid = 0;
9264         key.offset = 0;
9265         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9266         path = btrfs_alloc_path();
9267         if (!path)
9268                 return -ENOMEM;
9269         path->reada = 1;
9270
9271         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9272         if (btrfs_test_opt(root, SPACE_CACHE) &&
9273             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9274                 need_clear = 1;
9275         if (btrfs_test_opt(root, CLEAR_CACHE))
9276                 need_clear = 1;
9277
9278         while (1) {
9279                 ret = find_first_block_group(root, path, &key);
9280                 if (ret > 0)
9281                         break;
9282                 if (ret != 0)
9283                         goto error;
9284
9285                 leaf = path->nodes[0];
9286                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9287
9288                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9289                                                        found_key.offset);
9290                 if (!cache) {
9291                         ret = -ENOMEM;
9292                         goto error;
9293                 }
9294
9295                 if (need_clear) {
9296                         /*
9297                          * When we mount with old space cache, we need to
9298                          * set BTRFS_DC_CLEAR and set dirty flag.
9299                          *
9300                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9301                          *    truncate the old free space cache inode and
9302                          *    setup a new one.
9303                          * b) Setting 'dirty flag' makes sure that we flush
9304                          *    the new space cache info onto disk.
9305                          */
9306                         if (btrfs_test_opt(root, SPACE_CACHE))
9307                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9308                 }
9309
9310                 read_extent_buffer(leaf, &cache->item,
9311                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9312                                    sizeof(cache->item));
9313                 cache->flags = btrfs_block_group_flags(&cache->item);
9314
9315                 key.objectid = found_key.objectid + found_key.offset;
9316                 btrfs_release_path(path);
9317
9318                 /*
9319                  * We need to exclude the super stripes now so that the space
9320                  * info has super bytes accounted for, otherwise we'll think
9321                  * we have more space than we actually do.
9322                  */
9323                 ret = exclude_super_stripes(root, cache);
9324                 if (ret) {
9325                         /*
9326                          * We may have excluded something, so call this just in
9327                          * case.
9328                          */
9329                         free_excluded_extents(root, cache);
9330                         btrfs_put_block_group(cache);
9331                         goto error;
9332                 }
9333
9334                 /*
9335                  * check for two cases, either we are full, and therefore
9336                  * don't need to bother with the caching work since we won't
9337                  * find any space, or we are empty, and we can just add all
9338                  * the space in and be done with it.  This saves us _alot_ of
9339                  * time, particularly in the full case.
9340                  */
9341                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9342                         cache->last_byte_to_unpin = (u64)-1;
9343                         cache->cached = BTRFS_CACHE_FINISHED;
9344                         free_excluded_extents(root, cache);
9345                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9346                         cache->last_byte_to_unpin = (u64)-1;
9347                         cache->cached = BTRFS_CACHE_FINISHED;
9348                         add_new_free_space(cache, root->fs_info,
9349                                            found_key.objectid,
9350                                            found_key.objectid +
9351                                            found_key.offset);
9352                         free_excluded_extents(root, cache);
9353                 }
9354
9355                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9356                 if (ret) {
9357                         btrfs_remove_free_space_cache(cache);
9358                         btrfs_put_block_group(cache);
9359                         goto error;
9360                 }
9361
9362                 ret = update_space_info(info, cache->flags, found_key.offset,
9363                                         btrfs_block_group_used(&cache->item),
9364                                         &space_info);
9365                 if (ret) {
9366                         btrfs_remove_free_space_cache(cache);
9367                         spin_lock(&info->block_group_cache_lock);
9368                         rb_erase(&cache->cache_node,
9369                                  &info->block_group_cache_tree);
9370                         RB_CLEAR_NODE(&cache->cache_node);
9371                         spin_unlock(&info->block_group_cache_lock);
9372                         btrfs_put_block_group(cache);
9373                         goto error;
9374                 }
9375
9376                 cache->space_info = space_info;
9377                 spin_lock(&cache->space_info->lock);
9378                 cache->space_info->bytes_readonly += cache->bytes_super;
9379                 spin_unlock(&cache->space_info->lock);
9380
9381                 __link_block_group(space_info, cache);
9382
9383                 set_avail_alloc_bits(root->fs_info, cache->flags);
9384                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9385                         set_block_group_ro(cache, 1);
9386                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9387                         spin_lock(&info->unused_bgs_lock);
9388                         /* Should always be true but just in case. */
9389                         if (list_empty(&cache->bg_list)) {
9390                                 btrfs_get_block_group(cache);
9391                                 list_add_tail(&cache->bg_list,
9392                                               &info->unused_bgs);
9393                         }
9394                         spin_unlock(&info->unused_bgs_lock);
9395                 }
9396         }
9397
9398         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9399                 if (!(get_alloc_profile(root, space_info->flags) &
9400                       (BTRFS_BLOCK_GROUP_RAID10 |
9401                        BTRFS_BLOCK_GROUP_RAID1 |
9402                        BTRFS_BLOCK_GROUP_RAID5 |
9403                        BTRFS_BLOCK_GROUP_RAID6 |
9404                        BTRFS_BLOCK_GROUP_DUP)))
9405                         continue;
9406                 /*
9407                  * avoid allocating from un-mirrored block group if there are
9408                  * mirrored block groups.
9409                  */
9410                 list_for_each_entry(cache,
9411                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9412                                 list)
9413                         set_block_group_ro(cache, 1);
9414                 list_for_each_entry(cache,
9415                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9416                                 list)
9417                         set_block_group_ro(cache, 1);
9418         }
9419
9420         init_global_block_rsv(info);
9421         ret = 0;
9422 error:
9423         btrfs_free_path(path);
9424         return ret;
9425 }
9426
9427 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9428                                        struct btrfs_root *root)
9429 {
9430         struct btrfs_block_group_cache *block_group, *tmp;
9431         struct btrfs_root *extent_root = root->fs_info->extent_root;
9432         struct btrfs_block_group_item item;
9433         struct btrfs_key key;
9434         int ret = 0;
9435
9436         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9437                 if (ret)
9438                         goto next;
9439
9440                 spin_lock(&block_group->lock);
9441                 memcpy(&item, &block_group->item, sizeof(item));
9442                 memcpy(&key, &block_group->key, sizeof(key));
9443                 spin_unlock(&block_group->lock);
9444
9445                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9446                                         sizeof(item));
9447                 if (ret)
9448                         btrfs_abort_transaction(trans, extent_root, ret);
9449                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9450                                                key.objectid, key.offset);
9451                 if (ret)
9452                         btrfs_abort_transaction(trans, extent_root, ret);
9453 next:
9454                 list_del_init(&block_group->bg_list);
9455         }
9456 }
9457
9458 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9459                            struct btrfs_root *root, u64 bytes_used,
9460                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9461                            u64 size)
9462 {
9463         int ret;
9464         struct btrfs_root *extent_root;
9465         struct btrfs_block_group_cache *cache;
9466
9467         extent_root = root->fs_info->extent_root;
9468
9469         btrfs_set_log_full_commit(root->fs_info, trans);
9470
9471         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9472         if (!cache)
9473                 return -ENOMEM;
9474
9475         btrfs_set_block_group_used(&cache->item, bytes_used);
9476         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9477         btrfs_set_block_group_flags(&cache->item, type);
9478
9479         cache->flags = type;
9480         cache->last_byte_to_unpin = (u64)-1;
9481         cache->cached = BTRFS_CACHE_FINISHED;
9482         ret = exclude_super_stripes(root, cache);
9483         if (ret) {
9484                 /*
9485                  * We may have excluded something, so call this just in
9486                  * case.
9487                  */
9488                 free_excluded_extents(root, cache);
9489                 btrfs_put_block_group(cache);
9490                 return ret;
9491         }
9492
9493         add_new_free_space(cache, root->fs_info, chunk_offset,
9494                            chunk_offset + size);
9495
9496         free_excluded_extents(root, cache);
9497
9498         ret = btrfs_add_block_group_cache(root->fs_info, cache);
9499         if (ret) {
9500                 btrfs_remove_free_space_cache(cache);
9501                 btrfs_put_block_group(cache);
9502                 return ret;
9503         }
9504
9505         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
9506                                 &cache->space_info);
9507         if (ret) {
9508                 btrfs_remove_free_space_cache(cache);
9509                 spin_lock(&root->fs_info->block_group_cache_lock);
9510                 rb_erase(&cache->cache_node,
9511                          &root->fs_info->block_group_cache_tree);
9512                 RB_CLEAR_NODE(&cache->cache_node);
9513                 spin_unlock(&root->fs_info->block_group_cache_lock);
9514                 btrfs_put_block_group(cache);
9515                 return ret;
9516         }
9517         update_global_block_rsv(root->fs_info);
9518
9519         spin_lock(&cache->space_info->lock);
9520         cache->space_info->bytes_readonly += cache->bytes_super;
9521         spin_unlock(&cache->space_info->lock);
9522
9523         __link_block_group(cache->space_info, cache);
9524
9525         list_add_tail(&cache->bg_list, &trans->new_bgs);
9526
9527         set_avail_alloc_bits(extent_root->fs_info, type);
9528
9529         return 0;
9530 }
9531
9532 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9533 {
9534         u64 extra_flags = chunk_to_extended(flags) &
9535                                 BTRFS_EXTENDED_PROFILE_MASK;
9536
9537         write_seqlock(&fs_info->profiles_lock);
9538         if (flags & BTRFS_BLOCK_GROUP_DATA)
9539                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9540         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9541                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9542         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9543                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9544         write_sequnlock(&fs_info->profiles_lock);
9545 }
9546
9547 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9548                              struct btrfs_root *root, u64 group_start,
9549                              struct extent_map *em)
9550 {
9551         struct btrfs_path *path;
9552         struct btrfs_block_group_cache *block_group;
9553         struct btrfs_free_cluster *cluster;
9554         struct btrfs_root *tree_root = root->fs_info->tree_root;
9555         struct btrfs_key key;
9556         struct inode *inode;
9557         struct kobject *kobj = NULL;
9558         int ret;
9559         int index;
9560         int factor;
9561         struct btrfs_caching_control *caching_ctl = NULL;
9562         bool remove_em;
9563
9564         root = root->fs_info->extent_root;
9565
9566         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9567         BUG_ON(!block_group);
9568         BUG_ON(!block_group->ro);
9569
9570         /*
9571          * Free the reserved super bytes from this block group before
9572          * remove it.
9573          */
9574         free_excluded_extents(root, block_group);
9575
9576         memcpy(&key, &block_group->key, sizeof(key));
9577         index = get_block_group_index(block_group);
9578         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9579                                   BTRFS_BLOCK_GROUP_RAID1 |
9580                                   BTRFS_BLOCK_GROUP_RAID10))
9581                 factor = 2;
9582         else
9583                 factor = 1;
9584
9585         /* make sure this block group isn't part of an allocation cluster */
9586         cluster = &root->fs_info->data_alloc_cluster;
9587         spin_lock(&cluster->refill_lock);
9588         btrfs_return_cluster_to_free_space(block_group, cluster);
9589         spin_unlock(&cluster->refill_lock);
9590
9591         /*
9592          * make sure this block group isn't part of a metadata
9593          * allocation cluster
9594          */
9595         cluster = &root->fs_info->meta_alloc_cluster;
9596         spin_lock(&cluster->refill_lock);
9597         btrfs_return_cluster_to_free_space(block_group, cluster);
9598         spin_unlock(&cluster->refill_lock);
9599
9600         path = btrfs_alloc_path();
9601         if (!path) {
9602                 ret = -ENOMEM;
9603                 goto out;
9604         }
9605
9606         /*
9607          * get the inode first so any iput calls done for the io_list
9608          * aren't the final iput (no unlinks allowed now)
9609          */
9610         inode = lookup_free_space_inode(tree_root, block_group, path);
9611
9612         mutex_lock(&trans->transaction->cache_write_mutex);
9613         /*
9614          * make sure our free spache cache IO is done before remove the
9615          * free space inode
9616          */
9617         spin_lock(&trans->transaction->dirty_bgs_lock);
9618         if (!list_empty(&block_group->io_list)) {
9619                 list_del_init(&block_group->io_list);
9620
9621                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
9622
9623                 spin_unlock(&trans->transaction->dirty_bgs_lock);
9624                 btrfs_wait_cache_io(root, trans, block_group,
9625                                     &block_group->io_ctl, path,
9626                                     block_group->key.objectid);
9627                 btrfs_put_block_group(block_group);
9628                 spin_lock(&trans->transaction->dirty_bgs_lock);
9629         }
9630
9631         if (!list_empty(&block_group->dirty_list)) {
9632                 list_del_init(&block_group->dirty_list);
9633                 btrfs_put_block_group(block_group);
9634         }
9635         spin_unlock(&trans->transaction->dirty_bgs_lock);
9636         mutex_unlock(&trans->transaction->cache_write_mutex);
9637
9638         if (!IS_ERR(inode)) {
9639                 ret = btrfs_orphan_add(trans, inode);
9640                 if (ret) {
9641                         btrfs_add_delayed_iput(inode);
9642                         goto out;
9643                 }
9644                 clear_nlink(inode);
9645                 /* One for the block groups ref */
9646                 spin_lock(&block_group->lock);
9647                 if (block_group->iref) {
9648                         block_group->iref = 0;
9649                         block_group->inode = NULL;
9650                         spin_unlock(&block_group->lock);
9651                         iput(inode);
9652                 } else {
9653                         spin_unlock(&block_group->lock);
9654                 }
9655                 /* One for our lookup ref */
9656                 btrfs_add_delayed_iput(inode);
9657         }
9658
9659         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9660         key.offset = block_group->key.objectid;
9661         key.type = 0;
9662
9663         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9664         if (ret < 0)
9665                 goto out;
9666         if (ret > 0)
9667                 btrfs_release_path(path);
9668         if (ret == 0) {
9669                 ret = btrfs_del_item(trans, tree_root, path);
9670                 if (ret)
9671                         goto out;
9672                 btrfs_release_path(path);
9673         }
9674
9675         spin_lock(&root->fs_info->block_group_cache_lock);
9676         rb_erase(&block_group->cache_node,
9677                  &root->fs_info->block_group_cache_tree);
9678         RB_CLEAR_NODE(&block_group->cache_node);
9679
9680         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9681                 root->fs_info->first_logical_byte = (u64)-1;
9682         spin_unlock(&root->fs_info->block_group_cache_lock);
9683
9684         down_write(&block_group->space_info->groups_sem);
9685         /*
9686          * we must use list_del_init so people can check to see if they
9687          * are still on the list after taking the semaphore
9688          */
9689         list_del_init(&block_group->list);
9690         if (list_empty(&block_group->space_info->block_groups[index])) {
9691                 kobj = block_group->space_info->block_group_kobjs[index];
9692                 block_group->space_info->block_group_kobjs[index] = NULL;
9693                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9694         }
9695         up_write(&block_group->space_info->groups_sem);
9696         if (kobj) {
9697                 kobject_del(kobj);
9698                 kobject_put(kobj);
9699         }
9700
9701         if (block_group->has_caching_ctl)
9702                 caching_ctl = get_caching_control(block_group);
9703         if (block_group->cached == BTRFS_CACHE_STARTED)
9704                 wait_block_group_cache_done(block_group);
9705         if (block_group->has_caching_ctl) {
9706                 down_write(&root->fs_info->commit_root_sem);
9707                 if (!caching_ctl) {
9708                         struct btrfs_caching_control *ctl;
9709
9710                         list_for_each_entry(ctl,
9711                                     &root->fs_info->caching_block_groups, list)
9712                                 if (ctl->block_group == block_group) {
9713                                         caching_ctl = ctl;
9714                                         atomic_inc(&caching_ctl->count);
9715                                         break;
9716                                 }
9717                 }
9718                 if (caching_ctl)
9719                         list_del_init(&caching_ctl->list);
9720                 up_write(&root->fs_info->commit_root_sem);
9721                 if (caching_ctl) {
9722                         /* Once for the caching bgs list and once for us. */
9723                         put_caching_control(caching_ctl);
9724                         put_caching_control(caching_ctl);
9725                 }
9726         }
9727
9728         spin_lock(&trans->transaction->dirty_bgs_lock);
9729         if (!list_empty(&block_group->dirty_list)) {
9730                 WARN_ON(1);
9731         }
9732         if (!list_empty(&block_group->io_list)) {
9733                 WARN_ON(1);
9734         }
9735         spin_unlock(&trans->transaction->dirty_bgs_lock);
9736         btrfs_remove_free_space_cache(block_group);
9737
9738         spin_lock(&block_group->space_info->lock);
9739         list_del_init(&block_group->ro_list);
9740
9741         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
9742                 WARN_ON(block_group->space_info->total_bytes
9743                         < block_group->key.offset);
9744                 WARN_ON(block_group->space_info->bytes_readonly
9745                         < block_group->key.offset);
9746                 WARN_ON(block_group->space_info->disk_total
9747                         < block_group->key.offset * factor);
9748         }
9749         block_group->space_info->total_bytes -= block_group->key.offset;
9750         block_group->space_info->bytes_readonly -= block_group->key.offset;
9751         block_group->space_info->disk_total -= block_group->key.offset * factor;
9752
9753         spin_unlock(&block_group->space_info->lock);
9754
9755         memcpy(&key, &block_group->key, sizeof(key));
9756
9757         lock_chunks(root);
9758         if (!list_empty(&em->list)) {
9759                 /* We're in the transaction->pending_chunks list. */
9760                 free_extent_map(em);
9761         }
9762         spin_lock(&block_group->lock);
9763         block_group->removed = 1;
9764         /*
9765          * At this point trimming can't start on this block group, because we
9766          * removed the block group from the tree fs_info->block_group_cache_tree
9767          * so no one can't find it anymore and even if someone already got this
9768          * block group before we removed it from the rbtree, they have already
9769          * incremented block_group->trimming - if they didn't, they won't find
9770          * any free space entries because we already removed them all when we
9771          * called btrfs_remove_free_space_cache().
9772          *
9773          * And we must not remove the extent map from the fs_info->mapping_tree
9774          * to prevent the same logical address range and physical device space
9775          * ranges from being reused for a new block group. This is because our
9776          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
9777          * completely transactionless, so while it is trimming a range the
9778          * currently running transaction might finish and a new one start,
9779          * allowing for new block groups to be created that can reuse the same
9780          * physical device locations unless we take this special care.
9781          */
9782         remove_em = (atomic_read(&block_group->trimming) == 0);
9783         /*
9784          * Make sure a trimmer task always sees the em in the pinned_chunks list
9785          * if it sees block_group->removed == 1 (needs to lock block_group->lock
9786          * before checking block_group->removed).
9787          */
9788         if (!remove_em) {
9789                 /*
9790                  * Our em might be in trans->transaction->pending_chunks which
9791                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
9792                  * and so is the fs_info->pinned_chunks list.
9793                  *
9794                  * So at this point we must be holding the chunk_mutex to avoid
9795                  * any races with chunk allocation (more specifically at
9796                  * volumes.c:contains_pending_extent()), to ensure it always
9797                  * sees the em, either in the pending_chunks list or in the
9798                  * pinned_chunks list.
9799                  */
9800                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
9801         }
9802         spin_unlock(&block_group->lock);
9803
9804         if (remove_em) {
9805                 struct extent_map_tree *em_tree;
9806
9807                 em_tree = &root->fs_info->mapping_tree.map_tree;
9808                 write_lock(&em_tree->lock);
9809                 /*
9810                  * The em might be in the pending_chunks list, so make sure the
9811                  * chunk mutex is locked, since remove_extent_mapping() will
9812                  * delete us from that list.
9813                  */
9814                 remove_extent_mapping(em_tree, em);
9815                 write_unlock(&em_tree->lock);
9816                 /* once for the tree */
9817                 free_extent_map(em);
9818         }
9819
9820         unlock_chunks(root);
9821
9822         btrfs_put_block_group(block_group);
9823         btrfs_put_block_group(block_group);
9824
9825         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9826         if (ret > 0)
9827                 ret = -EIO;
9828         if (ret < 0)
9829                 goto out;
9830
9831         ret = btrfs_del_item(trans, root, path);
9832 out:
9833         btrfs_free_path(path);
9834         return ret;
9835 }
9836
9837 /*
9838  * Process the unused_bgs list and remove any that don't have any allocated
9839  * space inside of them.
9840  */
9841 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
9842 {
9843         struct btrfs_block_group_cache *block_group;
9844         struct btrfs_space_info *space_info;
9845         struct btrfs_root *root = fs_info->extent_root;
9846         struct btrfs_trans_handle *trans;
9847         int ret = 0;
9848
9849         if (!fs_info->open)
9850                 return;
9851
9852         spin_lock(&fs_info->unused_bgs_lock);
9853         while (!list_empty(&fs_info->unused_bgs)) {
9854                 u64 start, end;
9855
9856                 block_group = list_first_entry(&fs_info->unused_bgs,
9857                                                struct btrfs_block_group_cache,
9858                                                bg_list);
9859                 space_info = block_group->space_info;
9860                 list_del_init(&block_group->bg_list);
9861                 if (ret || btrfs_mixed_space_info(space_info)) {
9862                         btrfs_put_block_group(block_group);
9863                         continue;
9864                 }
9865                 spin_unlock(&fs_info->unused_bgs_lock);
9866
9867                 /* Don't want to race with allocators so take the groups_sem */
9868                 down_write(&space_info->groups_sem);
9869                 spin_lock(&block_group->lock);
9870                 if (block_group->reserved ||
9871                     btrfs_block_group_used(&block_group->item) ||
9872                     block_group->ro) {
9873                         /*
9874                          * We want to bail if we made new allocations or have
9875                          * outstanding allocations in this block group.  We do
9876                          * the ro check in case balance is currently acting on
9877                          * this block group.
9878                          */
9879                         spin_unlock(&block_group->lock);
9880                         up_write(&space_info->groups_sem);
9881                         goto next;
9882                 }
9883                 spin_unlock(&block_group->lock);
9884
9885                 /* We don't want to force the issue, only flip if it's ok. */
9886                 ret = set_block_group_ro(block_group, 0);
9887                 up_write(&space_info->groups_sem);
9888                 if (ret < 0) {
9889                         ret = 0;
9890                         goto next;
9891                 }
9892
9893                 /*
9894                  * Want to do this before we do anything else so we can recover
9895                  * properly if we fail to join the transaction.
9896                  */
9897                 /* 1 for btrfs_orphan_reserve_metadata() */
9898                 trans = btrfs_start_transaction(root, 1);
9899                 if (IS_ERR(trans)) {
9900                         btrfs_set_block_group_rw(root, block_group);
9901                         ret = PTR_ERR(trans);
9902                         goto next;
9903                 }
9904
9905                 /*
9906                  * We could have pending pinned extents for this block group,
9907                  * just delete them, we don't care about them anymore.
9908                  */
9909                 start = block_group->key.objectid;
9910                 end = start + block_group->key.offset - 1;
9911                 /*
9912                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
9913                  * btrfs_finish_extent_commit(). If we are at transaction N,
9914                  * another task might be running finish_extent_commit() for the
9915                  * previous transaction N - 1, and have seen a range belonging
9916                  * to the block group in freed_extents[] before we were able to
9917                  * clear the whole block group range from freed_extents[]. This
9918                  * means that task can lookup for the block group after we
9919                  * unpinned it from freed_extents[] and removed it, leading to
9920                  * a BUG_ON() at btrfs_unpin_extent_range().
9921                  */
9922                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
9923                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
9924                                   EXTENT_DIRTY, GFP_NOFS);
9925                 if (ret) {
9926                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9927                         btrfs_set_block_group_rw(root, block_group);
9928                         goto end_trans;
9929                 }
9930                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
9931                                   EXTENT_DIRTY, GFP_NOFS);
9932                 if (ret) {
9933                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9934                         btrfs_set_block_group_rw(root, block_group);
9935                         goto end_trans;
9936                 }
9937                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9938
9939                 /* Reset pinned so btrfs_put_block_group doesn't complain */
9940                 spin_lock(&space_info->lock);
9941                 spin_lock(&block_group->lock);
9942
9943                 space_info->bytes_pinned -= block_group->pinned;
9944                 space_info->bytes_readonly += block_group->pinned;
9945                 percpu_counter_add(&space_info->total_bytes_pinned,
9946                                    -block_group->pinned);
9947                 block_group->pinned = 0;
9948
9949                 spin_unlock(&block_group->lock);
9950                 spin_unlock(&space_info->lock);
9951
9952                 /*
9953                  * Btrfs_remove_chunk will abort the transaction if things go
9954                  * horribly wrong.
9955                  */
9956                 ret = btrfs_remove_chunk(trans, root,
9957                                          block_group->key.objectid);
9958 end_trans:
9959                 btrfs_end_transaction(trans, root);
9960 next:
9961                 btrfs_put_block_group(block_group);
9962                 spin_lock(&fs_info->unused_bgs_lock);
9963         }
9964         spin_unlock(&fs_info->unused_bgs_lock);
9965 }
9966
9967 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
9968 {
9969         struct btrfs_space_info *space_info;
9970         struct btrfs_super_block *disk_super;
9971         u64 features;
9972         u64 flags;
9973         int mixed = 0;
9974         int ret;
9975
9976         disk_super = fs_info->super_copy;
9977         if (!btrfs_super_root(disk_super))
9978                 return 1;
9979
9980         features = btrfs_super_incompat_flags(disk_super);
9981         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
9982                 mixed = 1;
9983
9984         flags = BTRFS_BLOCK_GROUP_SYSTEM;
9985         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9986         if (ret)
9987                 goto out;
9988
9989         if (mixed) {
9990                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
9991                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9992         } else {
9993                 flags = BTRFS_BLOCK_GROUP_METADATA;
9994                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9995                 if (ret)
9996                         goto out;
9997
9998                 flags = BTRFS_BLOCK_GROUP_DATA;
9999                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10000         }
10001 out:
10002         return ret;
10003 }
10004
10005 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10006 {
10007         return unpin_extent_range(root, start, end, false);
10008 }
10009
10010 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10011 {
10012         struct btrfs_fs_info *fs_info = root->fs_info;
10013         struct btrfs_block_group_cache *cache = NULL;
10014         u64 group_trimmed;
10015         u64 start;
10016         u64 end;
10017         u64 trimmed = 0;
10018         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10019         int ret = 0;
10020
10021         /*
10022          * try to trim all FS space, our block group may start from non-zero.
10023          */
10024         if (range->len == total_bytes)
10025                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10026         else
10027                 cache = btrfs_lookup_block_group(fs_info, range->start);
10028
10029         while (cache) {
10030                 if (cache->key.objectid >= (range->start + range->len)) {
10031                         btrfs_put_block_group(cache);
10032                         break;
10033                 }
10034
10035                 start = max(range->start, cache->key.objectid);
10036                 end = min(range->start + range->len,
10037                                 cache->key.objectid + cache->key.offset);
10038
10039                 if (end - start >= range->minlen) {
10040                         if (!block_group_cache_done(cache)) {
10041                                 ret = cache_block_group(cache, 0);
10042                                 if (ret) {
10043                                         btrfs_put_block_group(cache);
10044                                         break;
10045                                 }
10046                                 ret = wait_block_group_cache_done(cache);
10047                                 if (ret) {
10048                                         btrfs_put_block_group(cache);
10049                                         break;
10050                                 }
10051                         }
10052                         ret = btrfs_trim_block_group(cache,
10053                                                      &group_trimmed,
10054                                                      start,
10055                                                      end,
10056                                                      range->minlen);
10057
10058                         trimmed += group_trimmed;
10059                         if (ret) {
10060                                 btrfs_put_block_group(cache);
10061                                 break;
10062                         }
10063                 }
10064
10065                 cache = next_block_group(fs_info->tree_root, cache);
10066         }
10067
10068         range->len = trimmed;
10069         return ret;
10070 }
10071
10072 /*
10073  * btrfs_{start,end}_write_no_snapshoting() are similar to
10074  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10075  * data into the page cache through nocow before the subvolume is snapshoted,
10076  * but flush the data into disk after the snapshot creation, or to prevent
10077  * operations while snapshoting is ongoing and that cause the snapshot to be
10078  * inconsistent (writes followed by expanding truncates for example).
10079  */
10080 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10081 {
10082         percpu_counter_dec(&root->subv_writers->counter);
10083         /*
10084          * Make sure counter is updated before we wake up
10085          * waiters.
10086          */
10087         smp_mb();
10088         if (waitqueue_active(&root->subv_writers->wait))
10089                 wake_up(&root->subv_writers->wait);
10090 }
10091
10092 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10093 {
10094         if (atomic_read(&root->will_be_snapshoted))
10095                 return 0;
10096
10097         percpu_counter_inc(&root->subv_writers->counter);
10098         /*
10099          * Make sure counter is updated before we check for snapshot creation.
10100          */
10101         smp_mb();
10102         if (atomic_read(&root->will_be_snapshoted)) {
10103                 btrfs_end_write_no_snapshoting(root);
10104                 return 0;
10105         }
10106         return 1;
10107 }