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                         btrfs_delayed_ref_unlock(locked_ref);
2542                         locked_ref = NULL;
2543                 }
2544                 btrfs_put_delayed_ref(ref);
2545                 count++;
2546                 cond_resched();
2547         }
2548
2549         /*
2550          * We don't want to include ref heads since we can have empty ref heads
2551          * and those will drastically skew our runtime down since we just do
2552          * accounting, no actual extent tree updates.
2553          */
2554         if (actual_count > 0) {
2555                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2556                 u64 avg;
2557
2558                 /*
2559                  * We weigh the current average higher than our current runtime
2560                  * to avoid large swings in the average.
2561                  */
2562                 spin_lock(&delayed_refs->lock);
2563                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2564                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2565                 spin_unlock(&delayed_refs->lock);
2566         }
2567         return 0;
2568 }
2569
2570 #ifdef SCRAMBLE_DELAYED_REFS
2571 /*
2572  * Normally delayed refs get processed in ascending bytenr order. This
2573  * correlates in most cases to the order added. To expose dependencies on this
2574  * order, we start to process the tree in the middle instead of the beginning
2575  */
2576 static u64 find_middle(struct rb_root *root)
2577 {
2578         struct rb_node *n = root->rb_node;
2579         struct btrfs_delayed_ref_node *entry;
2580         int alt = 1;
2581         u64 middle;
2582         u64 first = 0, last = 0;
2583
2584         n = rb_first(root);
2585         if (n) {
2586                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2587                 first = entry->bytenr;
2588         }
2589         n = rb_last(root);
2590         if (n) {
2591                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2592                 last = entry->bytenr;
2593         }
2594         n = root->rb_node;
2595
2596         while (n) {
2597                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2598                 WARN_ON(!entry->in_tree);
2599
2600                 middle = entry->bytenr;
2601
2602                 if (alt)
2603                         n = n->rb_left;
2604                 else
2605                         n = n->rb_right;
2606
2607                 alt = 1 - alt;
2608         }
2609         return middle;
2610 }
2611 #endif
2612
2613 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2614 {
2615         u64 num_bytes;
2616
2617         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2618                              sizeof(struct btrfs_extent_inline_ref));
2619         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2620                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2621
2622         /*
2623          * We don't ever fill up leaves all the way so multiply by 2 just to be
2624          * closer to what we're really going to want to ouse.
2625          */
2626         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2627 }
2628
2629 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2630                                        struct btrfs_root *root)
2631 {
2632         struct btrfs_block_rsv *global_rsv;
2633         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2634         u64 num_bytes;
2635         int ret = 0;
2636
2637         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2638         num_heads = heads_to_leaves(root, num_heads);
2639         if (num_heads > 1)
2640                 num_bytes += (num_heads - 1) * root->nodesize;
2641         num_bytes <<= 1;
2642         global_rsv = &root->fs_info->global_block_rsv;
2643
2644         /*
2645          * If we can't allocate any more chunks lets make sure we have _lots_ of
2646          * wiggle room since running delayed refs can create more delayed refs.
2647          */
2648         if (global_rsv->space_info->full)
2649                 num_bytes <<= 1;
2650
2651         spin_lock(&global_rsv->lock);
2652         if (global_rsv->reserved <= num_bytes)
2653                 ret = 1;
2654         spin_unlock(&global_rsv->lock);
2655         return ret;
2656 }
2657
2658 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2659                                        struct btrfs_root *root)
2660 {
2661         struct btrfs_fs_info *fs_info = root->fs_info;
2662         u64 num_entries =
2663                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2664         u64 avg_runtime;
2665         u64 val;
2666
2667         smp_mb();
2668         avg_runtime = fs_info->avg_delayed_ref_runtime;
2669         val = num_entries * avg_runtime;
2670         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2671                 return 1;
2672         if (val >= NSEC_PER_SEC / 2)
2673                 return 2;
2674
2675         return btrfs_check_space_for_delayed_refs(trans, root);
2676 }
2677
2678 struct async_delayed_refs {
2679         struct btrfs_root *root;
2680         int count;
2681         int error;
2682         int sync;
2683         struct completion wait;
2684         struct btrfs_work work;
2685 };
2686
2687 static void delayed_ref_async_start(struct btrfs_work *work)
2688 {
2689         struct async_delayed_refs *async;
2690         struct btrfs_trans_handle *trans;
2691         int ret;
2692
2693         async = container_of(work, struct async_delayed_refs, work);
2694
2695         trans = btrfs_join_transaction(async->root);
2696         if (IS_ERR(trans)) {
2697                 async->error = PTR_ERR(trans);
2698                 goto done;
2699         }
2700
2701         /*
2702          * trans->sync means that when we call end_transaciton, we won't
2703          * wait on delayed refs
2704          */
2705         trans->sync = true;
2706         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2707         if (ret)
2708                 async->error = ret;
2709
2710         ret = btrfs_end_transaction(trans, async->root);
2711         if (ret && !async->error)
2712                 async->error = ret;
2713 done:
2714         if (async->sync)
2715                 complete(&async->wait);
2716         else
2717                 kfree(async);
2718 }
2719
2720 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2721                                  unsigned long count, int wait)
2722 {
2723         struct async_delayed_refs *async;
2724         int ret;
2725
2726         async = kmalloc(sizeof(*async), GFP_NOFS);
2727         if (!async)
2728                 return -ENOMEM;
2729
2730         async->root = root->fs_info->tree_root;
2731         async->count = count;
2732         async->error = 0;
2733         if (wait)
2734                 async->sync = 1;
2735         else
2736                 async->sync = 0;
2737         init_completion(&async->wait);
2738
2739         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2740                         delayed_ref_async_start, NULL, NULL);
2741
2742         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2743
2744         if (wait) {
2745                 wait_for_completion(&async->wait);
2746                 ret = async->error;
2747                 kfree(async);
2748                 return ret;
2749         }
2750         return 0;
2751 }
2752
2753 /*
2754  * this starts processing the delayed reference count updates and
2755  * extent insertions we have queued up so far.  count can be
2756  * 0, which means to process everything in the tree at the start
2757  * of the run (but not newly added entries), or it can be some target
2758  * number you'd like to process.
2759  *
2760  * Returns 0 on success or if called with an aborted transaction
2761  * Returns <0 on error and aborts the transaction
2762  */
2763 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2764                            struct btrfs_root *root, unsigned long count)
2765 {
2766         struct rb_node *node;
2767         struct btrfs_delayed_ref_root *delayed_refs;
2768         struct btrfs_delayed_ref_head *head;
2769         int ret;
2770         int run_all = count == (unsigned long)-1;
2771
2772         /* We'll clean this up in btrfs_cleanup_transaction */
2773         if (trans->aborted)
2774                 return 0;
2775
2776         if (root == root->fs_info->extent_root)
2777                 root = root->fs_info->tree_root;
2778
2779         delayed_refs = &trans->transaction->delayed_refs;
2780         if (count == 0)
2781                 count = atomic_read(&delayed_refs->num_entries) * 2;
2782
2783 again:
2784 #ifdef SCRAMBLE_DELAYED_REFS
2785         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2786 #endif
2787         ret = __btrfs_run_delayed_refs(trans, root, count);
2788         if (ret < 0) {
2789                 btrfs_abort_transaction(trans, root, ret);
2790                 return ret;
2791         }
2792
2793         if (run_all) {
2794                 if (!list_empty(&trans->new_bgs))
2795                         btrfs_create_pending_block_groups(trans, root);
2796
2797                 spin_lock(&delayed_refs->lock);
2798                 node = rb_first(&delayed_refs->href_root);
2799                 if (!node) {
2800                         spin_unlock(&delayed_refs->lock);
2801                         goto out;
2802                 }
2803                 count = (unsigned long)-1;
2804
2805                 while (node) {
2806                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2807                                         href_node);
2808                         if (btrfs_delayed_ref_is_head(&head->node)) {
2809                                 struct btrfs_delayed_ref_node *ref;
2810
2811                                 ref = &head->node;
2812                                 atomic_inc(&ref->refs);
2813
2814                                 spin_unlock(&delayed_refs->lock);
2815                                 /*
2816                                  * Mutex was contended, block until it's
2817                                  * released and try again
2818                                  */
2819                                 mutex_lock(&head->mutex);
2820                                 mutex_unlock(&head->mutex);
2821
2822                                 btrfs_put_delayed_ref(ref);
2823                                 cond_resched();
2824                                 goto again;
2825                         } else {
2826                                 WARN_ON(1);
2827                         }
2828                         node = rb_next(node);
2829                 }
2830                 spin_unlock(&delayed_refs->lock);
2831                 cond_resched();
2832                 goto again;
2833         }
2834 out:
2835         ret = btrfs_delayed_qgroup_accounting(trans, root->fs_info);
2836         if (ret)
2837                 return ret;
2838         assert_qgroups_uptodate(trans);
2839         return 0;
2840 }
2841
2842 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2843                                 struct btrfs_root *root,
2844                                 u64 bytenr, u64 num_bytes, u64 flags,
2845                                 int level, int is_data)
2846 {
2847         struct btrfs_delayed_extent_op *extent_op;
2848         int ret;
2849
2850         extent_op = btrfs_alloc_delayed_extent_op();
2851         if (!extent_op)
2852                 return -ENOMEM;
2853
2854         extent_op->flags_to_set = flags;
2855         extent_op->update_flags = 1;
2856         extent_op->update_key = 0;
2857         extent_op->is_data = is_data ? 1 : 0;
2858         extent_op->level = level;
2859
2860         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2861                                           num_bytes, extent_op);
2862         if (ret)
2863                 btrfs_free_delayed_extent_op(extent_op);
2864         return ret;
2865 }
2866
2867 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2868                                       struct btrfs_root *root,
2869                                       struct btrfs_path *path,
2870                                       u64 objectid, u64 offset, u64 bytenr)
2871 {
2872         struct btrfs_delayed_ref_head *head;
2873         struct btrfs_delayed_ref_node *ref;
2874         struct btrfs_delayed_data_ref *data_ref;
2875         struct btrfs_delayed_ref_root *delayed_refs;
2876         struct rb_node *node;
2877         int ret = 0;
2878
2879         delayed_refs = &trans->transaction->delayed_refs;
2880         spin_lock(&delayed_refs->lock);
2881         head = btrfs_find_delayed_ref_head(trans, bytenr);
2882         if (!head) {
2883                 spin_unlock(&delayed_refs->lock);
2884                 return 0;
2885         }
2886
2887         if (!mutex_trylock(&head->mutex)) {
2888                 atomic_inc(&head->node.refs);
2889                 spin_unlock(&delayed_refs->lock);
2890
2891                 btrfs_release_path(path);
2892
2893                 /*
2894                  * Mutex was contended, block until it's released and let
2895                  * caller try again
2896                  */
2897                 mutex_lock(&head->mutex);
2898                 mutex_unlock(&head->mutex);
2899                 btrfs_put_delayed_ref(&head->node);
2900                 return -EAGAIN;
2901         }
2902         spin_unlock(&delayed_refs->lock);
2903
2904         spin_lock(&head->lock);
2905         node = rb_first(&head->ref_root);
2906         while (node) {
2907                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2908                 node = rb_next(node);
2909
2910                 /* If it's a shared ref we know a cross reference exists */
2911                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2912                         ret = 1;
2913                         break;
2914                 }
2915
2916                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2917
2918                 /*
2919                  * If our ref doesn't match the one we're currently looking at
2920                  * then we have a cross reference.
2921                  */
2922                 if (data_ref->root != root->root_key.objectid ||
2923                     data_ref->objectid != objectid ||
2924                     data_ref->offset != offset) {
2925                         ret = 1;
2926                         break;
2927                 }
2928         }
2929         spin_unlock(&head->lock);
2930         mutex_unlock(&head->mutex);
2931         return ret;
2932 }
2933
2934 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2935                                         struct btrfs_root *root,
2936                                         struct btrfs_path *path,
2937                                         u64 objectid, u64 offset, u64 bytenr)
2938 {
2939         struct btrfs_root *extent_root = root->fs_info->extent_root;
2940         struct extent_buffer *leaf;
2941         struct btrfs_extent_data_ref *ref;
2942         struct btrfs_extent_inline_ref *iref;
2943         struct btrfs_extent_item *ei;
2944         struct btrfs_key key;
2945         u32 item_size;
2946         int ret;
2947
2948         key.objectid = bytenr;
2949         key.offset = (u64)-1;
2950         key.type = BTRFS_EXTENT_ITEM_KEY;
2951
2952         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2953         if (ret < 0)
2954                 goto out;
2955         BUG_ON(ret == 0); /* Corruption */
2956
2957         ret = -ENOENT;
2958         if (path->slots[0] == 0)
2959                 goto out;
2960
2961         path->slots[0]--;
2962         leaf = path->nodes[0];
2963         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2964
2965         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2966                 goto out;
2967
2968         ret = 1;
2969         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2970 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2971         if (item_size < sizeof(*ei)) {
2972                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2973                 goto out;
2974         }
2975 #endif
2976         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2977
2978         if (item_size != sizeof(*ei) +
2979             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2980                 goto out;
2981
2982         if (btrfs_extent_generation(leaf, ei) <=
2983             btrfs_root_last_snapshot(&root->root_item))
2984                 goto out;
2985
2986         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2987         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2988             BTRFS_EXTENT_DATA_REF_KEY)
2989                 goto out;
2990
2991         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2992         if (btrfs_extent_refs(leaf, ei) !=
2993             btrfs_extent_data_ref_count(leaf, ref) ||
2994             btrfs_extent_data_ref_root(leaf, ref) !=
2995             root->root_key.objectid ||
2996             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2997             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2998                 goto out;
2999
3000         ret = 0;
3001 out:
3002         return ret;
3003 }
3004
3005 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3006                           struct btrfs_root *root,
3007                           u64 objectid, u64 offset, u64 bytenr)
3008 {
3009         struct btrfs_path *path;
3010         int ret;
3011         int ret2;
3012
3013         path = btrfs_alloc_path();
3014         if (!path)
3015                 return -ENOENT;
3016
3017         do {
3018                 ret = check_committed_ref(trans, root, path, objectid,
3019                                           offset, bytenr);
3020                 if (ret && ret != -ENOENT)
3021                         goto out;
3022
3023                 ret2 = check_delayed_ref(trans, root, path, objectid,
3024                                          offset, bytenr);
3025         } while (ret2 == -EAGAIN);
3026
3027         if (ret2 && ret2 != -ENOENT) {
3028                 ret = ret2;
3029                 goto out;
3030         }
3031
3032         if (ret != -ENOENT || ret2 != -ENOENT)
3033                 ret = 0;
3034 out:
3035         btrfs_free_path(path);
3036         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3037                 WARN_ON(ret > 0);
3038         return ret;
3039 }
3040
3041 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3042                            struct btrfs_root *root,
3043                            struct extent_buffer *buf,
3044                            int full_backref, int inc)
3045 {
3046         u64 bytenr;
3047         u64 num_bytes;
3048         u64 parent;
3049         u64 ref_root;
3050         u32 nritems;
3051         struct btrfs_key key;
3052         struct btrfs_file_extent_item *fi;
3053         int i;
3054         int level;
3055         int ret = 0;
3056         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3057                             u64, u64, u64, u64, u64, u64, int);
3058
3059
3060         if (btrfs_test_is_dummy_root(root))
3061                 return 0;
3062
3063         ref_root = btrfs_header_owner(buf);
3064         nritems = btrfs_header_nritems(buf);
3065         level = btrfs_header_level(buf);
3066
3067         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3068                 return 0;
3069
3070         if (inc)
3071                 process_func = btrfs_inc_extent_ref;
3072         else
3073                 process_func = btrfs_free_extent;
3074
3075         if (full_backref)
3076                 parent = buf->start;
3077         else
3078                 parent = 0;
3079
3080         for (i = 0; i < nritems; i++) {
3081                 if (level == 0) {
3082                         btrfs_item_key_to_cpu(buf, &key, i);
3083                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3084                                 continue;
3085                         fi = btrfs_item_ptr(buf, i,
3086                                             struct btrfs_file_extent_item);
3087                         if (btrfs_file_extent_type(buf, fi) ==
3088                             BTRFS_FILE_EXTENT_INLINE)
3089                                 continue;
3090                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3091                         if (bytenr == 0)
3092                                 continue;
3093
3094                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3095                         key.offset -= btrfs_file_extent_offset(buf, fi);
3096                         ret = process_func(trans, root, bytenr, num_bytes,
3097                                            parent, ref_root, key.objectid,
3098                                            key.offset, 1);
3099                         if (ret)
3100                                 goto fail;
3101                 } else {
3102                         bytenr = btrfs_node_blockptr(buf, i);
3103                         num_bytes = root->nodesize;
3104                         ret = process_func(trans, root, bytenr, num_bytes,
3105                                            parent, ref_root, level - 1, 0,
3106                                            1);
3107                         if (ret)
3108                                 goto fail;
3109                 }
3110         }
3111         return 0;
3112 fail:
3113         return ret;
3114 }
3115
3116 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3117                   struct extent_buffer *buf, int full_backref)
3118 {
3119         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3120 }
3121
3122 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3123                   struct extent_buffer *buf, int full_backref)
3124 {
3125         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3126 }
3127
3128 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3129                                  struct btrfs_root *root,
3130                                  struct btrfs_path *path,
3131                                  struct btrfs_block_group_cache *cache)
3132 {
3133         int ret;
3134         struct btrfs_root *extent_root = root->fs_info->extent_root;
3135         unsigned long bi;
3136         struct extent_buffer *leaf;
3137
3138         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3139         if (ret) {
3140                 if (ret > 0)
3141                         ret = -ENOENT;
3142                 goto fail;
3143         }
3144
3145         leaf = path->nodes[0];
3146         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3147         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3148         btrfs_mark_buffer_dirty(leaf);
3149         btrfs_release_path(path);
3150 fail:
3151         if (ret)
3152                 btrfs_abort_transaction(trans, root, ret);
3153         return ret;
3154
3155 }
3156
3157 static struct btrfs_block_group_cache *
3158 next_block_group(struct btrfs_root *root,
3159                  struct btrfs_block_group_cache *cache)
3160 {
3161         struct rb_node *node;
3162
3163         spin_lock(&root->fs_info->block_group_cache_lock);
3164
3165         /* If our block group was removed, we need a full search. */
3166         if (RB_EMPTY_NODE(&cache->cache_node)) {
3167                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3168
3169                 spin_unlock(&root->fs_info->block_group_cache_lock);
3170                 btrfs_put_block_group(cache);
3171                 cache = btrfs_lookup_first_block_group(root->fs_info,
3172                                                        next_bytenr);
3173                 return cache;
3174         }
3175         node = rb_next(&cache->cache_node);
3176         btrfs_put_block_group(cache);
3177         if (node) {
3178                 cache = rb_entry(node, struct btrfs_block_group_cache,
3179                                  cache_node);
3180                 btrfs_get_block_group(cache);
3181         } else
3182                 cache = NULL;
3183         spin_unlock(&root->fs_info->block_group_cache_lock);
3184         return cache;
3185 }
3186
3187 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3188                             struct btrfs_trans_handle *trans,
3189                             struct btrfs_path *path)
3190 {
3191         struct btrfs_root *root = block_group->fs_info->tree_root;
3192         struct inode *inode = NULL;
3193         u64 alloc_hint = 0;
3194         int dcs = BTRFS_DC_ERROR;
3195         u64 num_pages = 0;
3196         int retries = 0;
3197         int ret = 0;
3198
3199         /*
3200          * If this block group is smaller than 100 megs don't bother caching the
3201          * block group.
3202          */
3203         if (block_group->key.offset < (100 * 1024 * 1024)) {
3204                 spin_lock(&block_group->lock);
3205                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3206                 spin_unlock(&block_group->lock);
3207                 return 0;
3208         }
3209
3210         if (trans->aborted)
3211                 return 0;
3212 again:
3213         inode = lookup_free_space_inode(root, block_group, path);
3214         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3215                 ret = PTR_ERR(inode);
3216                 btrfs_release_path(path);
3217                 goto out;
3218         }
3219
3220         if (IS_ERR(inode)) {
3221                 BUG_ON(retries);
3222                 retries++;
3223
3224                 if (block_group->ro)
3225                         goto out_free;
3226
3227                 ret = create_free_space_inode(root, trans, block_group, path);
3228                 if (ret)
3229                         goto out_free;
3230                 goto again;
3231         }
3232
3233         /* We've already setup this transaction, go ahead and exit */
3234         if (block_group->cache_generation == trans->transid &&
3235             i_size_read(inode)) {
3236                 dcs = BTRFS_DC_SETUP;
3237                 goto out_put;
3238         }
3239
3240         /*
3241          * We want to set the generation to 0, that way if anything goes wrong
3242          * from here on out we know not to trust this cache when we load up next
3243          * time.
3244          */
3245         BTRFS_I(inode)->generation = 0;
3246         ret = btrfs_update_inode(trans, root, inode);
3247         if (ret) {
3248                 /*
3249                  * So theoretically we could recover from this, simply set the
3250                  * super cache generation to 0 so we know to invalidate the
3251                  * cache, but then we'd have to keep track of the block groups
3252                  * that fail this way so we know we _have_ to reset this cache
3253                  * before the next commit or risk reading stale cache.  So to
3254                  * limit our exposure to horrible edge cases lets just abort the
3255                  * transaction, this only happens in really bad situations
3256                  * anyway.
3257                  */
3258                 btrfs_abort_transaction(trans, root, ret);
3259                 goto out_put;
3260         }
3261         WARN_ON(ret);
3262
3263         if (i_size_read(inode) > 0) {
3264                 ret = btrfs_check_trunc_cache_free_space(root,
3265                                         &root->fs_info->global_block_rsv);
3266                 if (ret)
3267                         goto out_put;
3268
3269                 ret = btrfs_truncate_free_space_cache(root, trans, inode);
3270                 if (ret)
3271                         goto out_put;
3272         }
3273
3274         spin_lock(&block_group->lock);
3275         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3276             !btrfs_test_opt(root, SPACE_CACHE) ||
3277             block_group->delalloc_bytes) {
3278                 /*
3279                  * don't bother trying to write stuff out _if_
3280                  * a) we're not cached,
3281                  * b) we're with nospace_cache mount option.
3282                  */
3283                 dcs = BTRFS_DC_WRITTEN;
3284                 spin_unlock(&block_group->lock);
3285                 goto out_put;
3286         }
3287         spin_unlock(&block_group->lock);
3288
3289         /*
3290          * Try to preallocate enough space based on how big the block group is.
3291          * Keep in mind this has to include any pinned space which could end up
3292          * taking up quite a bit since it's not folded into the other space
3293          * cache.
3294          */
3295         num_pages = div_u64(block_group->key.offset, 256 * 1024 * 1024);
3296         if (!num_pages)
3297                 num_pages = 1;
3298
3299         num_pages *= 16;
3300         num_pages *= PAGE_CACHE_SIZE;
3301
3302         ret = btrfs_check_data_free_space(inode, num_pages);
3303         if (ret)
3304                 goto out_put;
3305
3306         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3307                                               num_pages, num_pages,
3308                                               &alloc_hint);
3309         if (!ret)
3310                 dcs = BTRFS_DC_SETUP;
3311         btrfs_free_reserved_data_space(inode, num_pages);
3312
3313 out_put:
3314         iput(inode);
3315 out_free:
3316         btrfs_release_path(path);
3317 out:
3318         spin_lock(&block_group->lock);
3319         if (!ret && dcs == BTRFS_DC_SETUP)
3320                 block_group->cache_generation = trans->transid;
3321         block_group->disk_cache_state = dcs;
3322         spin_unlock(&block_group->lock);
3323
3324         return ret;
3325 }
3326
3327 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3328                             struct btrfs_root *root)
3329 {
3330         struct btrfs_block_group_cache *cache, *tmp;
3331         struct btrfs_transaction *cur_trans = trans->transaction;
3332         struct btrfs_path *path;
3333
3334         if (list_empty(&cur_trans->dirty_bgs) ||
3335             !btrfs_test_opt(root, SPACE_CACHE))
3336                 return 0;
3337
3338         path = btrfs_alloc_path();
3339         if (!path)
3340                 return -ENOMEM;
3341
3342         /* Could add new block groups, use _safe just in case */
3343         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3344                                  dirty_list) {
3345                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3346                         cache_save_setup(cache, trans, path);
3347         }
3348
3349         btrfs_free_path(path);
3350         return 0;
3351 }
3352
3353 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3354                                    struct btrfs_root *root)
3355 {
3356         struct btrfs_block_group_cache *cache;
3357         struct btrfs_transaction *cur_trans = trans->transaction;
3358         int ret = 0;
3359         struct btrfs_path *path;
3360
3361         if (list_empty(&cur_trans->dirty_bgs))
3362                 return 0;
3363
3364         path = btrfs_alloc_path();
3365         if (!path)
3366                 return -ENOMEM;
3367
3368         /*
3369          * We don't need the lock here since we are protected by the transaction
3370          * commit.  We want to do the cache_save_setup first and then run the
3371          * delayed refs to make sure we have the best chance at doing this all
3372          * in one shot.
3373          */
3374         while (!list_empty(&cur_trans->dirty_bgs)) {
3375                 cache = list_first_entry(&cur_trans->dirty_bgs,
3376                                          struct btrfs_block_group_cache,
3377                                          dirty_list);
3378                 list_del_init(&cache->dirty_list);
3379                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3380                         cache_save_setup(cache, trans, path);
3381                 if (!ret)
3382                         ret = btrfs_run_delayed_refs(trans, root,
3383                                                      (unsigned long) -1);
3384                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP)
3385                         btrfs_write_out_cache(root, trans, cache, path);
3386                 if (!ret)
3387                         ret = write_one_cache_group(trans, root, path, cache);
3388                 btrfs_put_block_group(cache);
3389         }
3390
3391         btrfs_free_path(path);
3392         return ret;
3393 }
3394
3395 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3396 {
3397         struct btrfs_block_group_cache *block_group;
3398         int readonly = 0;
3399
3400         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3401         if (!block_group || block_group->ro)
3402                 readonly = 1;
3403         if (block_group)
3404                 btrfs_put_block_group(block_group);
3405         return readonly;
3406 }
3407
3408 static const char *alloc_name(u64 flags)
3409 {
3410         switch (flags) {
3411         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3412                 return "mixed";
3413         case BTRFS_BLOCK_GROUP_METADATA:
3414                 return "metadata";
3415         case BTRFS_BLOCK_GROUP_DATA:
3416                 return "data";
3417         case BTRFS_BLOCK_GROUP_SYSTEM:
3418                 return "system";
3419         default:
3420                 WARN_ON(1);
3421                 return "invalid-combination";
3422         };
3423 }
3424
3425 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3426                              u64 total_bytes, u64 bytes_used,
3427                              struct btrfs_space_info **space_info)
3428 {
3429         struct btrfs_space_info *found;
3430         int i;
3431         int factor;
3432         int ret;
3433
3434         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3435                      BTRFS_BLOCK_GROUP_RAID10))
3436                 factor = 2;
3437         else
3438                 factor = 1;
3439
3440         found = __find_space_info(info, flags);
3441         if (found) {
3442                 spin_lock(&found->lock);
3443                 found->total_bytes += total_bytes;
3444                 found->disk_total += total_bytes * factor;
3445                 found->bytes_used += bytes_used;
3446                 found->disk_used += bytes_used * factor;
3447                 found->full = 0;
3448                 spin_unlock(&found->lock);
3449                 *space_info = found;
3450                 return 0;
3451         }
3452         found = kzalloc(sizeof(*found), GFP_NOFS);
3453         if (!found)
3454                 return -ENOMEM;
3455
3456         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3457         if (ret) {
3458                 kfree(found);
3459                 return ret;
3460         }
3461
3462         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3463                 INIT_LIST_HEAD(&found->block_groups[i]);
3464         init_rwsem(&found->groups_sem);
3465         spin_lock_init(&found->lock);
3466         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3467         found->total_bytes = total_bytes;
3468         found->disk_total = total_bytes * factor;
3469         found->bytes_used = bytes_used;
3470         found->disk_used = bytes_used * factor;
3471         found->bytes_pinned = 0;
3472         found->bytes_reserved = 0;
3473         found->bytes_readonly = 0;
3474         found->bytes_may_use = 0;
3475         found->full = 0;
3476         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3477         found->chunk_alloc = 0;
3478         found->flush = 0;
3479         init_waitqueue_head(&found->wait);
3480         INIT_LIST_HEAD(&found->ro_bgs);
3481
3482         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3483                                     info->space_info_kobj, "%s",
3484                                     alloc_name(found->flags));
3485         if (ret) {
3486                 kfree(found);
3487                 return ret;
3488         }
3489
3490         *space_info = found;
3491         list_add_rcu(&found->list, &info->space_info);
3492         if (flags & BTRFS_BLOCK_GROUP_DATA)
3493                 info->data_sinfo = found;
3494
3495         return ret;
3496 }
3497
3498 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3499 {
3500         u64 extra_flags = chunk_to_extended(flags) &
3501                                 BTRFS_EXTENDED_PROFILE_MASK;
3502
3503         write_seqlock(&fs_info->profiles_lock);
3504         if (flags & BTRFS_BLOCK_GROUP_DATA)
3505                 fs_info->avail_data_alloc_bits |= extra_flags;
3506         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3507                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3508         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3509                 fs_info->avail_system_alloc_bits |= extra_flags;
3510         write_sequnlock(&fs_info->profiles_lock);
3511 }
3512
3513 /*
3514  * returns target flags in extended format or 0 if restripe for this
3515  * chunk_type is not in progress
3516  *
3517  * should be called with either volume_mutex or balance_lock held
3518  */
3519 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3520 {
3521         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3522         u64 target = 0;
3523
3524         if (!bctl)
3525                 return 0;
3526
3527         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3528             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3529                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3530         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3531                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3532                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3533         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3534                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3535                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3536         }
3537
3538         return target;
3539 }
3540
3541 /*
3542  * @flags: available profiles in extended format (see ctree.h)
3543  *
3544  * Returns reduced profile in chunk format.  If profile changing is in
3545  * progress (either running or paused) picks the target profile (if it's
3546  * already available), otherwise falls back to plain reducing.
3547  */
3548 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3549 {
3550         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3551         u64 target;
3552         u64 tmp;
3553
3554         /*
3555          * see if restripe for this chunk_type is in progress, if so
3556          * try to reduce to the target profile
3557          */
3558         spin_lock(&root->fs_info->balance_lock);
3559         target = get_restripe_target(root->fs_info, flags);
3560         if (target) {
3561                 /* pick target profile only if it's already available */
3562                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3563                         spin_unlock(&root->fs_info->balance_lock);
3564                         return extended_to_chunk(target);
3565                 }
3566         }
3567         spin_unlock(&root->fs_info->balance_lock);
3568
3569         /* First, mask out the RAID levels which aren't possible */
3570         if (num_devices == 1)
3571                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3572                            BTRFS_BLOCK_GROUP_RAID5);
3573         if (num_devices < 3)
3574                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3575         if (num_devices < 4)
3576                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3577
3578         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3579                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3580                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3581         flags &= ~tmp;
3582
3583         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3584                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3585         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3586                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3587         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3588                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3589         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3590                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3591         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3592                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3593
3594         return extended_to_chunk(flags | tmp);
3595 }
3596
3597 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3598 {
3599         unsigned seq;
3600         u64 flags;
3601
3602         do {
3603                 flags = orig_flags;
3604                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3605
3606                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3607                         flags |= root->fs_info->avail_data_alloc_bits;
3608                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3609                         flags |= root->fs_info->avail_system_alloc_bits;
3610                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3611                         flags |= root->fs_info->avail_metadata_alloc_bits;
3612         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3613
3614         return btrfs_reduce_alloc_profile(root, flags);
3615 }
3616
3617 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3618 {
3619         u64 flags;
3620         u64 ret;
3621
3622         if (data)
3623                 flags = BTRFS_BLOCK_GROUP_DATA;
3624         else if (root == root->fs_info->chunk_root)
3625                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3626         else
3627                 flags = BTRFS_BLOCK_GROUP_METADATA;
3628
3629         ret = get_alloc_profile(root, flags);
3630         return ret;
3631 }
3632
3633 /*
3634  * This will check the space that the inode allocates from to make sure we have
3635  * enough space for bytes.
3636  */
3637 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3638 {
3639         struct btrfs_space_info *data_sinfo;
3640         struct btrfs_root *root = BTRFS_I(inode)->root;
3641         struct btrfs_fs_info *fs_info = root->fs_info;
3642         u64 used;
3643         int ret = 0, committed = 0;
3644
3645         /* make sure bytes are sectorsize aligned */
3646         bytes = ALIGN(bytes, root->sectorsize);
3647
3648         if (btrfs_is_free_space_inode(inode)) {
3649                 committed = 1;
3650                 ASSERT(current->journal_info);
3651         }
3652
3653         data_sinfo = fs_info->data_sinfo;
3654         if (!data_sinfo)
3655                 goto alloc;
3656
3657 again:
3658         /* make sure we have enough space to handle the data first */
3659         spin_lock(&data_sinfo->lock);
3660         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3661                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3662                 data_sinfo->bytes_may_use;
3663
3664         if (used + bytes > data_sinfo->total_bytes) {
3665                 struct btrfs_trans_handle *trans;
3666
3667                 /*
3668                  * if we don't have enough free bytes in this space then we need
3669                  * to alloc a new chunk.
3670                  */
3671                 if (!data_sinfo->full) {
3672                         u64 alloc_target;
3673
3674                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3675                         spin_unlock(&data_sinfo->lock);
3676 alloc:
3677                         alloc_target = btrfs_get_alloc_profile(root, 1);
3678                         /*
3679                          * It is ugly that we don't call nolock join
3680                          * transaction for the free space inode case here.
3681                          * But it is safe because we only do the data space
3682                          * reservation for the free space cache in the
3683                          * transaction context, the common join transaction
3684                          * just increase the counter of the current transaction
3685                          * handler, doesn't try to acquire the trans_lock of
3686                          * the fs.
3687                          */
3688                         trans = btrfs_join_transaction(root);
3689                         if (IS_ERR(trans))
3690                                 return PTR_ERR(trans);
3691
3692                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3693                                              alloc_target,
3694                                              CHUNK_ALLOC_NO_FORCE);
3695                         btrfs_end_transaction(trans, root);
3696                         if (ret < 0) {
3697                                 if (ret != -ENOSPC)
3698                                         return ret;
3699                                 else
3700                                         goto commit_trans;
3701                         }
3702
3703                         if (!data_sinfo)
3704                                 data_sinfo = fs_info->data_sinfo;
3705
3706                         goto again;
3707                 }
3708
3709                 /*
3710                  * If we don't have enough pinned space to deal with this
3711                  * allocation don't bother committing the transaction.
3712                  */
3713                 if (percpu_counter_compare(&data_sinfo->total_bytes_pinned,
3714                                            bytes) < 0)
3715                         committed = 1;
3716                 spin_unlock(&data_sinfo->lock);
3717
3718                 /* commit the current transaction and try again */
3719 commit_trans:
3720                 if (!committed &&
3721                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3722                         committed = 1;
3723
3724                         trans = btrfs_join_transaction(root);
3725                         if (IS_ERR(trans))
3726                                 return PTR_ERR(trans);
3727                         ret = btrfs_commit_transaction(trans, root);
3728                         if (ret)
3729                                 return ret;
3730                         goto again;
3731                 }
3732
3733                 trace_btrfs_space_reservation(root->fs_info,
3734                                               "space_info:enospc",
3735                                               data_sinfo->flags, bytes, 1);
3736                 return -ENOSPC;
3737         }
3738         data_sinfo->bytes_may_use += bytes;
3739         trace_btrfs_space_reservation(root->fs_info, "space_info",
3740                                       data_sinfo->flags, bytes, 1);
3741         spin_unlock(&data_sinfo->lock);
3742
3743         return 0;
3744 }
3745
3746 /*
3747  * Called if we need to clear a data reservation for this inode.
3748  */
3749 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3750 {
3751         struct btrfs_root *root = BTRFS_I(inode)->root;
3752         struct btrfs_space_info *data_sinfo;
3753
3754         /* make sure bytes are sectorsize aligned */
3755         bytes = ALIGN(bytes, root->sectorsize);
3756
3757         data_sinfo = root->fs_info->data_sinfo;
3758         spin_lock(&data_sinfo->lock);
3759         WARN_ON(data_sinfo->bytes_may_use < bytes);
3760         data_sinfo->bytes_may_use -= bytes;
3761         trace_btrfs_space_reservation(root->fs_info, "space_info",
3762                                       data_sinfo->flags, bytes, 0);
3763         spin_unlock(&data_sinfo->lock);
3764 }
3765
3766 static void force_metadata_allocation(struct btrfs_fs_info *info)
3767 {
3768         struct list_head *head = &info->space_info;
3769         struct btrfs_space_info *found;
3770
3771         rcu_read_lock();
3772         list_for_each_entry_rcu(found, head, list) {
3773                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3774                         found->force_alloc = CHUNK_ALLOC_FORCE;
3775         }
3776         rcu_read_unlock();
3777 }
3778
3779 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3780 {
3781         return (global->size << 1);
3782 }
3783
3784 static int should_alloc_chunk(struct btrfs_root *root,
3785                               struct btrfs_space_info *sinfo, int force)
3786 {
3787         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3788         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3789         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3790         u64 thresh;
3791
3792         if (force == CHUNK_ALLOC_FORCE)
3793                 return 1;
3794
3795         /*
3796          * We need to take into account the global rsv because for all intents
3797          * and purposes it's used space.  Don't worry about locking the
3798          * global_rsv, it doesn't change except when the transaction commits.
3799          */
3800         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3801                 num_allocated += calc_global_rsv_need_space(global_rsv);
3802
3803         /*
3804          * in limited mode, we want to have some free space up to
3805          * about 1% of the FS size.
3806          */
3807         if (force == CHUNK_ALLOC_LIMITED) {
3808                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3809                 thresh = max_t(u64, 64 * 1024 * 1024,
3810                                div_factor_fine(thresh, 1));
3811
3812                 if (num_bytes - num_allocated < thresh)
3813                         return 1;
3814         }
3815
3816         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3817                 return 0;
3818         return 1;
3819 }
3820
3821 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3822 {
3823         u64 num_dev;
3824
3825         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3826                     BTRFS_BLOCK_GROUP_RAID0 |
3827                     BTRFS_BLOCK_GROUP_RAID5 |
3828                     BTRFS_BLOCK_GROUP_RAID6))
3829                 num_dev = root->fs_info->fs_devices->rw_devices;
3830         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3831                 num_dev = 2;
3832         else
3833                 num_dev = 1;    /* DUP or single */
3834
3835         /* metadata for updaing devices and chunk tree */
3836         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3837 }
3838
3839 static void check_system_chunk(struct btrfs_trans_handle *trans,
3840                                struct btrfs_root *root, u64 type)
3841 {
3842         struct btrfs_space_info *info;
3843         u64 left;
3844         u64 thresh;
3845
3846         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3847         spin_lock(&info->lock);
3848         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3849                 info->bytes_reserved - info->bytes_readonly;
3850         spin_unlock(&info->lock);
3851
3852         thresh = get_system_chunk_thresh(root, type);
3853         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3854                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3855                         left, thresh, type);
3856                 dump_space_info(info, 0, 0);
3857         }
3858
3859         if (left < thresh) {
3860                 u64 flags;
3861
3862                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3863                 btrfs_alloc_chunk(trans, root, flags);
3864         }
3865 }
3866
3867 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3868                           struct btrfs_root *extent_root, u64 flags, int force)
3869 {
3870         struct btrfs_space_info *space_info;
3871         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3872         int wait_for_alloc = 0;
3873         int ret = 0;
3874
3875         /* Don't re-enter if we're already allocating a chunk */
3876         if (trans->allocating_chunk)
3877                 return -ENOSPC;
3878
3879         space_info = __find_space_info(extent_root->fs_info, flags);
3880         if (!space_info) {
3881                 ret = update_space_info(extent_root->fs_info, flags,
3882                                         0, 0, &space_info);
3883                 BUG_ON(ret); /* -ENOMEM */
3884         }
3885         BUG_ON(!space_info); /* Logic error */
3886
3887 again:
3888         spin_lock(&space_info->lock);
3889         if (force < space_info->force_alloc)
3890                 force = space_info->force_alloc;
3891         if (space_info->full) {
3892                 if (should_alloc_chunk(extent_root, space_info, force))
3893                         ret = -ENOSPC;
3894                 else
3895                         ret = 0;
3896                 spin_unlock(&space_info->lock);
3897                 return ret;
3898         }
3899
3900         if (!should_alloc_chunk(extent_root, space_info, force)) {
3901                 spin_unlock(&space_info->lock);
3902                 return 0;
3903         } else if (space_info->chunk_alloc) {
3904                 wait_for_alloc = 1;
3905         } else {
3906                 space_info->chunk_alloc = 1;
3907         }
3908
3909         spin_unlock(&space_info->lock);
3910
3911         mutex_lock(&fs_info->chunk_mutex);
3912
3913         /*
3914          * The chunk_mutex is held throughout the entirety of a chunk
3915          * allocation, so once we've acquired the chunk_mutex we know that the
3916          * other guy is done and we need to recheck and see if we should
3917          * allocate.
3918          */
3919         if (wait_for_alloc) {
3920                 mutex_unlock(&fs_info->chunk_mutex);
3921                 wait_for_alloc = 0;
3922                 goto again;
3923         }
3924
3925         trans->allocating_chunk = true;
3926
3927         /*
3928          * If we have mixed data/metadata chunks we want to make sure we keep
3929          * allocating mixed chunks instead of individual chunks.
3930          */
3931         if (btrfs_mixed_space_info(space_info))
3932                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3933
3934         /*
3935          * if we're doing a data chunk, go ahead and make sure that
3936          * we keep a reasonable number of metadata chunks allocated in the
3937          * FS as well.
3938          */
3939         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3940                 fs_info->data_chunk_allocations++;
3941                 if (!(fs_info->data_chunk_allocations %
3942                       fs_info->metadata_ratio))
3943                         force_metadata_allocation(fs_info);
3944         }
3945
3946         /*
3947          * Check if we have enough space in SYSTEM chunk because we may need
3948          * to update devices.
3949          */
3950         check_system_chunk(trans, extent_root, flags);
3951
3952         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3953         trans->allocating_chunk = false;
3954
3955         spin_lock(&space_info->lock);
3956         if (ret < 0 && ret != -ENOSPC)
3957                 goto out;
3958         if (ret)
3959                 space_info->full = 1;
3960         else
3961                 ret = 1;
3962
3963         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3964 out:
3965         space_info->chunk_alloc = 0;
3966         spin_unlock(&space_info->lock);
3967         mutex_unlock(&fs_info->chunk_mutex);
3968         return ret;
3969 }
3970
3971 static int can_overcommit(struct btrfs_root *root,
3972                           struct btrfs_space_info *space_info, u64 bytes,
3973                           enum btrfs_reserve_flush_enum flush)
3974 {
3975         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3976         u64 profile = btrfs_get_alloc_profile(root, 0);
3977         u64 space_size;
3978         u64 avail;
3979         u64 used;
3980
3981         used = space_info->bytes_used + space_info->bytes_reserved +
3982                 space_info->bytes_pinned + space_info->bytes_readonly;
3983
3984         /*
3985          * We only want to allow over committing if we have lots of actual space
3986          * free, but if we don't have enough space to handle the global reserve
3987          * space then we could end up having a real enospc problem when trying
3988          * to allocate a chunk or some other such important allocation.
3989          */
3990         spin_lock(&global_rsv->lock);
3991         space_size = calc_global_rsv_need_space(global_rsv);
3992         spin_unlock(&global_rsv->lock);
3993         if (used + space_size >= space_info->total_bytes)
3994                 return 0;
3995
3996         used += space_info->bytes_may_use;
3997
3998         spin_lock(&root->fs_info->free_chunk_lock);
3999         avail = root->fs_info->free_chunk_space;
4000         spin_unlock(&root->fs_info->free_chunk_lock);
4001
4002         /*
4003          * If we have dup, raid1 or raid10 then only half of the free
4004          * space is actually useable.  For raid56, the space info used
4005          * doesn't include the parity drive, so we don't have to
4006          * change the math
4007          */
4008         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4009                        BTRFS_BLOCK_GROUP_RAID1 |
4010                        BTRFS_BLOCK_GROUP_RAID10))
4011                 avail >>= 1;
4012
4013         /*
4014          * If we aren't flushing all things, let us overcommit up to
4015          * 1/2th of the space. If we can flush, don't let us overcommit
4016          * too much, let it overcommit up to 1/8 of the space.
4017          */
4018         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4019                 avail >>= 3;
4020         else
4021                 avail >>= 1;
4022
4023         if (used + bytes < space_info->total_bytes + avail)
4024                 return 1;
4025         return 0;
4026 }
4027
4028 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4029                                          unsigned long nr_pages, int nr_items)
4030 {
4031         struct super_block *sb = root->fs_info->sb;
4032
4033         if (down_read_trylock(&sb->s_umount)) {
4034                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4035                 up_read(&sb->s_umount);
4036         } else {
4037                 /*
4038                  * We needn't worry the filesystem going from r/w to r/o though
4039                  * we don't acquire ->s_umount mutex, because the filesystem
4040                  * should guarantee the delalloc inodes list be empty after
4041                  * the filesystem is readonly(all dirty pages are written to
4042                  * the disk).
4043                  */
4044                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4045                 if (!current->journal_info)
4046                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4047         }
4048 }
4049
4050 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4051 {
4052         u64 bytes;
4053         int nr;
4054
4055         bytes = btrfs_calc_trans_metadata_size(root, 1);
4056         nr = (int)div64_u64(to_reclaim, bytes);
4057         if (!nr)
4058                 nr = 1;
4059         return nr;
4060 }
4061
4062 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4063
4064 /*
4065  * shrink metadata reservation for delalloc
4066  */
4067 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4068                             bool wait_ordered)
4069 {
4070         struct btrfs_block_rsv *block_rsv;
4071         struct btrfs_space_info *space_info;
4072         struct btrfs_trans_handle *trans;
4073         u64 delalloc_bytes;
4074         u64 max_reclaim;
4075         long time_left;
4076         unsigned long nr_pages;
4077         int loops;
4078         int items;
4079         enum btrfs_reserve_flush_enum flush;
4080
4081         /* Calc the number of the pages we need flush for space reservation */
4082         items = calc_reclaim_items_nr(root, to_reclaim);
4083         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4084
4085         trans = (struct btrfs_trans_handle *)current->journal_info;
4086         block_rsv = &root->fs_info->delalloc_block_rsv;
4087         space_info = block_rsv->space_info;
4088
4089         delalloc_bytes = percpu_counter_sum_positive(
4090                                                 &root->fs_info->delalloc_bytes);
4091         if (delalloc_bytes == 0) {
4092                 if (trans)
4093                         return;
4094                 if (wait_ordered)
4095                         btrfs_wait_ordered_roots(root->fs_info, items);
4096                 return;
4097         }
4098
4099         loops = 0;
4100         while (delalloc_bytes && loops < 3) {
4101                 max_reclaim = min(delalloc_bytes, to_reclaim);
4102                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4103                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4104                 /*
4105                  * We need to wait for the async pages to actually start before
4106                  * we do anything.
4107                  */
4108                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4109                 if (!max_reclaim)
4110                         goto skip_async;
4111
4112                 if (max_reclaim <= nr_pages)
4113                         max_reclaim = 0;
4114                 else
4115                         max_reclaim -= nr_pages;
4116
4117                 wait_event(root->fs_info->async_submit_wait,
4118                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4119                            (int)max_reclaim);
4120 skip_async:
4121                 if (!trans)
4122                         flush = BTRFS_RESERVE_FLUSH_ALL;
4123                 else
4124                         flush = BTRFS_RESERVE_NO_FLUSH;
4125                 spin_lock(&space_info->lock);
4126                 if (can_overcommit(root, space_info, orig, flush)) {
4127                         spin_unlock(&space_info->lock);
4128                         break;
4129                 }
4130                 spin_unlock(&space_info->lock);
4131
4132                 loops++;
4133                 if (wait_ordered && !trans) {
4134                         btrfs_wait_ordered_roots(root->fs_info, items);
4135                 } else {
4136                         time_left = schedule_timeout_killable(1);
4137                         if (time_left)
4138                                 break;
4139                 }
4140                 delalloc_bytes = percpu_counter_sum_positive(
4141                                                 &root->fs_info->delalloc_bytes);
4142         }
4143 }
4144
4145 /**
4146  * maybe_commit_transaction - possibly commit the transaction if its ok to
4147  * @root - the root we're allocating for
4148  * @bytes - the number of bytes we want to reserve
4149  * @force - force the commit
4150  *
4151  * This will check to make sure that committing the transaction will actually
4152  * get us somewhere and then commit the transaction if it does.  Otherwise it
4153  * will return -ENOSPC.
4154  */
4155 static int may_commit_transaction(struct btrfs_root *root,
4156                                   struct btrfs_space_info *space_info,
4157                                   u64 bytes, int force)
4158 {
4159         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4160         struct btrfs_trans_handle *trans;
4161
4162         trans = (struct btrfs_trans_handle *)current->journal_info;
4163         if (trans)
4164                 return -EAGAIN;
4165
4166         if (force)
4167                 goto commit;
4168
4169         /* See if there is enough pinned space to make this reservation */
4170         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4171                                    bytes) >= 0)
4172                 goto commit;
4173
4174         /*
4175          * See if there is some space in the delayed insertion reservation for
4176          * this reservation.
4177          */
4178         if (space_info != delayed_rsv->space_info)
4179                 return -ENOSPC;
4180
4181         spin_lock(&delayed_rsv->lock);
4182         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4183                                    bytes - delayed_rsv->size) >= 0) {
4184                 spin_unlock(&delayed_rsv->lock);
4185                 return -ENOSPC;
4186         }
4187         spin_unlock(&delayed_rsv->lock);
4188
4189 commit:
4190         trans = btrfs_join_transaction(root);
4191         if (IS_ERR(trans))
4192                 return -ENOSPC;
4193
4194         return btrfs_commit_transaction(trans, root);
4195 }
4196
4197 enum flush_state {
4198         FLUSH_DELAYED_ITEMS_NR  =       1,
4199         FLUSH_DELAYED_ITEMS     =       2,
4200         FLUSH_DELALLOC          =       3,
4201         FLUSH_DELALLOC_WAIT     =       4,
4202         ALLOC_CHUNK             =       5,
4203         COMMIT_TRANS            =       6,
4204 };
4205
4206 static int flush_space(struct btrfs_root *root,
4207                        struct btrfs_space_info *space_info, u64 num_bytes,
4208                        u64 orig_bytes, int state)
4209 {
4210         struct btrfs_trans_handle *trans;
4211         int nr;
4212         int ret = 0;
4213
4214         switch (state) {
4215         case FLUSH_DELAYED_ITEMS_NR:
4216         case FLUSH_DELAYED_ITEMS:
4217                 if (state == FLUSH_DELAYED_ITEMS_NR)
4218                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4219                 else
4220                         nr = -1;
4221
4222                 trans = btrfs_join_transaction(root);
4223                 if (IS_ERR(trans)) {
4224                         ret = PTR_ERR(trans);
4225                         break;
4226                 }
4227                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4228                 btrfs_end_transaction(trans, root);
4229                 break;
4230         case FLUSH_DELALLOC:
4231         case FLUSH_DELALLOC_WAIT:
4232                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4233                                 state == FLUSH_DELALLOC_WAIT);
4234                 break;
4235         case ALLOC_CHUNK:
4236                 trans = btrfs_join_transaction(root);
4237                 if (IS_ERR(trans)) {
4238                         ret = PTR_ERR(trans);
4239                         break;
4240                 }
4241                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4242                                      btrfs_get_alloc_profile(root, 0),
4243                                      CHUNK_ALLOC_NO_FORCE);
4244                 btrfs_end_transaction(trans, root);
4245                 if (ret == -ENOSPC)
4246                         ret = 0;
4247                 break;
4248         case COMMIT_TRANS:
4249                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4250                 break;
4251         default:
4252                 ret = -ENOSPC;
4253                 break;
4254         }
4255
4256         return ret;
4257 }
4258
4259 static inline u64
4260 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4261                                  struct btrfs_space_info *space_info)
4262 {
4263         u64 used;
4264         u64 expected;
4265         u64 to_reclaim;
4266
4267         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4268                                 16 * 1024 * 1024);
4269         spin_lock(&space_info->lock);
4270         if (can_overcommit(root, space_info, to_reclaim,
4271                            BTRFS_RESERVE_FLUSH_ALL)) {
4272                 to_reclaim = 0;
4273                 goto out;
4274         }
4275
4276         used = space_info->bytes_used + space_info->bytes_reserved +
4277                space_info->bytes_pinned + space_info->bytes_readonly +
4278                space_info->bytes_may_use;
4279         if (can_overcommit(root, space_info, 1024 * 1024,
4280                            BTRFS_RESERVE_FLUSH_ALL))
4281                 expected = div_factor_fine(space_info->total_bytes, 95);
4282         else
4283                 expected = div_factor_fine(space_info->total_bytes, 90);
4284
4285         if (used > expected)
4286                 to_reclaim = used - expected;
4287         else
4288                 to_reclaim = 0;
4289         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4290                                      space_info->bytes_reserved);
4291 out:
4292         spin_unlock(&space_info->lock);
4293
4294         return to_reclaim;
4295 }
4296
4297 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4298                                         struct btrfs_fs_info *fs_info, u64 used)
4299 {
4300         return (used >= div_factor_fine(space_info->total_bytes, 98) &&
4301                 !btrfs_fs_closing(fs_info) &&
4302                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4303 }
4304
4305 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4306                                        struct btrfs_fs_info *fs_info,
4307                                        int flush_state)
4308 {
4309         u64 used;
4310
4311         spin_lock(&space_info->lock);
4312         /*
4313          * We run out of space and have not got any free space via flush_space,
4314          * so don't bother doing async reclaim.
4315          */
4316         if (flush_state > COMMIT_TRANS && space_info->full) {
4317                 spin_unlock(&space_info->lock);
4318                 return 0;
4319         }
4320
4321         used = space_info->bytes_used + space_info->bytes_reserved +
4322                space_info->bytes_pinned + space_info->bytes_readonly +
4323                space_info->bytes_may_use;
4324         if (need_do_async_reclaim(space_info, fs_info, used)) {
4325                 spin_unlock(&space_info->lock);
4326                 return 1;
4327         }
4328         spin_unlock(&space_info->lock);
4329
4330         return 0;
4331 }
4332
4333 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4334 {
4335         struct btrfs_fs_info *fs_info;
4336         struct btrfs_space_info *space_info;
4337         u64 to_reclaim;
4338         int flush_state;
4339
4340         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4341         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4342
4343         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4344                                                       space_info);
4345         if (!to_reclaim)
4346                 return;
4347
4348         flush_state = FLUSH_DELAYED_ITEMS_NR;
4349         do {
4350                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4351                             to_reclaim, flush_state);
4352                 flush_state++;
4353                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4354                                                  flush_state))
4355                         return;
4356         } while (flush_state <= COMMIT_TRANS);
4357
4358         if (btrfs_need_do_async_reclaim(space_info, fs_info, flush_state))
4359                 queue_work(system_unbound_wq, work);
4360 }
4361
4362 void btrfs_init_async_reclaim_work(struct work_struct *work)
4363 {
4364         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4365 }
4366
4367 /**
4368  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4369  * @root - the root we're allocating for
4370  * @block_rsv - the block_rsv we're allocating for
4371  * @orig_bytes - the number of bytes we want
4372  * @flush - whether or not we can flush to make our reservation
4373  *
4374  * This will reserve orgi_bytes number of bytes from the space info associated
4375  * with the block_rsv.  If there is not enough space it will make an attempt to
4376  * flush out space to make room.  It will do this by flushing delalloc if
4377  * possible or committing the transaction.  If flush is 0 then no attempts to
4378  * regain reservations will be made and this will fail if there is not enough
4379  * space already.
4380  */
4381 static int reserve_metadata_bytes(struct btrfs_root *root,
4382                                   struct btrfs_block_rsv *block_rsv,
4383                                   u64 orig_bytes,
4384                                   enum btrfs_reserve_flush_enum flush)
4385 {
4386         struct btrfs_space_info *space_info = block_rsv->space_info;
4387         u64 used;
4388         u64 num_bytes = orig_bytes;
4389         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4390         int ret = 0;
4391         bool flushing = false;
4392
4393 again:
4394         ret = 0;
4395         spin_lock(&space_info->lock);
4396         /*
4397          * We only want to wait if somebody other than us is flushing and we
4398          * are actually allowed to flush all things.
4399          */
4400         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4401                space_info->flush) {
4402                 spin_unlock(&space_info->lock);
4403                 /*
4404                  * If we have a trans handle we can't wait because the flusher
4405                  * may have to commit the transaction, which would mean we would
4406                  * deadlock since we are waiting for the flusher to finish, but
4407                  * hold the current transaction open.
4408                  */
4409                 if (current->journal_info)
4410                         return -EAGAIN;
4411                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4412                 /* Must have been killed, return */
4413                 if (ret)
4414                         return -EINTR;
4415
4416                 spin_lock(&space_info->lock);
4417         }
4418
4419         ret = -ENOSPC;
4420         used = space_info->bytes_used + space_info->bytes_reserved +
4421                 space_info->bytes_pinned + space_info->bytes_readonly +
4422                 space_info->bytes_may_use;
4423
4424         /*
4425          * The idea here is that we've not already over-reserved the block group
4426          * then we can go ahead and save our reservation first and then start
4427          * flushing if we need to.  Otherwise if we've already overcommitted
4428          * lets start flushing stuff first and then come back and try to make
4429          * our reservation.
4430          */
4431         if (used <= space_info->total_bytes) {
4432                 if (used + orig_bytes <= space_info->total_bytes) {
4433                         space_info->bytes_may_use += orig_bytes;
4434                         trace_btrfs_space_reservation(root->fs_info,
4435                                 "space_info", space_info->flags, orig_bytes, 1);
4436                         ret = 0;
4437                 } else {
4438                         /*
4439                          * Ok set num_bytes to orig_bytes since we aren't
4440                          * overocmmitted, this way we only try and reclaim what
4441                          * we need.
4442                          */
4443                         num_bytes = orig_bytes;
4444                 }
4445         } else {
4446                 /*
4447                  * Ok we're over committed, set num_bytes to the overcommitted
4448                  * amount plus the amount of bytes that we need for this
4449                  * reservation.
4450                  */
4451                 num_bytes = used - space_info->total_bytes +
4452                         (orig_bytes * 2);
4453         }
4454
4455         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4456                 space_info->bytes_may_use += orig_bytes;
4457                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4458                                               space_info->flags, orig_bytes,
4459                                               1);
4460                 ret = 0;
4461         }
4462
4463         /*
4464          * Couldn't make our reservation, save our place so while we're trying
4465          * to reclaim space we can actually use it instead of somebody else
4466          * stealing it from us.
4467          *
4468          * We make the other tasks wait for the flush only when we can flush
4469          * all things.
4470          */
4471         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4472                 flushing = true;
4473                 space_info->flush = 1;
4474         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4475                 used += orig_bytes;
4476                 /*
4477                  * We will do the space reservation dance during log replay,
4478                  * which means we won't have fs_info->fs_root set, so don't do
4479                  * the async reclaim as we will panic.
4480                  */
4481                 if (!root->fs_info->log_root_recovering &&
4482                     need_do_async_reclaim(space_info, root->fs_info, used) &&
4483                     !work_busy(&root->fs_info->async_reclaim_work))
4484                         queue_work(system_unbound_wq,
4485                                    &root->fs_info->async_reclaim_work);
4486         }
4487         spin_unlock(&space_info->lock);
4488
4489         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4490                 goto out;
4491
4492         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4493                           flush_state);
4494         flush_state++;
4495
4496         /*
4497          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4498          * would happen. So skip delalloc flush.
4499          */
4500         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4501             (flush_state == FLUSH_DELALLOC ||
4502              flush_state == FLUSH_DELALLOC_WAIT))
4503                 flush_state = ALLOC_CHUNK;
4504
4505         if (!ret)
4506                 goto again;
4507         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4508                  flush_state < COMMIT_TRANS)
4509                 goto again;
4510         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4511                  flush_state <= COMMIT_TRANS)
4512                 goto again;
4513
4514 out:
4515         if (ret == -ENOSPC &&
4516             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4517                 struct btrfs_block_rsv *global_rsv =
4518                         &root->fs_info->global_block_rsv;
4519
4520                 if (block_rsv != global_rsv &&
4521                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4522                         ret = 0;
4523         }
4524         if (ret == -ENOSPC)
4525                 trace_btrfs_space_reservation(root->fs_info,
4526                                               "space_info:enospc",
4527                                               space_info->flags, orig_bytes, 1);
4528         if (flushing) {
4529                 spin_lock(&space_info->lock);
4530                 space_info->flush = 0;
4531                 wake_up_all(&space_info->wait);
4532                 spin_unlock(&space_info->lock);
4533         }
4534         return ret;
4535 }
4536
4537 static struct btrfs_block_rsv *get_block_rsv(
4538                                         const struct btrfs_trans_handle *trans,
4539                                         const struct btrfs_root *root)
4540 {
4541         struct btrfs_block_rsv *block_rsv = NULL;
4542
4543         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4544                 block_rsv = trans->block_rsv;
4545
4546         if (root == root->fs_info->csum_root && trans->adding_csums)
4547                 block_rsv = trans->block_rsv;
4548
4549         if (root == root->fs_info->uuid_root)
4550                 block_rsv = trans->block_rsv;
4551
4552         if (!block_rsv)
4553                 block_rsv = root->block_rsv;
4554
4555         if (!block_rsv)
4556                 block_rsv = &root->fs_info->empty_block_rsv;
4557
4558         return block_rsv;
4559 }
4560
4561 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4562                                u64 num_bytes)
4563 {
4564         int ret = -ENOSPC;
4565         spin_lock(&block_rsv->lock);
4566         if (block_rsv->reserved >= num_bytes) {
4567                 block_rsv->reserved -= num_bytes;
4568                 if (block_rsv->reserved < block_rsv->size)
4569                         block_rsv->full = 0;
4570                 ret = 0;
4571         }
4572         spin_unlock(&block_rsv->lock);
4573         return ret;
4574 }
4575
4576 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4577                                 u64 num_bytes, int update_size)
4578 {
4579         spin_lock(&block_rsv->lock);
4580         block_rsv->reserved += num_bytes;
4581         if (update_size)
4582                 block_rsv->size += num_bytes;
4583         else if (block_rsv->reserved >= block_rsv->size)
4584                 block_rsv->full = 1;
4585         spin_unlock(&block_rsv->lock);
4586 }
4587
4588 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4589                              struct btrfs_block_rsv *dest, u64 num_bytes,
4590                              int min_factor)
4591 {
4592         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4593         u64 min_bytes;
4594
4595         if (global_rsv->space_info != dest->space_info)
4596                 return -ENOSPC;
4597
4598         spin_lock(&global_rsv->lock);
4599         min_bytes = div_factor(global_rsv->size, min_factor);
4600         if (global_rsv->reserved < min_bytes + num_bytes) {
4601                 spin_unlock(&global_rsv->lock);
4602                 return -ENOSPC;
4603         }
4604         global_rsv->reserved -= num_bytes;
4605         if (global_rsv->reserved < global_rsv->size)
4606                 global_rsv->full = 0;
4607         spin_unlock(&global_rsv->lock);
4608
4609         block_rsv_add_bytes(dest, num_bytes, 1);
4610         return 0;
4611 }
4612
4613 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4614                                     struct btrfs_block_rsv *block_rsv,
4615                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4616 {
4617         struct btrfs_space_info *space_info = block_rsv->space_info;
4618
4619         spin_lock(&block_rsv->lock);
4620         if (num_bytes == (u64)-1)
4621                 num_bytes = block_rsv->size;
4622         block_rsv->size -= num_bytes;
4623         if (block_rsv->reserved >= block_rsv->size) {
4624                 num_bytes = block_rsv->reserved - block_rsv->size;
4625                 block_rsv->reserved = block_rsv->size;
4626                 block_rsv->full = 1;
4627         } else {
4628                 num_bytes = 0;
4629         }
4630         spin_unlock(&block_rsv->lock);
4631
4632         if (num_bytes > 0) {
4633                 if (dest) {
4634                         spin_lock(&dest->lock);
4635                         if (!dest->full) {
4636                                 u64 bytes_to_add;
4637
4638                                 bytes_to_add = dest->size - dest->reserved;
4639                                 bytes_to_add = min(num_bytes, bytes_to_add);
4640                                 dest->reserved += bytes_to_add;
4641                                 if (dest->reserved >= dest->size)
4642                                         dest->full = 1;
4643                                 num_bytes -= bytes_to_add;
4644                         }
4645                         spin_unlock(&dest->lock);
4646                 }
4647                 if (num_bytes) {
4648                         spin_lock(&space_info->lock);
4649                         space_info->bytes_may_use -= num_bytes;
4650                         trace_btrfs_space_reservation(fs_info, "space_info",
4651                                         space_info->flags, num_bytes, 0);
4652                         spin_unlock(&space_info->lock);
4653                 }
4654         }
4655 }
4656
4657 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4658                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4659 {
4660         int ret;
4661
4662         ret = block_rsv_use_bytes(src, num_bytes);
4663         if (ret)
4664                 return ret;
4665
4666         block_rsv_add_bytes(dst, num_bytes, 1);
4667         return 0;
4668 }
4669
4670 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4671 {
4672         memset(rsv, 0, sizeof(*rsv));
4673         spin_lock_init(&rsv->lock);
4674         rsv->type = type;
4675 }
4676
4677 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4678                                               unsigned short type)
4679 {
4680         struct btrfs_block_rsv *block_rsv;
4681         struct btrfs_fs_info *fs_info = root->fs_info;
4682
4683         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4684         if (!block_rsv)
4685                 return NULL;
4686
4687         btrfs_init_block_rsv(block_rsv, type);
4688         block_rsv->space_info = __find_space_info(fs_info,
4689                                                   BTRFS_BLOCK_GROUP_METADATA);
4690         return block_rsv;
4691 }
4692
4693 void btrfs_free_block_rsv(struct btrfs_root *root,
4694                           struct btrfs_block_rsv *rsv)
4695 {
4696         if (!rsv)
4697                 return;
4698         btrfs_block_rsv_release(root, rsv, (u64)-1);
4699         kfree(rsv);
4700 }
4701
4702 int btrfs_block_rsv_add(struct btrfs_root *root,
4703                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4704                         enum btrfs_reserve_flush_enum flush)
4705 {
4706         int ret;
4707
4708         if (num_bytes == 0)
4709                 return 0;
4710
4711         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4712         if (!ret) {
4713                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4714                 return 0;
4715         }
4716
4717         return ret;
4718 }
4719
4720 int btrfs_block_rsv_check(struct btrfs_root *root,
4721                           struct btrfs_block_rsv *block_rsv, int min_factor)
4722 {
4723         u64 num_bytes = 0;
4724         int ret = -ENOSPC;
4725
4726         if (!block_rsv)
4727                 return 0;
4728
4729         spin_lock(&block_rsv->lock);
4730         num_bytes = div_factor(block_rsv->size, min_factor);
4731         if (block_rsv->reserved >= num_bytes)
4732                 ret = 0;
4733         spin_unlock(&block_rsv->lock);
4734
4735         return ret;
4736 }
4737
4738 int btrfs_block_rsv_refill(struct btrfs_root *root,
4739                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4740                            enum btrfs_reserve_flush_enum flush)
4741 {
4742         u64 num_bytes = 0;
4743         int ret = -ENOSPC;
4744
4745         if (!block_rsv)
4746                 return 0;
4747
4748         spin_lock(&block_rsv->lock);
4749         num_bytes = min_reserved;
4750         if (block_rsv->reserved >= num_bytes)
4751                 ret = 0;
4752         else
4753                 num_bytes -= block_rsv->reserved;
4754         spin_unlock(&block_rsv->lock);
4755
4756         if (!ret)
4757                 return 0;
4758
4759         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4760         if (!ret) {
4761                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4762                 return 0;
4763         }
4764
4765         return ret;
4766 }
4767
4768 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4769                             struct btrfs_block_rsv *dst_rsv,
4770                             u64 num_bytes)
4771 {
4772         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4773 }
4774
4775 void btrfs_block_rsv_release(struct btrfs_root *root,
4776                              struct btrfs_block_rsv *block_rsv,
4777                              u64 num_bytes)
4778 {
4779         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4780         if (global_rsv == block_rsv ||
4781             block_rsv->space_info != global_rsv->space_info)
4782                 global_rsv = NULL;
4783         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4784                                 num_bytes);
4785 }
4786
4787 /*
4788  * helper to calculate size of global block reservation.
4789  * the desired value is sum of space used by extent tree,
4790  * checksum tree and root tree
4791  */
4792 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4793 {
4794         struct btrfs_space_info *sinfo;
4795         u64 num_bytes;
4796         u64 meta_used;
4797         u64 data_used;
4798         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4799
4800         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4801         spin_lock(&sinfo->lock);
4802         data_used = sinfo->bytes_used;
4803         spin_unlock(&sinfo->lock);
4804
4805         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4806         spin_lock(&sinfo->lock);
4807         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4808                 data_used = 0;
4809         meta_used = sinfo->bytes_used;
4810         spin_unlock(&sinfo->lock);
4811
4812         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4813                     csum_size * 2;
4814         num_bytes += div_u64(data_used + meta_used, 50);
4815
4816         if (num_bytes * 3 > meta_used)
4817                 num_bytes = div_u64(meta_used, 3);
4818
4819         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
4820 }
4821
4822 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4823 {
4824         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4825         struct btrfs_space_info *sinfo = block_rsv->space_info;
4826         u64 num_bytes;
4827
4828         num_bytes = calc_global_metadata_size(fs_info);
4829
4830         spin_lock(&sinfo->lock);
4831         spin_lock(&block_rsv->lock);
4832
4833         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4834
4835         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4836                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4837                     sinfo->bytes_may_use;
4838
4839         if (sinfo->total_bytes > num_bytes) {
4840                 num_bytes = sinfo->total_bytes - num_bytes;
4841                 block_rsv->reserved += num_bytes;
4842                 sinfo->bytes_may_use += num_bytes;
4843                 trace_btrfs_space_reservation(fs_info, "space_info",
4844                                       sinfo->flags, num_bytes, 1);
4845         }
4846
4847         if (block_rsv->reserved >= block_rsv->size) {
4848                 num_bytes = block_rsv->reserved - block_rsv->size;
4849                 sinfo->bytes_may_use -= num_bytes;
4850                 trace_btrfs_space_reservation(fs_info, "space_info",
4851                                       sinfo->flags, num_bytes, 0);
4852                 block_rsv->reserved = block_rsv->size;
4853                 block_rsv->full = 1;
4854         }
4855
4856         spin_unlock(&block_rsv->lock);
4857         spin_unlock(&sinfo->lock);
4858 }
4859
4860 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4861 {
4862         struct btrfs_space_info *space_info;
4863
4864         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4865         fs_info->chunk_block_rsv.space_info = space_info;
4866
4867         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4868         fs_info->global_block_rsv.space_info = space_info;
4869         fs_info->delalloc_block_rsv.space_info = space_info;
4870         fs_info->trans_block_rsv.space_info = space_info;
4871         fs_info->empty_block_rsv.space_info = space_info;
4872         fs_info->delayed_block_rsv.space_info = space_info;
4873
4874         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4875         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4876         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4877         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4878         if (fs_info->quota_root)
4879                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
4880         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4881
4882         update_global_block_rsv(fs_info);
4883 }
4884
4885 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4886 {
4887         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4888                                 (u64)-1);
4889         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4890         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4891         WARN_ON(fs_info->trans_block_rsv.size > 0);
4892         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4893         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4894         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4895         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4896         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4897 }
4898
4899 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4900                                   struct btrfs_root *root)
4901 {
4902         if (!trans->block_rsv)
4903                 return;
4904
4905         if (!trans->bytes_reserved)
4906                 return;
4907
4908         trace_btrfs_space_reservation(root->fs_info, "transaction",
4909                                       trans->transid, trans->bytes_reserved, 0);
4910         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4911         trans->bytes_reserved = 0;
4912 }
4913
4914 /* Can only return 0 or -ENOSPC */
4915 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4916                                   struct inode *inode)
4917 {
4918         struct btrfs_root *root = BTRFS_I(inode)->root;
4919         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4920         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4921
4922         /*
4923          * We need to hold space in order to delete our orphan item once we've
4924          * added it, so this takes the reservation so we can release it later
4925          * when we are truly done with the orphan item.
4926          */
4927         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4928         trace_btrfs_space_reservation(root->fs_info, "orphan",
4929                                       btrfs_ino(inode), num_bytes, 1);
4930         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4931 }
4932
4933 void btrfs_orphan_release_metadata(struct inode *inode)
4934 {
4935         struct btrfs_root *root = BTRFS_I(inode)->root;
4936         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4937         trace_btrfs_space_reservation(root->fs_info, "orphan",
4938                                       btrfs_ino(inode), num_bytes, 0);
4939         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4940 }
4941
4942 /*
4943  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4944  * root: the root of the parent directory
4945  * rsv: block reservation
4946  * items: the number of items that we need do reservation
4947  * qgroup_reserved: used to return the reserved size in qgroup
4948  *
4949  * This function is used to reserve the space for snapshot/subvolume
4950  * creation and deletion. Those operations are different with the
4951  * common file/directory operations, they change two fs/file trees
4952  * and root tree, the number of items that the qgroup reserves is
4953  * different with the free space reservation. So we can not use
4954  * the space reseravtion mechanism in start_transaction().
4955  */
4956 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4957                                      struct btrfs_block_rsv *rsv,
4958                                      int items,
4959                                      u64 *qgroup_reserved,
4960                                      bool use_global_rsv)
4961 {
4962         u64 num_bytes;
4963         int ret;
4964         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4965
4966         if (root->fs_info->quota_enabled) {
4967                 /* One for parent inode, two for dir entries */
4968                 num_bytes = 3 * root->nodesize;
4969                 ret = btrfs_qgroup_reserve(root, num_bytes);
4970                 if (ret)
4971                         return ret;
4972         } else {
4973                 num_bytes = 0;
4974         }
4975
4976         *qgroup_reserved = num_bytes;
4977
4978         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4979         rsv->space_info = __find_space_info(root->fs_info,
4980                                             BTRFS_BLOCK_GROUP_METADATA);
4981         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4982                                   BTRFS_RESERVE_FLUSH_ALL);
4983
4984         if (ret == -ENOSPC && use_global_rsv)
4985                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
4986
4987         if (ret) {
4988                 if (*qgroup_reserved)
4989                         btrfs_qgroup_free(root, *qgroup_reserved);
4990         }
4991
4992         return ret;
4993 }
4994
4995 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4996                                       struct btrfs_block_rsv *rsv,
4997                                       u64 qgroup_reserved)
4998 {
4999         btrfs_block_rsv_release(root, rsv, (u64)-1);
5000         if (qgroup_reserved)
5001                 btrfs_qgroup_free(root, qgroup_reserved);
5002 }
5003
5004 /**
5005  * drop_outstanding_extent - drop an outstanding extent
5006  * @inode: the inode we're dropping the extent for
5007  * @num_bytes: the number of bytes we're relaseing.
5008  *
5009  * This is called when we are freeing up an outstanding extent, either called
5010  * after an error or after an extent is written.  This will return the number of
5011  * reserved extents that need to be freed.  This must be called with
5012  * BTRFS_I(inode)->lock held.
5013  */
5014 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5015 {
5016         unsigned drop_inode_space = 0;
5017         unsigned dropped_extents = 0;
5018         unsigned num_extents = 0;
5019
5020         num_extents = (unsigned)div64_u64(num_bytes +
5021                                           BTRFS_MAX_EXTENT_SIZE - 1,
5022                                           BTRFS_MAX_EXTENT_SIZE);
5023         ASSERT(num_extents);
5024         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5025         BTRFS_I(inode)->outstanding_extents -= num_extents;
5026
5027         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5028             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5029                                &BTRFS_I(inode)->runtime_flags))
5030                 drop_inode_space = 1;
5031
5032         /*
5033          * If we have more or the same amount of outsanding extents than we have
5034          * reserved then we need to leave the reserved extents count alone.
5035          */
5036         if (BTRFS_I(inode)->outstanding_extents >=
5037             BTRFS_I(inode)->reserved_extents)
5038                 return drop_inode_space;
5039
5040         dropped_extents = BTRFS_I(inode)->reserved_extents -
5041                 BTRFS_I(inode)->outstanding_extents;
5042         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5043         return dropped_extents + drop_inode_space;
5044 }
5045
5046 /**
5047  * calc_csum_metadata_size - return the amount of metada space that must be
5048  *      reserved/free'd for the given bytes.
5049  * @inode: the inode we're manipulating
5050  * @num_bytes: the number of bytes in question
5051  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5052  *
5053  * This adjusts the number of csum_bytes in the inode and then returns the
5054  * correct amount of metadata that must either be reserved or freed.  We
5055  * calculate how many checksums we can fit into one leaf and then divide the
5056  * number of bytes that will need to be checksumed by this value to figure out
5057  * how many checksums will be required.  If we are adding bytes then the number
5058  * may go up and we will return the number of additional bytes that must be
5059  * reserved.  If it is going down we will return the number of bytes that must
5060  * be freed.
5061  *
5062  * This must be called with BTRFS_I(inode)->lock held.
5063  */
5064 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5065                                    int reserve)
5066 {
5067         struct btrfs_root *root = BTRFS_I(inode)->root;
5068         u64 csum_size;
5069         int num_csums_per_leaf;
5070         int num_csums;
5071         int old_csums;
5072
5073         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5074             BTRFS_I(inode)->csum_bytes == 0)
5075                 return 0;
5076
5077         old_csums = (int)div_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
5078         if (reserve)
5079                 BTRFS_I(inode)->csum_bytes += num_bytes;
5080         else
5081                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5082         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
5083         num_csums_per_leaf = (int)div_u64(csum_size,
5084                                             sizeof(struct btrfs_csum_item) +
5085                                             sizeof(struct btrfs_disk_key));
5086         num_csums = (int)div_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
5087         num_csums = num_csums + num_csums_per_leaf - 1;
5088         num_csums = num_csums / num_csums_per_leaf;
5089
5090         old_csums = old_csums + num_csums_per_leaf - 1;
5091         old_csums = old_csums / num_csums_per_leaf;
5092
5093         /* No change, no need to reserve more */
5094         if (old_csums == num_csums)
5095                 return 0;
5096
5097         if (reserve)
5098                 return btrfs_calc_trans_metadata_size(root,
5099                                                       num_csums - old_csums);
5100
5101         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5102 }
5103
5104 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5105 {
5106         struct btrfs_root *root = BTRFS_I(inode)->root;
5107         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5108         u64 to_reserve = 0;
5109         u64 csum_bytes;
5110         unsigned nr_extents = 0;
5111         int extra_reserve = 0;
5112         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5113         int ret = 0;
5114         bool delalloc_lock = true;
5115         u64 to_free = 0;
5116         unsigned dropped;
5117
5118         /* If we are a free space inode we need to not flush since we will be in
5119          * the middle of a transaction commit.  We also don't need the delalloc
5120          * mutex since we won't race with anybody.  We need this mostly to make
5121          * lockdep shut its filthy mouth.
5122          */
5123         if (btrfs_is_free_space_inode(inode)) {
5124                 flush = BTRFS_RESERVE_NO_FLUSH;
5125                 delalloc_lock = false;
5126         }
5127
5128         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5129             btrfs_transaction_in_commit(root->fs_info))
5130                 schedule_timeout(1);
5131
5132         if (delalloc_lock)
5133                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5134
5135         num_bytes = ALIGN(num_bytes, root->sectorsize);
5136
5137         spin_lock(&BTRFS_I(inode)->lock);
5138         nr_extents = (unsigned)div64_u64(num_bytes +
5139                                          BTRFS_MAX_EXTENT_SIZE - 1,
5140                                          BTRFS_MAX_EXTENT_SIZE);
5141         BTRFS_I(inode)->outstanding_extents += nr_extents;
5142         nr_extents = 0;
5143
5144         if (BTRFS_I(inode)->outstanding_extents >
5145             BTRFS_I(inode)->reserved_extents)
5146                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5147                         BTRFS_I(inode)->reserved_extents;
5148
5149         /*
5150          * Add an item to reserve for updating the inode when we complete the
5151          * delalloc io.
5152          */
5153         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5154                       &BTRFS_I(inode)->runtime_flags)) {
5155                 nr_extents++;
5156                 extra_reserve = 1;
5157         }
5158
5159         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5160         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5161         csum_bytes = BTRFS_I(inode)->csum_bytes;
5162         spin_unlock(&BTRFS_I(inode)->lock);
5163
5164         if (root->fs_info->quota_enabled) {
5165                 ret = btrfs_qgroup_reserve(root, num_bytes +
5166                                            nr_extents * root->nodesize);
5167                 if (ret)
5168                         goto out_fail;
5169         }
5170
5171         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5172         if (unlikely(ret)) {
5173                 if (root->fs_info->quota_enabled)
5174                         btrfs_qgroup_free(root, num_bytes +
5175                                                 nr_extents * root->nodesize);
5176                 goto out_fail;
5177         }
5178
5179         spin_lock(&BTRFS_I(inode)->lock);
5180         if (extra_reserve) {
5181                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5182                         &BTRFS_I(inode)->runtime_flags);
5183                 nr_extents--;
5184         }
5185         BTRFS_I(inode)->reserved_extents += nr_extents;
5186         spin_unlock(&BTRFS_I(inode)->lock);
5187
5188         if (delalloc_lock)
5189                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5190
5191         if (to_reserve)
5192                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5193                                               btrfs_ino(inode), to_reserve, 1);
5194         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5195
5196         return 0;
5197
5198 out_fail:
5199         spin_lock(&BTRFS_I(inode)->lock);
5200         dropped = drop_outstanding_extent(inode, num_bytes);
5201         /*
5202          * If the inodes csum_bytes is the same as the original
5203          * csum_bytes then we know we haven't raced with any free()ers
5204          * so we can just reduce our inodes csum bytes and carry on.
5205          */
5206         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5207                 calc_csum_metadata_size(inode, num_bytes, 0);
5208         } else {
5209                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5210                 u64 bytes;
5211
5212                 /*
5213                  * This is tricky, but first we need to figure out how much we
5214                  * free'd from any free-ers that occured during this
5215                  * reservation, so we reset ->csum_bytes to the csum_bytes
5216                  * before we dropped our lock, and then call the free for the
5217                  * number of bytes that were freed while we were trying our
5218                  * reservation.
5219                  */
5220                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5221                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5222                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5223
5224
5225                 /*
5226                  * Now we need to see how much we would have freed had we not
5227                  * been making this reservation and our ->csum_bytes were not
5228                  * artificially inflated.
5229                  */
5230                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5231                 bytes = csum_bytes - orig_csum_bytes;
5232                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5233
5234                 /*
5235                  * Now reset ->csum_bytes to what it should be.  If bytes is
5236                  * more than to_free then we would have free'd more space had we
5237                  * not had an artificially high ->csum_bytes, so we need to free
5238                  * the remainder.  If bytes is the same or less then we don't
5239                  * need to do anything, the other free-ers did the correct
5240                  * thing.
5241                  */
5242                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5243                 if (bytes > to_free)
5244                         to_free = bytes - to_free;
5245                 else
5246                         to_free = 0;
5247         }
5248         spin_unlock(&BTRFS_I(inode)->lock);
5249         if (dropped)
5250                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5251
5252         if (to_free) {
5253                 btrfs_block_rsv_release(root, block_rsv, to_free);
5254                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5255                                               btrfs_ino(inode), to_free, 0);
5256         }
5257         if (delalloc_lock)
5258                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5259         return ret;
5260 }
5261
5262 /**
5263  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5264  * @inode: the inode to release the reservation for
5265  * @num_bytes: the number of bytes we're releasing
5266  *
5267  * This will release the metadata reservation for an inode.  This can be called
5268  * once we complete IO for a given set of bytes to release their metadata
5269  * reservations.
5270  */
5271 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5272 {
5273         struct btrfs_root *root = BTRFS_I(inode)->root;
5274         u64 to_free = 0;
5275         unsigned dropped;
5276
5277         num_bytes = ALIGN(num_bytes, root->sectorsize);
5278         spin_lock(&BTRFS_I(inode)->lock);
5279         dropped = drop_outstanding_extent(inode, num_bytes);
5280
5281         if (num_bytes)
5282                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5283         spin_unlock(&BTRFS_I(inode)->lock);
5284         if (dropped > 0)
5285                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5286
5287         if (btrfs_test_is_dummy_root(root))
5288                 return;
5289
5290         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5291                                       btrfs_ino(inode), to_free, 0);
5292         if (root->fs_info->quota_enabled) {
5293                 btrfs_qgroup_free(root, num_bytes +
5294                                         dropped * root->nodesize);
5295         }
5296
5297         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5298                                 to_free);
5299 }
5300
5301 /**
5302  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5303  * @inode: inode we're writing to
5304  * @num_bytes: the number of bytes we want to allocate
5305  *
5306  * This will do the following things
5307  *
5308  * o reserve space in the data space info for num_bytes
5309  * o reserve space in the metadata space info based on number of outstanding
5310  *   extents and how much csums will be needed
5311  * o add to the inodes ->delalloc_bytes
5312  * o add it to the fs_info's delalloc inodes list.
5313  *
5314  * This will return 0 for success and -ENOSPC if there is no space left.
5315  */
5316 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5317 {
5318         int ret;
5319
5320         ret = btrfs_check_data_free_space(inode, num_bytes);
5321         if (ret)
5322                 return ret;
5323
5324         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5325         if (ret) {
5326                 btrfs_free_reserved_data_space(inode, num_bytes);
5327                 return ret;
5328         }
5329
5330         return 0;
5331 }
5332
5333 /**
5334  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5335  * @inode: inode we're releasing space for
5336  * @num_bytes: the number of bytes we want to free up
5337  *
5338  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5339  * called in the case that we don't need the metadata AND data reservations
5340  * anymore.  So if there is an error or we insert an inline extent.
5341  *
5342  * This function will release the metadata space that was not used and will
5343  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5344  * list if there are no delalloc bytes left.
5345  */
5346 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5347 {
5348         btrfs_delalloc_release_metadata(inode, num_bytes);
5349         btrfs_free_reserved_data_space(inode, num_bytes);
5350 }
5351
5352 static int update_block_group(struct btrfs_trans_handle *trans,
5353                               struct btrfs_root *root, u64 bytenr,
5354                               u64 num_bytes, int alloc)
5355 {
5356         struct btrfs_block_group_cache *cache = NULL;
5357         struct btrfs_fs_info *info = root->fs_info;
5358         u64 total = num_bytes;
5359         u64 old_val;
5360         u64 byte_in_group;
5361         int factor;
5362
5363         /* block accounting for super block */
5364         spin_lock(&info->delalloc_root_lock);
5365         old_val = btrfs_super_bytes_used(info->super_copy);
5366         if (alloc)
5367                 old_val += num_bytes;
5368         else
5369                 old_val -= num_bytes;
5370         btrfs_set_super_bytes_used(info->super_copy, old_val);
5371         spin_unlock(&info->delalloc_root_lock);
5372
5373         while (total) {
5374                 cache = btrfs_lookup_block_group(info, bytenr);
5375                 if (!cache)
5376                         return -ENOENT;
5377                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5378                                     BTRFS_BLOCK_GROUP_RAID1 |
5379                                     BTRFS_BLOCK_GROUP_RAID10))
5380                         factor = 2;
5381                 else
5382                         factor = 1;
5383                 /*
5384                  * If this block group has free space cache written out, we
5385                  * need to make sure to load it if we are removing space.  This
5386                  * is because we need the unpinning stage to actually add the
5387                  * space back to the block group, otherwise we will leak space.
5388                  */
5389                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5390                         cache_block_group(cache, 1);
5391
5392                 spin_lock(&trans->transaction->dirty_bgs_lock);
5393                 if (list_empty(&cache->dirty_list)) {
5394                         list_add_tail(&cache->dirty_list,
5395                                       &trans->transaction->dirty_bgs);
5396                         btrfs_get_block_group(cache);
5397                 }
5398                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5399
5400                 byte_in_group = bytenr - cache->key.objectid;
5401                 WARN_ON(byte_in_group > cache->key.offset);
5402
5403                 spin_lock(&cache->space_info->lock);
5404                 spin_lock(&cache->lock);
5405
5406                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5407                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5408                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5409
5410                 old_val = btrfs_block_group_used(&cache->item);
5411                 num_bytes = min(total, cache->key.offset - byte_in_group);
5412                 if (alloc) {
5413                         old_val += num_bytes;
5414                         btrfs_set_block_group_used(&cache->item, old_val);
5415                         cache->reserved -= num_bytes;
5416                         cache->space_info->bytes_reserved -= num_bytes;
5417                         cache->space_info->bytes_used += num_bytes;
5418                         cache->space_info->disk_used += num_bytes * factor;
5419                         spin_unlock(&cache->lock);
5420                         spin_unlock(&cache->space_info->lock);
5421                 } else {
5422                         old_val -= num_bytes;
5423                         btrfs_set_block_group_used(&cache->item, old_val);
5424                         cache->pinned += num_bytes;
5425                         cache->space_info->bytes_pinned += num_bytes;
5426                         cache->space_info->bytes_used -= num_bytes;
5427                         cache->space_info->disk_used -= num_bytes * factor;
5428                         spin_unlock(&cache->lock);
5429                         spin_unlock(&cache->space_info->lock);
5430
5431                         set_extent_dirty(info->pinned_extents,
5432                                          bytenr, bytenr + num_bytes - 1,
5433                                          GFP_NOFS | __GFP_NOFAIL);
5434                         /*
5435                          * No longer have used bytes in this block group, queue
5436                          * it for deletion.
5437                          */
5438                         if (old_val == 0) {
5439                                 spin_lock(&info->unused_bgs_lock);
5440                                 if (list_empty(&cache->bg_list)) {
5441                                         btrfs_get_block_group(cache);
5442                                         list_add_tail(&cache->bg_list,
5443                                                       &info->unused_bgs);
5444                                 }
5445                                 spin_unlock(&info->unused_bgs_lock);
5446                         }
5447                 }
5448                 btrfs_put_block_group(cache);
5449                 total -= num_bytes;
5450                 bytenr += num_bytes;
5451         }
5452         return 0;
5453 }
5454
5455 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5456 {
5457         struct btrfs_block_group_cache *cache;
5458         u64 bytenr;
5459
5460         spin_lock(&root->fs_info->block_group_cache_lock);
5461         bytenr = root->fs_info->first_logical_byte;
5462         spin_unlock(&root->fs_info->block_group_cache_lock);
5463
5464         if (bytenr < (u64)-1)
5465                 return bytenr;
5466
5467         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5468         if (!cache)
5469                 return 0;
5470
5471         bytenr = cache->key.objectid;
5472         btrfs_put_block_group(cache);
5473
5474         return bytenr;
5475 }
5476
5477 static int pin_down_extent(struct btrfs_root *root,
5478                            struct btrfs_block_group_cache *cache,
5479                            u64 bytenr, u64 num_bytes, int reserved)
5480 {
5481         spin_lock(&cache->space_info->lock);
5482         spin_lock(&cache->lock);
5483         cache->pinned += num_bytes;
5484         cache->space_info->bytes_pinned += num_bytes;
5485         if (reserved) {
5486                 cache->reserved -= num_bytes;
5487                 cache->space_info->bytes_reserved -= num_bytes;
5488         }
5489         spin_unlock(&cache->lock);
5490         spin_unlock(&cache->space_info->lock);
5491
5492         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5493                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5494         if (reserved)
5495                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5496         return 0;
5497 }
5498
5499 /*
5500  * this function must be called within transaction
5501  */
5502 int btrfs_pin_extent(struct btrfs_root *root,
5503                      u64 bytenr, u64 num_bytes, int reserved)
5504 {
5505         struct btrfs_block_group_cache *cache;
5506
5507         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5508         BUG_ON(!cache); /* Logic error */
5509
5510         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5511
5512         btrfs_put_block_group(cache);
5513         return 0;
5514 }
5515
5516 /*
5517  * this function must be called within transaction
5518  */
5519 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5520                                     u64 bytenr, u64 num_bytes)
5521 {
5522         struct btrfs_block_group_cache *cache;
5523         int ret;
5524
5525         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5526         if (!cache)
5527                 return -EINVAL;
5528
5529         /*
5530          * pull in the free space cache (if any) so that our pin
5531          * removes the free space from the cache.  We have load_only set
5532          * to one because the slow code to read in the free extents does check
5533          * the pinned extents.
5534          */
5535         cache_block_group(cache, 1);
5536
5537         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5538
5539         /* remove us from the free space cache (if we're there at all) */
5540         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5541         btrfs_put_block_group(cache);
5542         return ret;
5543 }
5544
5545 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5546 {
5547         int ret;
5548         struct btrfs_block_group_cache *block_group;
5549         struct btrfs_caching_control *caching_ctl;
5550
5551         block_group = btrfs_lookup_block_group(root->fs_info, start);
5552         if (!block_group)
5553                 return -EINVAL;
5554
5555         cache_block_group(block_group, 0);
5556         caching_ctl = get_caching_control(block_group);
5557
5558         if (!caching_ctl) {
5559                 /* Logic error */
5560                 BUG_ON(!block_group_cache_done(block_group));
5561                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5562         } else {
5563                 mutex_lock(&caching_ctl->mutex);
5564
5565                 if (start >= caching_ctl->progress) {
5566                         ret = add_excluded_extent(root, start, num_bytes);
5567                 } else if (start + num_bytes <= caching_ctl->progress) {
5568                         ret = btrfs_remove_free_space(block_group,
5569                                                       start, num_bytes);
5570                 } else {
5571                         num_bytes = caching_ctl->progress - start;
5572                         ret = btrfs_remove_free_space(block_group,
5573                                                       start, num_bytes);
5574                         if (ret)
5575                                 goto out_lock;
5576
5577                         num_bytes = (start + num_bytes) -
5578                                 caching_ctl->progress;
5579                         start = caching_ctl->progress;
5580                         ret = add_excluded_extent(root, start, num_bytes);
5581                 }
5582 out_lock:
5583                 mutex_unlock(&caching_ctl->mutex);
5584                 put_caching_control(caching_ctl);
5585         }
5586         btrfs_put_block_group(block_group);
5587         return ret;
5588 }
5589
5590 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5591                                  struct extent_buffer *eb)
5592 {
5593         struct btrfs_file_extent_item *item;
5594         struct btrfs_key key;
5595         int found_type;
5596         int i;
5597
5598         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5599                 return 0;
5600
5601         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5602                 btrfs_item_key_to_cpu(eb, &key, i);
5603                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5604                         continue;
5605                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5606                 found_type = btrfs_file_extent_type(eb, item);
5607                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5608                         continue;
5609                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5610                         continue;
5611                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5612                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5613                 __exclude_logged_extent(log, key.objectid, key.offset);
5614         }
5615
5616         return 0;
5617 }
5618
5619 /**
5620  * btrfs_update_reserved_bytes - update the block_group and space info counters
5621  * @cache:      The cache we are manipulating
5622  * @num_bytes:  The number of bytes in question
5623  * @reserve:    One of the reservation enums
5624  * @delalloc:   The blocks are allocated for the delalloc write
5625  *
5626  * This is called by the allocator when it reserves space, or by somebody who is
5627  * freeing space that was never actually used on disk.  For example if you
5628  * reserve some space for a new leaf in transaction A and before transaction A
5629  * commits you free that leaf, you call this with reserve set to 0 in order to
5630  * clear the reservation.
5631  *
5632  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5633  * ENOSPC accounting.  For data we handle the reservation through clearing the
5634  * delalloc bits in the io_tree.  We have to do this since we could end up
5635  * allocating less disk space for the amount of data we have reserved in the
5636  * case of compression.
5637  *
5638  * If this is a reservation and the block group has become read only we cannot
5639  * make the reservation and return -EAGAIN, otherwise this function always
5640  * succeeds.
5641  */
5642 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5643                                        u64 num_bytes, int reserve, int delalloc)
5644 {
5645         struct btrfs_space_info *space_info = cache->space_info;
5646         int ret = 0;
5647
5648         spin_lock(&space_info->lock);
5649         spin_lock(&cache->lock);
5650         if (reserve != RESERVE_FREE) {
5651                 if (cache->ro) {
5652                         ret = -EAGAIN;
5653                 } else {
5654                         cache->reserved += num_bytes;
5655                         space_info->bytes_reserved += num_bytes;
5656                         if (reserve == RESERVE_ALLOC) {
5657                                 trace_btrfs_space_reservation(cache->fs_info,
5658                                                 "space_info", space_info->flags,
5659                                                 num_bytes, 0);
5660                                 space_info->bytes_may_use -= num_bytes;
5661                         }
5662
5663                         if (delalloc)
5664                                 cache->delalloc_bytes += num_bytes;
5665                 }
5666         } else {
5667                 if (cache->ro)
5668                         space_info->bytes_readonly += num_bytes;
5669                 cache->reserved -= num_bytes;
5670                 space_info->bytes_reserved -= num_bytes;
5671
5672                 if (delalloc)
5673                         cache->delalloc_bytes -= num_bytes;
5674         }
5675         spin_unlock(&cache->lock);
5676         spin_unlock(&space_info->lock);
5677         return ret;
5678 }
5679
5680 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5681                                 struct btrfs_root *root)
5682 {
5683         struct btrfs_fs_info *fs_info = root->fs_info;
5684         struct btrfs_caching_control *next;
5685         struct btrfs_caching_control *caching_ctl;
5686         struct btrfs_block_group_cache *cache;
5687
5688         down_write(&fs_info->commit_root_sem);
5689
5690         list_for_each_entry_safe(caching_ctl, next,
5691                                  &fs_info->caching_block_groups, list) {
5692                 cache = caching_ctl->block_group;
5693                 if (block_group_cache_done(cache)) {
5694                         cache->last_byte_to_unpin = (u64)-1;
5695                         list_del_init(&caching_ctl->list);
5696                         put_caching_control(caching_ctl);
5697                 } else {
5698                         cache->last_byte_to_unpin = caching_ctl->progress;
5699                 }
5700         }
5701
5702         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5703                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5704         else
5705                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5706
5707         up_write(&fs_info->commit_root_sem);
5708
5709         update_global_block_rsv(fs_info);
5710 }
5711
5712 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
5713                               const bool return_free_space)
5714 {
5715         struct btrfs_fs_info *fs_info = root->fs_info;
5716         struct btrfs_block_group_cache *cache = NULL;
5717         struct btrfs_space_info *space_info;
5718         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5719         u64 len;
5720         bool readonly;
5721
5722         while (start <= end) {
5723                 readonly = false;
5724                 if (!cache ||
5725                     start >= cache->key.objectid + cache->key.offset) {
5726                         if (cache)
5727                                 btrfs_put_block_group(cache);
5728                         cache = btrfs_lookup_block_group(fs_info, start);
5729                         BUG_ON(!cache); /* Logic error */
5730                 }
5731
5732                 len = cache->key.objectid + cache->key.offset - start;
5733                 len = min(len, end + 1 - start);
5734
5735                 if (start < cache->last_byte_to_unpin) {
5736                         len = min(len, cache->last_byte_to_unpin - start);
5737                         if (return_free_space)
5738                                 btrfs_add_free_space(cache, start, len);
5739                 }
5740
5741                 start += len;
5742                 space_info = cache->space_info;
5743
5744                 spin_lock(&space_info->lock);
5745                 spin_lock(&cache->lock);
5746                 cache->pinned -= len;
5747                 space_info->bytes_pinned -= len;
5748                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
5749                 if (cache->ro) {
5750                         space_info->bytes_readonly += len;
5751                         readonly = true;
5752                 }
5753                 spin_unlock(&cache->lock);
5754                 if (!readonly && global_rsv->space_info == space_info) {
5755                         spin_lock(&global_rsv->lock);
5756                         if (!global_rsv->full) {
5757                                 len = min(len, global_rsv->size -
5758                                           global_rsv->reserved);
5759                                 global_rsv->reserved += len;
5760                                 space_info->bytes_may_use += len;
5761                                 if (global_rsv->reserved >= global_rsv->size)
5762                                         global_rsv->full = 1;
5763                         }
5764                         spin_unlock(&global_rsv->lock);
5765                 }
5766                 spin_unlock(&space_info->lock);
5767         }
5768
5769         if (cache)
5770                 btrfs_put_block_group(cache);
5771         return 0;
5772 }
5773
5774 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5775                                struct btrfs_root *root)
5776 {
5777         struct btrfs_fs_info *fs_info = root->fs_info;
5778         struct extent_io_tree *unpin;
5779         u64 start;
5780         u64 end;
5781         int ret;
5782
5783         if (trans->aborted)
5784                 return 0;
5785
5786         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5787                 unpin = &fs_info->freed_extents[1];
5788         else
5789                 unpin = &fs_info->freed_extents[0];
5790
5791         while (1) {
5792                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
5793                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5794                                             EXTENT_DIRTY, NULL);
5795                 if (ret) {
5796                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
5797                         break;
5798                 }
5799
5800                 if (btrfs_test_opt(root, DISCARD))
5801                         ret = btrfs_discard_extent(root, start,
5802                                                    end + 1 - start, NULL);
5803
5804                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5805                 unpin_extent_range(root, start, end, true);
5806                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
5807                 cond_resched();
5808         }
5809
5810         return 0;
5811 }
5812
5813 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
5814                              u64 owner, u64 root_objectid)
5815 {
5816         struct btrfs_space_info *space_info;
5817         u64 flags;
5818
5819         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5820                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
5821                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
5822                 else
5823                         flags = BTRFS_BLOCK_GROUP_METADATA;
5824         } else {
5825                 flags = BTRFS_BLOCK_GROUP_DATA;
5826         }
5827
5828         space_info = __find_space_info(fs_info, flags);
5829         BUG_ON(!space_info); /* Logic bug */
5830         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
5831 }
5832
5833
5834 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5835                                 struct btrfs_root *root,
5836                                 u64 bytenr, u64 num_bytes, u64 parent,
5837                                 u64 root_objectid, u64 owner_objectid,
5838                                 u64 owner_offset, int refs_to_drop,
5839                                 struct btrfs_delayed_extent_op *extent_op,
5840                                 int no_quota)
5841 {
5842         struct btrfs_key key;
5843         struct btrfs_path *path;
5844         struct btrfs_fs_info *info = root->fs_info;
5845         struct btrfs_root *extent_root = info->extent_root;
5846         struct extent_buffer *leaf;
5847         struct btrfs_extent_item *ei;
5848         struct btrfs_extent_inline_ref *iref;
5849         int ret;
5850         int is_data;
5851         int extent_slot = 0;
5852         int found_extent = 0;
5853         int num_to_del = 1;
5854         u32 item_size;
5855         u64 refs;
5856         int last_ref = 0;
5857         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_SUB_EXCL;
5858         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5859                                                  SKINNY_METADATA);
5860
5861         if (!info->quota_enabled || !is_fstree(root_objectid))
5862                 no_quota = 1;
5863
5864         path = btrfs_alloc_path();
5865         if (!path)
5866                 return -ENOMEM;
5867
5868         path->reada = 1;
5869         path->leave_spinning = 1;
5870
5871         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5872         BUG_ON(!is_data && refs_to_drop != 1);
5873
5874         if (is_data)
5875                 skinny_metadata = 0;
5876
5877         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5878                                     bytenr, num_bytes, parent,
5879                                     root_objectid, owner_objectid,
5880                                     owner_offset);
5881         if (ret == 0) {
5882                 extent_slot = path->slots[0];
5883                 while (extent_slot >= 0) {
5884                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5885                                               extent_slot);
5886                         if (key.objectid != bytenr)
5887                                 break;
5888                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5889                             key.offset == num_bytes) {
5890                                 found_extent = 1;
5891                                 break;
5892                         }
5893                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5894                             key.offset == owner_objectid) {
5895                                 found_extent = 1;
5896                                 break;
5897                         }
5898                         if (path->slots[0] - extent_slot > 5)
5899                                 break;
5900                         extent_slot--;
5901                 }
5902 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5903                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5904                 if (found_extent && item_size < sizeof(*ei))
5905                         found_extent = 0;
5906 #endif
5907                 if (!found_extent) {
5908                         BUG_ON(iref);
5909                         ret = remove_extent_backref(trans, extent_root, path,
5910                                                     NULL, refs_to_drop,
5911                                                     is_data, &last_ref);
5912                         if (ret) {
5913                                 btrfs_abort_transaction(trans, extent_root, ret);
5914                                 goto out;
5915                         }
5916                         btrfs_release_path(path);
5917                         path->leave_spinning = 1;
5918
5919                         key.objectid = bytenr;
5920                         key.type = BTRFS_EXTENT_ITEM_KEY;
5921                         key.offset = num_bytes;
5922
5923                         if (!is_data && skinny_metadata) {
5924                                 key.type = BTRFS_METADATA_ITEM_KEY;
5925                                 key.offset = owner_objectid;
5926                         }
5927
5928                         ret = btrfs_search_slot(trans, extent_root,
5929                                                 &key, path, -1, 1);
5930                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5931                                 /*
5932                                  * Couldn't find our skinny metadata item,
5933                                  * see if we have ye olde extent item.
5934                                  */
5935                                 path->slots[0]--;
5936                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5937                                                       path->slots[0]);
5938                                 if (key.objectid == bytenr &&
5939                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5940                                     key.offset == num_bytes)
5941                                         ret = 0;
5942                         }
5943
5944                         if (ret > 0 && skinny_metadata) {
5945                                 skinny_metadata = false;
5946                                 key.objectid = bytenr;
5947                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5948                                 key.offset = num_bytes;
5949                                 btrfs_release_path(path);
5950                                 ret = btrfs_search_slot(trans, extent_root,
5951                                                         &key, path, -1, 1);
5952                         }
5953
5954                         if (ret) {
5955                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5956                                         ret, bytenr);
5957                                 if (ret > 0)
5958                                         btrfs_print_leaf(extent_root,
5959                                                          path->nodes[0]);
5960                         }
5961                         if (ret < 0) {
5962                                 btrfs_abort_transaction(trans, extent_root, ret);
5963                                 goto out;
5964                         }
5965                         extent_slot = path->slots[0];
5966                 }
5967         } else if (WARN_ON(ret == -ENOENT)) {
5968                 btrfs_print_leaf(extent_root, path->nodes[0]);
5969                 btrfs_err(info,
5970                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5971                         bytenr, parent, root_objectid, owner_objectid,
5972                         owner_offset);
5973                 btrfs_abort_transaction(trans, extent_root, ret);
5974                 goto out;
5975         } else {
5976                 btrfs_abort_transaction(trans, extent_root, ret);
5977                 goto out;
5978         }
5979
5980         leaf = path->nodes[0];
5981         item_size = btrfs_item_size_nr(leaf, extent_slot);
5982 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5983         if (item_size < sizeof(*ei)) {
5984                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5985                 ret = convert_extent_item_v0(trans, extent_root, path,
5986                                              owner_objectid, 0);
5987                 if (ret < 0) {
5988                         btrfs_abort_transaction(trans, extent_root, ret);
5989                         goto out;
5990                 }
5991
5992                 btrfs_release_path(path);
5993                 path->leave_spinning = 1;
5994
5995                 key.objectid = bytenr;
5996                 key.type = BTRFS_EXTENT_ITEM_KEY;
5997                 key.offset = num_bytes;
5998
5999                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6000                                         -1, 1);
6001                 if (ret) {
6002                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6003                                 ret, bytenr);
6004                         btrfs_print_leaf(extent_root, path->nodes[0]);
6005                 }
6006                 if (ret < 0) {
6007                         btrfs_abort_transaction(trans, extent_root, ret);
6008                         goto out;
6009                 }
6010
6011                 extent_slot = path->slots[0];
6012                 leaf = path->nodes[0];
6013                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6014         }
6015 #endif
6016         BUG_ON(item_size < sizeof(*ei));
6017         ei = btrfs_item_ptr(leaf, extent_slot,
6018                             struct btrfs_extent_item);
6019         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6020             key.type == BTRFS_EXTENT_ITEM_KEY) {
6021                 struct btrfs_tree_block_info *bi;
6022                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6023                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6024                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6025         }
6026
6027         refs = btrfs_extent_refs(leaf, ei);
6028         if (refs < refs_to_drop) {
6029                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6030                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6031                 ret = -EINVAL;
6032                 btrfs_abort_transaction(trans, extent_root, ret);
6033                 goto out;
6034         }
6035         refs -= refs_to_drop;
6036
6037         if (refs > 0) {
6038                 type = BTRFS_QGROUP_OPER_SUB_SHARED;
6039                 if (extent_op)
6040                         __run_delayed_extent_op(extent_op, leaf, ei);
6041                 /*
6042                  * In the case of inline back ref, reference count will
6043                  * be updated by remove_extent_backref
6044                  */
6045                 if (iref) {
6046                         BUG_ON(!found_extent);
6047                 } else {
6048                         btrfs_set_extent_refs(leaf, ei, refs);
6049                         btrfs_mark_buffer_dirty(leaf);
6050                 }
6051                 if (found_extent) {
6052                         ret = remove_extent_backref(trans, extent_root, path,
6053                                                     iref, refs_to_drop,
6054                                                     is_data, &last_ref);
6055                         if (ret) {
6056                                 btrfs_abort_transaction(trans, extent_root, ret);
6057                                 goto out;
6058                         }
6059                 }
6060                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6061                                  root_objectid);
6062         } else {
6063                 if (found_extent) {
6064                         BUG_ON(is_data && refs_to_drop !=
6065                                extent_data_ref_count(root, path, iref));
6066                         if (iref) {
6067                                 BUG_ON(path->slots[0] != extent_slot);
6068                         } else {
6069                                 BUG_ON(path->slots[0] != extent_slot + 1);
6070                                 path->slots[0] = extent_slot;
6071                                 num_to_del = 2;
6072                         }
6073                 }
6074
6075                 last_ref = 1;
6076                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6077                                       num_to_del);
6078                 if (ret) {
6079                         btrfs_abort_transaction(trans, extent_root, ret);
6080                         goto out;
6081                 }
6082                 btrfs_release_path(path);
6083
6084                 if (is_data) {
6085                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6086                         if (ret) {
6087                                 btrfs_abort_transaction(trans, extent_root, ret);
6088                                 goto out;
6089                         }
6090                 }
6091
6092                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6093                 if (ret) {
6094                         btrfs_abort_transaction(trans, extent_root, ret);
6095                         goto out;
6096                 }
6097         }
6098         btrfs_release_path(path);
6099
6100         /* Deal with the quota accounting */
6101         if (!ret && last_ref && !no_quota) {
6102                 int mod_seq = 0;
6103
6104                 if (owner_objectid >= BTRFS_FIRST_FREE_OBJECTID &&
6105                     type == BTRFS_QGROUP_OPER_SUB_SHARED)
6106                         mod_seq = 1;
6107
6108                 ret = btrfs_qgroup_record_ref(trans, info, root_objectid,
6109                                               bytenr, num_bytes, type,
6110                                               mod_seq);
6111         }
6112 out:
6113         btrfs_free_path(path);
6114         return ret;
6115 }
6116
6117 /*
6118  * when we free an block, it is possible (and likely) that we free the last
6119  * delayed ref for that extent as well.  This searches the delayed ref tree for
6120  * a given extent, and if there are no other delayed refs to be processed, it
6121  * removes it from the tree.
6122  */
6123 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6124                                       struct btrfs_root *root, u64 bytenr)
6125 {
6126         struct btrfs_delayed_ref_head *head;
6127         struct btrfs_delayed_ref_root *delayed_refs;
6128         int ret = 0;
6129
6130         delayed_refs = &trans->transaction->delayed_refs;
6131         spin_lock(&delayed_refs->lock);
6132         head = btrfs_find_delayed_ref_head(trans, bytenr);
6133         if (!head)
6134                 goto out_delayed_unlock;
6135
6136         spin_lock(&head->lock);
6137         if (rb_first(&head->ref_root))
6138                 goto out;
6139
6140         if (head->extent_op) {
6141                 if (!head->must_insert_reserved)
6142                         goto out;
6143                 btrfs_free_delayed_extent_op(head->extent_op);
6144                 head->extent_op = NULL;
6145         }
6146
6147         /*
6148          * waiting for the lock here would deadlock.  If someone else has it
6149          * locked they are already in the process of dropping it anyway
6150          */
6151         if (!mutex_trylock(&head->mutex))
6152                 goto out;
6153
6154         /*
6155          * at this point we have a head with no other entries.  Go
6156          * ahead and process it.
6157          */
6158         head->node.in_tree = 0;
6159         rb_erase(&head->href_node, &delayed_refs->href_root);
6160
6161         atomic_dec(&delayed_refs->num_entries);
6162
6163         /*
6164          * we don't take a ref on the node because we're removing it from the
6165          * tree, so we just steal the ref the tree was holding.
6166          */
6167         delayed_refs->num_heads--;
6168         if (head->processing == 0)
6169                 delayed_refs->num_heads_ready--;
6170         head->processing = 0;
6171         spin_unlock(&head->lock);
6172         spin_unlock(&delayed_refs->lock);
6173
6174         BUG_ON(head->extent_op);
6175         if (head->must_insert_reserved)
6176                 ret = 1;
6177
6178         mutex_unlock(&head->mutex);
6179         btrfs_put_delayed_ref(&head->node);
6180         return ret;
6181 out:
6182         spin_unlock(&head->lock);
6183
6184 out_delayed_unlock:
6185         spin_unlock(&delayed_refs->lock);
6186         return 0;
6187 }
6188
6189 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6190                            struct btrfs_root *root,
6191                            struct extent_buffer *buf,
6192                            u64 parent, int last_ref)
6193 {
6194         int pin = 1;
6195         int ret;
6196
6197         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6198                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6199                                         buf->start, buf->len,
6200                                         parent, root->root_key.objectid,
6201                                         btrfs_header_level(buf),
6202                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6203                 BUG_ON(ret); /* -ENOMEM */
6204         }
6205
6206         if (!last_ref)
6207                 return;
6208
6209         if (btrfs_header_generation(buf) == trans->transid) {
6210                 struct btrfs_block_group_cache *cache;
6211
6212                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6213                         ret = check_ref_cleanup(trans, root, buf->start);
6214                         if (!ret)
6215                                 goto out;
6216                 }
6217
6218                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6219
6220                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6221                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6222                         btrfs_put_block_group(cache);
6223                         goto out;
6224                 }
6225
6226                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6227
6228                 btrfs_add_free_space(cache, buf->start, buf->len);
6229                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6230                 btrfs_put_block_group(cache);
6231                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6232                 pin = 0;
6233         }
6234 out:
6235         if (pin)
6236                 add_pinned_bytes(root->fs_info, buf->len,
6237                                  btrfs_header_level(buf),
6238                                  root->root_key.objectid);
6239
6240         /*
6241          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6242          * anymore.
6243          */
6244         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6245 }
6246
6247 /* Can return -ENOMEM */
6248 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6249                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6250                       u64 owner, u64 offset, int no_quota)
6251 {
6252         int ret;
6253         struct btrfs_fs_info *fs_info = root->fs_info;
6254
6255         if (btrfs_test_is_dummy_root(root))
6256                 return 0;
6257
6258         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6259
6260         /*
6261          * tree log blocks never actually go into the extent allocation
6262          * tree, just update pinning info and exit early.
6263          */
6264         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6265                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6266                 /* unlocks the pinned mutex */
6267                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6268                 ret = 0;
6269         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6270                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6271                                         num_bytes,
6272                                         parent, root_objectid, (int)owner,
6273                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6274         } else {
6275                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6276                                                 num_bytes,
6277                                                 parent, root_objectid, owner,
6278                                                 offset, BTRFS_DROP_DELAYED_REF,
6279                                                 NULL, no_quota);
6280         }
6281         return ret;
6282 }
6283
6284 /*
6285  * when we wait for progress in the block group caching, its because
6286  * our allocation attempt failed at least once.  So, we must sleep
6287  * and let some progress happen before we try again.
6288  *
6289  * This function will sleep at least once waiting for new free space to
6290  * show up, and then it will check the block group free space numbers
6291  * for our min num_bytes.  Another option is to have it go ahead
6292  * and look in the rbtree for a free extent of a given size, but this
6293  * is a good start.
6294  *
6295  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6296  * any of the information in this block group.
6297  */
6298 static noinline void
6299 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6300                                 u64 num_bytes)
6301 {
6302         struct btrfs_caching_control *caching_ctl;
6303
6304         caching_ctl = get_caching_control(cache);
6305         if (!caching_ctl)
6306                 return;
6307
6308         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6309                    (cache->free_space_ctl->free_space >= num_bytes));
6310
6311         put_caching_control(caching_ctl);
6312 }
6313
6314 static noinline int
6315 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6316 {
6317         struct btrfs_caching_control *caching_ctl;
6318         int ret = 0;
6319
6320         caching_ctl = get_caching_control(cache);
6321         if (!caching_ctl)
6322                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6323
6324         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6325         if (cache->cached == BTRFS_CACHE_ERROR)
6326                 ret = -EIO;
6327         put_caching_control(caching_ctl);
6328         return ret;
6329 }
6330
6331 int __get_raid_index(u64 flags)
6332 {
6333         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6334                 return BTRFS_RAID_RAID10;
6335         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6336                 return BTRFS_RAID_RAID1;
6337         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6338                 return BTRFS_RAID_DUP;
6339         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6340                 return BTRFS_RAID_RAID0;
6341         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6342                 return BTRFS_RAID_RAID5;
6343         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6344                 return BTRFS_RAID_RAID6;
6345
6346         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6347 }
6348
6349 int get_block_group_index(struct btrfs_block_group_cache *cache)
6350 {
6351         return __get_raid_index(cache->flags);
6352 }
6353
6354 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6355         [BTRFS_RAID_RAID10]     = "raid10",
6356         [BTRFS_RAID_RAID1]      = "raid1",
6357         [BTRFS_RAID_DUP]        = "dup",
6358         [BTRFS_RAID_RAID0]      = "raid0",
6359         [BTRFS_RAID_SINGLE]     = "single",
6360         [BTRFS_RAID_RAID5]      = "raid5",
6361         [BTRFS_RAID_RAID6]      = "raid6",
6362 };
6363
6364 static const char *get_raid_name(enum btrfs_raid_types type)
6365 {
6366         if (type >= BTRFS_NR_RAID_TYPES)
6367                 return NULL;
6368
6369         return btrfs_raid_type_names[type];
6370 }
6371
6372 enum btrfs_loop_type {
6373         LOOP_CACHING_NOWAIT = 0,
6374         LOOP_CACHING_WAIT = 1,
6375         LOOP_ALLOC_CHUNK = 2,
6376         LOOP_NO_EMPTY_SIZE = 3,
6377 };
6378
6379 static inline void
6380 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6381                        int delalloc)
6382 {
6383         if (delalloc)
6384                 down_read(&cache->data_rwsem);
6385 }
6386
6387 static inline void
6388 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6389                        int delalloc)
6390 {
6391         btrfs_get_block_group(cache);
6392         if (delalloc)
6393                 down_read(&cache->data_rwsem);
6394 }
6395
6396 static struct btrfs_block_group_cache *
6397 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6398                    struct btrfs_free_cluster *cluster,
6399                    int delalloc)
6400 {
6401         struct btrfs_block_group_cache *used_bg;
6402         bool locked = false;
6403 again:
6404         spin_lock(&cluster->refill_lock);
6405         if (locked) {
6406                 if (used_bg == cluster->block_group)
6407                         return used_bg;
6408
6409                 up_read(&used_bg->data_rwsem);
6410                 btrfs_put_block_group(used_bg);
6411         }
6412
6413         used_bg = cluster->block_group;
6414         if (!used_bg)
6415                 return NULL;
6416
6417         if (used_bg == block_group)
6418                 return used_bg;
6419
6420         btrfs_get_block_group(used_bg);
6421
6422         if (!delalloc)
6423                 return used_bg;
6424
6425         if (down_read_trylock(&used_bg->data_rwsem))
6426                 return used_bg;
6427
6428         spin_unlock(&cluster->refill_lock);
6429         down_read(&used_bg->data_rwsem);
6430         locked = true;
6431         goto again;
6432 }
6433
6434 static inline void
6435 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6436                          int delalloc)
6437 {
6438         if (delalloc)
6439                 up_read(&cache->data_rwsem);
6440         btrfs_put_block_group(cache);
6441 }
6442
6443 /*
6444  * walks the btree of allocated extents and find a hole of a given size.
6445  * The key ins is changed to record the hole:
6446  * ins->objectid == start position
6447  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6448  * ins->offset == the size of the hole.
6449  * Any available blocks before search_start are skipped.
6450  *
6451  * If there is no suitable free space, we will record the max size of
6452  * the free space extent currently.
6453  */
6454 static noinline int find_free_extent(struct btrfs_root *orig_root,
6455                                      u64 num_bytes, u64 empty_size,
6456                                      u64 hint_byte, struct btrfs_key *ins,
6457                                      u64 flags, int delalloc)
6458 {
6459         int ret = 0;
6460         struct btrfs_root *root = orig_root->fs_info->extent_root;
6461         struct btrfs_free_cluster *last_ptr = NULL;
6462         struct btrfs_block_group_cache *block_group = NULL;
6463         u64 search_start = 0;
6464         u64 max_extent_size = 0;
6465         int empty_cluster = 2 * 1024 * 1024;
6466         struct btrfs_space_info *space_info;
6467         int loop = 0;
6468         int index = __get_raid_index(flags);
6469         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6470                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6471         bool failed_cluster_refill = false;
6472         bool failed_alloc = false;
6473         bool use_cluster = true;
6474         bool have_caching_bg = false;
6475
6476         WARN_ON(num_bytes < root->sectorsize);
6477         ins->type = BTRFS_EXTENT_ITEM_KEY;
6478         ins->objectid = 0;
6479         ins->offset = 0;
6480
6481         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6482
6483         space_info = __find_space_info(root->fs_info, flags);
6484         if (!space_info) {
6485                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6486                 return -ENOSPC;
6487         }
6488
6489         /*
6490          * If the space info is for both data and metadata it means we have a
6491          * small filesystem and we can't use the clustering stuff.
6492          */
6493         if (btrfs_mixed_space_info(space_info))
6494                 use_cluster = false;
6495
6496         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6497                 last_ptr = &root->fs_info->meta_alloc_cluster;
6498                 if (!btrfs_test_opt(root, SSD))
6499                         empty_cluster = 64 * 1024;
6500         }
6501
6502         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6503             btrfs_test_opt(root, SSD)) {
6504                 last_ptr = &root->fs_info->data_alloc_cluster;
6505         }
6506
6507         if (last_ptr) {
6508                 spin_lock(&last_ptr->lock);
6509                 if (last_ptr->block_group)
6510                         hint_byte = last_ptr->window_start;
6511                 spin_unlock(&last_ptr->lock);
6512         }
6513
6514         search_start = max(search_start, first_logical_byte(root, 0));
6515         search_start = max(search_start, hint_byte);
6516
6517         if (!last_ptr)
6518                 empty_cluster = 0;
6519
6520         if (search_start == hint_byte) {
6521                 block_group = btrfs_lookup_block_group(root->fs_info,
6522                                                        search_start);
6523                 /*
6524                  * we don't want to use the block group if it doesn't match our
6525                  * allocation bits, or if its not cached.
6526                  *
6527                  * However if we are re-searching with an ideal block group
6528                  * picked out then we don't care that the block group is cached.
6529                  */
6530                 if (block_group && block_group_bits(block_group, flags) &&
6531                     block_group->cached != BTRFS_CACHE_NO) {
6532                         down_read(&space_info->groups_sem);
6533                         if (list_empty(&block_group->list) ||
6534                             block_group->ro) {
6535                                 /*
6536                                  * someone is removing this block group,
6537                                  * we can't jump into the have_block_group
6538                                  * target because our list pointers are not
6539                                  * valid
6540                                  */
6541                                 btrfs_put_block_group(block_group);
6542                                 up_read(&space_info->groups_sem);
6543                         } else {
6544                                 index = get_block_group_index(block_group);
6545                                 btrfs_lock_block_group(block_group, delalloc);
6546                                 goto have_block_group;
6547                         }
6548                 } else if (block_group) {
6549                         btrfs_put_block_group(block_group);
6550                 }
6551         }
6552 search:
6553         have_caching_bg = false;
6554         down_read(&space_info->groups_sem);
6555         list_for_each_entry(block_group, &space_info->block_groups[index],
6556                             list) {
6557                 u64 offset;
6558                 int cached;
6559
6560                 btrfs_grab_block_group(block_group, delalloc);
6561                 search_start = block_group->key.objectid;
6562
6563                 /*
6564                  * this can happen if we end up cycling through all the
6565                  * raid types, but we want to make sure we only allocate
6566                  * for the proper type.
6567                  */
6568                 if (!block_group_bits(block_group, flags)) {
6569                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6570                                 BTRFS_BLOCK_GROUP_RAID1 |
6571                                 BTRFS_BLOCK_GROUP_RAID5 |
6572                                 BTRFS_BLOCK_GROUP_RAID6 |
6573                                 BTRFS_BLOCK_GROUP_RAID10;
6574
6575                         /*
6576                          * if they asked for extra copies and this block group
6577                          * doesn't provide them, bail.  This does allow us to
6578                          * fill raid0 from raid1.
6579                          */
6580                         if ((flags & extra) && !(block_group->flags & extra))
6581                                 goto loop;
6582                 }
6583
6584 have_block_group:
6585                 cached = block_group_cache_done(block_group);
6586                 if (unlikely(!cached)) {
6587                         ret = cache_block_group(block_group, 0);
6588                         BUG_ON(ret < 0);
6589                         ret = 0;
6590                 }
6591
6592                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6593                         goto loop;
6594                 if (unlikely(block_group->ro))
6595                         goto loop;
6596
6597                 /*
6598                  * Ok we want to try and use the cluster allocator, so
6599                  * lets look there
6600                  */
6601                 if (last_ptr) {
6602                         struct btrfs_block_group_cache *used_block_group;
6603                         unsigned long aligned_cluster;
6604                         /*
6605                          * the refill lock keeps out other
6606                          * people trying to start a new cluster
6607                          */
6608                         used_block_group = btrfs_lock_cluster(block_group,
6609                                                               last_ptr,
6610                                                               delalloc);
6611                         if (!used_block_group)
6612                                 goto refill_cluster;
6613
6614                         if (used_block_group != block_group &&
6615                             (used_block_group->ro ||
6616                              !block_group_bits(used_block_group, flags)))
6617                                 goto release_cluster;
6618
6619                         offset = btrfs_alloc_from_cluster(used_block_group,
6620                                                 last_ptr,
6621                                                 num_bytes,
6622                                                 used_block_group->key.objectid,
6623                                                 &max_extent_size);
6624                         if (offset) {
6625                                 /* we have a block, we're done */
6626                                 spin_unlock(&last_ptr->refill_lock);
6627                                 trace_btrfs_reserve_extent_cluster(root,
6628                                                 used_block_group,
6629                                                 search_start, num_bytes);
6630                                 if (used_block_group != block_group) {
6631                                         btrfs_release_block_group(block_group,
6632                                                                   delalloc);
6633                                         block_group = used_block_group;
6634                                 }
6635                                 goto checks;
6636                         }
6637
6638                         WARN_ON(last_ptr->block_group != used_block_group);
6639 release_cluster:
6640                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6641                          * set up a new clusters, so lets just skip it
6642                          * and let the allocator find whatever block
6643                          * it can find.  If we reach this point, we
6644                          * will have tried the cluster allocator
6645                          * plenty of times and not have found
6646                          * anything, so we are likely way too
6647                          * fragmented for the clustering stuff to find
6648                          * anything.
6649                          *
6650                          * However, if the cluster is taken from the
6651                          * current block group, release the cluster
6652                          * first, so that we stand a better chance of
6653                          * succeeding in the unclustered
6654                          * allocation.  */
6655                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6656                             used_block_group != block_group) {
6657                                 spin_unlock(&last_ptr->refill_lock);
6658                                 btrfs_release_block_group(used_block_group,
6659                                                           delalloc);
6660                                 goto unclustered_alloc;
6661                         }
6662
6663                         /*
6664                          * this cluster didn't work out, free it and
6665                          * start over
6666                          */
6667                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6668
6669                         if (used_block_group != block_group)
6670                                 btrfs_release_block_group(used_block_group,
6671                                                           delalloc);
6672 refill_cluster:
6673                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6674                                 spin_unlock(&last_ptr->refill_lock);
6675                                 goto unclustered_alloc;
6676                         }
6677
6678                         aligned_cluster = max_t(unsigned long,
6679                                                 empty_cluster + empty_size,
6680                                               block_group->full_stripe_len);
6681
6682                         /* allocate a cluster in this block group */
6683                         ret = btrfs_find_space_cluster(root, block_group,
6684                                                        last_ptr, search_start,
6685                                                        num_bytes,
6686                                                        aligned_cluster);
6687                         if (ret == 0) {
6688                                 /*
6689                                  * now pull our allocation out of this
6690                                  * cluster
6691                                  */
6692                                 offset = btrfs_alloc_from_cluster(block_group,
6693                                                         last_ptr,
6694                                                         num_bytes,
6695                                                         search_start,
6696                                                         &max_extent_size);
6697                                 if (offset) {
6698                                         /* we found one, proceed */
6699                                         spin_unlock(&last_ptr->refill_lock);
6700                                         trace_btrfs_reserve_extent_cluster(root,
6701                                                 block_group, search_start,
6702                                                 num_bytes);
6703                                         goto checks;
6704                                 }
6705                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6706                                    && !failed_cluster_refill) {
6707                                 spin_unlock(&last_ptr->refill_lock);
6708
6709                                 failed_cluster_refill = true;
6710                                 wait_block_group_cache_progress(block_group,
6711                                        num_bytes + empty_cluster + empty_size);
6712                                 goto have_block_group;
6713                         }
6714
6715                         /*
6716                          * at this point we either didn't find a cluster
6717                          * or we weren't able to allocate a block from our
6718                          * cluster.  Free the cluster we've been trying
6719                          * to use, and go to the next block group
6720                          */
6721                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6722                         spin_unlock(&last_ptr->refill_lock);
6723                         goto loop;
6724                 }
6725
6726 unclustered_alloc:
6727                 spin_lock(&block_group->free_space_ctl->tree_lock);
6728                 if (cached &&
6729                     block_group->free_space_ctl->free_space <
6730                     num_bytes + empty_cluster + empty_size) {
6731                         if (block_group->free_space_ctl->free_space >
6732                             max_extent_size)
6733                                 max_extent_size =
6734                                         block_group->free_space_ctl->free_space;
6735                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6736                         goto loop;
6737                 }
6738                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6739
6740                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6741                                                     num_bytes, empty_size,
6742                                                     &max_extent_size);
6743                 /*
6744                  * If we didn't find a chunk, and we haven't failed on this
6745                  * block group before, and this block group is in the middle of
6746                  * caching and we are ok with waiting, then go ahead and wait
6747                  * for progress to be made, and set failed_alloc to true.
6748                  *
6749                  * If failed_alloc is true then we've already waited on this
6750                  * block group once and should move on to the next block group.
6751                  */
6752                 if (!offset && !failed_alloc && !cached &&
6753                     loop > LOOP_CACHING_NOWAIT) {
6754                         wait_block_group_cache_progress(block_group,
6755                                                 num_bytes + empty_size);
6756                         failed_alloc = true;
6757                         goto have_block_group;
6758                 } else if (!offset) {
6759                         if (!cached)
6760                                 have_caching_bg = true;
6761                         goto loop;
6762                 }
6763 checks:
6764                 search_start = ALIGN(offset, root->stripesize);
6765
6766                 /* move on to the next group */
6767                 if (search_start + num_bytes >
6768                     block_group->key.objectid + block_group->key.offset) {
6769                         btrfs_add_free_space(block_group, offset, num_bytes);
6770                         goto loop;
6771                 }
6772
6773                 if (offset < search_start)
6774                         btrfs_add_free_space(block_group, offset,
6775                                              search_start - offset);
6776                 BUG_ON(offset > search_start);
6777
6778                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
6779                                                   alloc_type, delalloc);
6780                 if (ret == -EAGAIN) {
6781                         btrfs_add_free_space(block_group, offset, num_bytes);
6782                         goto loop;
6783                 }
6784
6785                 /* we are all good, lets return */
6786                 ins->objectid = search_start;
6787                 ins->offset = num_bytes;
6788
6789                 trace_btrfs_reserve_extent(orig_root, block_group,
6790                                            search_start, num_bytes);
6791                 btrfs_release_block_group(block_group, delalloc);
6792                 break;
6793 loop:
6794                 failed_cluster_refill = false;
6795                 failed_alloc = false;
6796                 BUG_ON(index != get_block_group_index(block_group));
6797                 btrfs_release_block_group(block_group, delalloc);
6798         }
6799         up_read(&space_info->groups_sem);
6800
6801         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6802                 goto search;
6803
6804         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6805                 goto search;
6806
6807         /*
6808          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6809          *                      caching kthreads as we move along
6810          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6811          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6812          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6813          *                      again
6814          */
6815         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6816                 index = 0;
6817                 loop++;
6818                 if (loop == LOOP_ALLOC_CHUNK) {
6819                         struct btrfs_trans_handle *trans;
6820                         int exist = 0;
6821
6822                         trans = current->journal_info;
6823                         if (trans)
6824                                 exist = 1;
6825                         else
6826                                 trans = btrfs_join_transaction(root);
6827
6828                         if (IS_ERR(trans)) {
6829                                 ret = PTR_ERR(trans);
6830                                 goto out;
6831                         }
6832
6833                         ret = do_chunk_alloc(trans, root, flags,
6834                                              CHUNK_ALLOC_FORCE);
6835                         /*
6836                          * Do not bail out on ENOSPC since we
6837                          * can do more things.
6838                          */
6839                         if (ret < 0 && ret != -ENOSPC)
6840                                 btrfs_abort_transaction(trans,
6841                                                         root, ret);
6842                         else
6843                                 ret = 0;
6844                         if (!exist)
6845                                 btrfs_end_transaction(trans, root);
6846                         if (ret)
6847                                 goto out;
6848                 }
6849
6850                 if (loop == LOOP_NO_EMPTY_SIZE) {
6851                         empty_size = 0;
6852                         empty_cluster = 0;
6853                 }
6854
6855                 goto search;
6856         } else if (!ins->objectid) {
6857                 ret = -ENOSPC;
6858         } else if (ins->objectid) {
6859                 ret = 0;
6860         }
6861 out:
6862         if (ret == -ENOSPC)
6863                 ins->offset = max_extent_size;
6864         return ret;
6865 }
6866
6867 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6868                             int dump_block_groups)
6869 {
6870         struct btrfs_block_group_cache *cache;
6871         int index = 0;
6872
6873         spin_lock(&info->lock);
6874         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
6875                info->flags,
6876                info->total_bytes - info->bytes_used - info->bytes_pinned -
6877                info->bytes_reserved - info->bytes_readonly,
6878                (info->full) ? "" : "not ");
6879         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
6880                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6881                info->total_bytes, info->bytes_used, info->bytes_pinned,
6882                info->bytes_reserved, info->bytes_may_use,
6883                info->bytes_readonly);
6884         spin_unlock(&info->lock);
6885
6886         if (!dump_block_groups)
6887                 return;
6888
6889         down_read(&info->groups_sem);
6890 again:
6891         list_for_each_entry(cache, &info->block_groups[index], list) {
6892                 spin_lock(&cache->lock);
6893                 printk(KERN_INFO "BTRFS: "
6894                            "block group %llu has %llu bytes, "
6895                            "%llu used %llu pinned %llu reserved %s\n",
6896                        cache->key.objectid, cache->key.offset,
6897                        btrfs_block_group_used(&cache->item), cache->pinned,
6898                        cache->reserved, cache->ro ? "[readonly]" : "");
6899                 btrfs_dump_free_space(cache, bytes);
6900                 spin_unlock(&cache->lock);
6901         }
6902         if (++index < BTRFS_NR_RAID_TYPES)
6903                 goto again;
6904         up_read(&info->groups_sem);
6905 }
6906
6907 int btrfs_reserve_extent(struct btrfs_root *root,
6908                          u64 num_bytes, u64 min_alloc_size,
6909                          u64 empty_size, u64 hint_byte,
6910                          struct btrfs_key *ins, int is_data, int delalloc)
6911 {
6912         bool final_tried = false;
6913         u64 flags;
6914         int ret;
6915
6916         flags = btrfs_get_alloc_profile(root, is_data);
6917 again:
6918         WARN_ON(num_bytes < root->sectorsize);
6919         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
6920                                flags, delalloc);
6921
6922         if (ret == -ENOSPC) {
6923                 if (!final_tried && ins->offset) {
6924                         num_bytes = min(num_bytes >> 1, ins->offset);
6925                         num_bytes = round_down(num_bytes, root->sectorsize);
6926                         num_bytes = max(num_bytes, min_alloc_size);
6927                         if (num_bytes == min_alloc_size)
6928                                 final_tried = true;
6929                         goto again;
6930                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6931                         struct btrfs_space_info *sinfo;
6932
6933                         sinfo = __find_space_info(root->fs_info, flags);
6934                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6935                                 flags, num_bytes);
6936                         if (sinfo)
6937                                 dump_space_info(sinfo, num_bytes, 1);
6938                 }
6939         }
6940
6941         return ret;
6942 }
6943
6944 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6945                                         u64 start, u64 len,
6946                                         int pin, int delalloc)
6947 {
6948         struct btrfs_block_group_cache *cache;
6949         int ret = 0;
6950
6951         cache = btrfs_lookup_block_group(root->fs_info, start);
6952         if (!cache) {
6953                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6954                         start);
6955                 return -ENOSPC;
6956         }
6957
6958         if (btrfs_test_opt(root, DISCARD))
6959                 ret = btrfs_discard_extent(root, start, len, NULL);
6960
6961         if (pin)
6962                 pin_down_extent(root, cache, start, len, 1);
6963         else {
6964                 btrfs_add_free_space(cache, start, len);
6965                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
6966         }
6967         btrfs_put_block_group(cache);
6968
6969         trace_btrfs_reserved_extent_free(root, start, len);
6970
6971         return ret;
6972 }
6973
6974 int btrfs_free_reserved_extent(struct btrfs_root *root,
6975                                u64 start, u64 len, int delalloc)
6976 {
6977         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
6978 }
6979
6980 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6981                                        u64 start, u64 len)
6982 {
6983         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
6984 }
6985
6986 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6987                                       struct btrfs_root *root,
6988                                       u64 parent, u64 root_objectid,
6989                                       u64 flags, u64 owner, u64 offset,
6990                                       struct btrfs_key *ins, int ref_mod)
6991 {
6992         int ret;
6993         struct btrfs_fs_info *fs_info = root->fs_info;
6994         struct btrfs_extent_item *extent_item;
6995         struct btrfs_extent_inline_ref *iref;
6996         struct btrfs_path *path;
6997         struct extent_buffer *leaf;
6998         int type;
6999         u32 size;
7000
7001         if (parent > 0)
7002                 type = BTRFS_SHARED_DATA_REF_KEY;
7003         else
7004                 type = BTRFS_EXTENT_DATA_REF_KEY;
7005
7006         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7007
7008         path = btrfs_alloc_path();
7009         if (!path)
7010                 return -ENOMEM;
7011
7012         path->leave_spinning = 1;
7013         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7014                                       ins, size);
7015         if (ret) {
7016                 btrfs_free_path(path);
7017                 return ret;
7018         }
7019
7020         leaf = path->nodes[0];
7021         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7022                                      struct btrfs_extent_item);
7023         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7024         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7025         btrfs_set_extent_flags(leaf, extent_item,
7026                                flags | BTRFS_EXTENT_FLAG_DATA);
7027
7028         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7029         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7030         if (parent > 0) {
7031                 struct btrfs_shared_data_ref *ref;
7032                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7033                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7034                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7035         } else {
7036                 struct btrfs_extent_data_ref *ref;
7037                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7038                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7039                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7040                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7041                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7042         }
7043
7044         btrfs_mark_buffer_dirty(path->nodes[0]);
7045         btrfs_free_path(path);
7046
7047         /* Always set parent to 0 here since its exclusive anyway. */
7048         ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7049                                       ins->objectid, ins->offset,
7050                                       BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7051         if (ret)
7052                 return ret;
7053
7054         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7055         if (ret) { /* -ENOENT, logic error */
7056                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7057                         ins->objectid, ins->offset);
7058                 BUG();
7059         }
7060         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7061         return ret;
7062 }
7063
7064 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7065                                      struct btrfs_root *root,
7066                                      u64 parent, u64 root_objectid,
7067                                      u64 flags, struct btrfs_disk_key *key,
7068                                      int level, struct btrfs_key *ins,
7069                                      int no_quota)
7070 {
7071         int ret;
7072         struct btrfs_fs_info *fs_info = root->fs_info;
7073         struct btrfs_extent_item *extent_item;
7074         struct btrfs_tree_block_info *block_info;
7075         struct btrfs_extent_inline_ref *iref;
7076         struct btrfs_path *path;
7077         struct extent_buffer *leaf;
7078         u32 size = sizeof(*extent_item) + sizeof(*iref);
7079         u64 num_bytes = ins->offset;
7080         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7081                                                  SKINNY_METADATA);
7082
7083         if (!skinny_metadata)
7084                 size += sizeof(*block_info);
7085
7086         path = btrfs_alloc_path();
7087         if (!path) {
7088                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7089                                                    root->nodesize);
7090                 return -ENOMEM;
7091         }
7092
7093         path->leave_spinning = 1;
7094         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7095                                       ins, size);
7096         if (ret) {
7097                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7098                                                    root->nodesize);
7099                 btrfs_free_path(path);
7100                 return ret;
7101         }
7102
7103         leaf = path->nodes[0];
7104         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7105                                      struct btrfs_extent_item);
7106         btrfs_set_extent_refs(leaf, extent_item, 1);
7107         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7108         btrfs_set_extent_flags(leaf, extent_item,
7109                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7110
7111         if (skinny_metadata) {
7112                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7113                 num_bytes = root->nodesize;
7114         } else {
7115                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7116                 btrfs_set_tree_block_key(leaf, block_info, key);
7117                 btrfs_set_tree_block_level(leaf, block_info, level);
7118                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7119         }
7120
7121         if (parent > 0) {
7122                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7123                 btrfs_set_extent_inline_ref_type(leaf, iref,
7124                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7125                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7126         } else {
7127                 btrfs_set_extent_inline_ref_type(leaf, iref,
7128                                                  BTRFS_TREE_BLOCK_REF_KEY);
7129                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7130         }
7131
7132         btrfs_mark_buffer_dirty(leaf);
7133         btrfs_free_path(path);
7134
7135         if (!no_quota) {
7136                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7137                                               ins->objectid, num_bytes,
7138                                               BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7139                 if (ret)
7140                         return ret;
7141         }
7142
7143         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7144                                  1);
7145         if (ret) { /* -ENOENT, logic error */
7146                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7147                         ins->objectid, ins->offset);
7148                 BUG();
7149         }
7150
7151         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7152         return ret;
7153 }
7154
7155 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7156                                      struct btrfs_root *root,
7157                                      u64 root_objectid, u64 owner,
7158                                      u64 offset, struct btrfs_key *ins)
7159 {
7160         int ret;
7161
7162         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7163
7164         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7165                                          ins->offset, 0,
7166                                          root_objectid, owner, offset,
7167                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7168         return ret;
7169 }
7170
7171 /*
7172  * this is used by the tree logging recovery code.  It records that
7173  * an extent has been allocated and makes sure to clear the free
7174  * space cache bits as well
7175  */
7176 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7177                                    struct btrfs_root *root,
7178                                    u64 root_objectid, u64 owner, u64 offset,
7179                                    struct btrfs_key *ins)
7180 {
7181         int ret;
7182         struct btrfs_block_group_cache *block_group;
7183
7184         /*
7185          * Mixed block groups will exclude before processing the log so we only
7186          * need to do the exlude dance if this fs isn't mixed.
7187          */
7188         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7189                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7190                 if (ret)
7191                         return ret;
7192         }
7193
7194         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7195         if (!block_group)
7196                 return -EINVAL;
7197
7198         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7199                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7200         BUG_ON(ret); /* logic error */
7201         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7202                                          0, owner, offset, ins, 1);
7203         btrfs_put_block_group(block_group);
7204         return ret;
7205 }
7206
7207 static struct extent_buffer *
7208 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7209                       u64 bytenr, int level)
7210 {
7211         struct extent_buffer *buf;
7212
7213         buf = btrfs_find_create_tree_block(root, bytenr);
7214         if (!buf)
7215                 return ERR_PTR(-ENOMEM);
7216         btrfs_set_header_generation(buf, trans->transid);
7217         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7218         btrfs_tree_lock(buf);
7219         clean_tree_block(trans, root->fs_info, buf);
7220         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7221
7222         btrfs_set_lock_blocking(buf);
7223         btrfs_set_buffer_uptodate(buf);
7224
7225         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7226                 buf->log_index = root->log_transid % 2;
7227                 /*
7228                  * we allow two log transactions at a time, use different
7229                  * EXENT bit to differentiate dirty pages.
7230                  */
7231                 if (buf->log_index == 0)
7232                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7233                                         buf->start + buf->len - 1, GFP_NOFS);
7234                 else
7235                         set_extent_new(&root->dirty_log_pages, buf->start,
7236                                         buf->start + buf->len - 1, GFP_NOFS);
7237         } else {
7238                 buf->log_index = -1;
7239                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7240                          buf->start + buf->len - 1, GFP_NOFS);
7241         }
7242         trans->blocks_used++;
7243         /* this returns a buffer locked for blocking */
7244         return buf;
7245 }
7246
7247 static struct btrfs_block_rsv *
7248 use_block_rsv(struct btrfs_trans_handle *trans,
7249               struct btrfs_root *root, u32 blocksize)
7250 {
7251         struct btrfs_block_rsv *block_rsv;
7252         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7253         int ret;
7254         bool global_updated = false;
7255
7256         block_rsv = get_block_rsv(trans, root);
7257
7258         if (unlikely(block_rsv->size == 0))
7259                 goto try_reserve;
7260 again:
7261         ret = block_rsv_use_bytes(block_rsv, blocksize);
7262         if (!ret)
7263                 return block_rsv;
7264
7265         if (block_rsv->failfast)
7266                 return ERR_PTR(ret);
7267
7268         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7269                 global_updated = true;
7270                 update_global_block_rsv(root->fs_info);
7271                 goto again;
7272         }
7273
7274         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7275                 static DEFINE_RATELIMIT_STATE(_rs,
7276                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7277                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7278                 if (__ratelimit(&_rs))
7279                         WARN(1, KERN_DEBUG
7280                                 "BTRFS: block rsv returned %d\n", ret);
7281         }
7282 try_reserve:
7283         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7284                                      BTRFS_RESERVE_NO_FLUSH);
7285         if (!ret)
7286                 return block_rsv;
7287         /*
7288          * If we couldn't reserve metadata bytes try and use some from
7289          * the global reserve if its space type is the same as the global
7290          * reservation.
7291          */
7292         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7293             block_rsv->space_info == global_rsv->space_info) {
7294                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7295                 if (!ret)
7296                         return global_rsv;
7297         }
7298         return ERR_PTR(ret);
7299 }
7300
7301 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7302                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7303 {
7304         block_rsv_add_bytes(block_rsv, blocksize, 0);
7305         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7306 }
7307
7308 /*
7309  * finds a free extent and does all the dirty work required for allocation
7310  * returns the key for the extent through ins, and a tree buffer for
7311  * the first block of the extent through buf.
7312  *
7313  * returns the tree buffer or NULL.
7314  */
7315 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7316                                         struct btrfs_root *root,
7317                                         u64 parent, u64 root_objectid,
7318                                         struct btrfs_disk_key *key, int level,
7319                                         u64 hint, u64 empty_size)
7320 {
7321         struct btrfs_key ins;
7322         struct btrfs_block_rsv *block_rsv;
7323         struct extent_buffer *buf;
7324         u64 flags = 0;
7325         int ret;
7326         u32 blocksize = root->nodesize;
7327         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7328                                                  SKINNY_METADATA);
7329
7330         if (btrfs_test_is_dummy_root(root)) {
7331                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7332                                             level);
7333                 if (!IS_ERR(buf))
7334                         root->alloc_bytenr += blocksize;
7335                 return buf;
7336         }
7337
7338         block_rsv = use_block_rsv(trans, root, blocksize);
7339         if (IS_ERR(block_rsv))
7340                 return ERR_CAST(block_rsv);
7341
7342         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7343                                    empty_size, hint, &ins, 0, 0);
7344         if (ret) {
7345                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7346                 return ERR_PTR(ret);
7347         }
7348
7349         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
7350         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
7351
7352         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7353                 if (parent == 0)
7354                         parent = ins.objectid;
7355                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7356         } else
7357                 BUG_ON(parent > 0);
7358
7359         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7360                 struct btrfs_delayed_extent_op *extent_op;
7361                 extent_op = btrfs_alloc_delayed_extent_op();
7362                 BUG_ON(!extent_op); /* -ENOMEM */
7363                 if (key)
7364                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7365                 else
7366                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7367                 extent_op->flags_to_set = flags;
7368                 if (skinny_metadata)
7369                         extent_op->update_key = 0;
7370                 else
7371                         extent_op->update_key = 1;
7372                 extent_op->update_flags = 1;
7373                 extent_op->is_data = 0;
7374                 extent_op->level = level;
7375
7376                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7377                                         ins.objectid,
7378                                         ins.offset, parent, root_objectid,
7379                                         level, BTRFS_ADD_DELAYED_EXTENT,
7380                                         extent_op, 0);
7381                 BUG_ON(ret); /* -ENOMEM */
7382         }
7383         return buf;
7384 }
7385
7386 struct walk_control {
7387         u64 refs[BTRFS_MAX_LEVEL];
7388         u64 flags[BTRFS_MAX_LEVEL];
7389         struct btrfs_key update_progress;
7390         int stage;
7391         int level;
7392         int shared_level;
7393         int update_ref;
7394         int keep_locks;
7395         int reada_slot;
7396         int reada_count;
7397         int for_reloc;
7398 };
7399
7400 #define DROP_REFERENCE  1
7401 #define UPDATE_BACKREF  2
7402
7403 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7404                                      struct btrfs_root *root,
7405                                      struct walk_control *wc,
7406                                      struct btrfs_path *path)
7407 {
7408         u64 bytenr;
7409         u64 generation;
7410         u64 refs;
7411         u64 flags;
7412         u32 nritems;
7413         u32 blocksize;
7414         struct btrfs_key key;
7415         struct extent_buffer *eb;
7416         int ret;
7417         int slot;
7418         int nread = 0;
7419
7420         if (path->slots[wc->level] < wc->reada_slot) {
7421                 wc->reada_count = wc->reada_count * 2 / 3;
7422                 wc->reada_count = max(wc->reada_count, 2);
7423         } else {
7424                 wc->reada_count = wc->reada_count * 3 / 2;
7425                 wc->reada_count = min_t(int, wc->reada_count,
7426                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7427         }
7428
7429         eb = path->nodes[wc->level];
7430         nritems = btrfs_header_nritems(eb);
7431         blocksize = root->nodesize;
7432
7433         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7434                 if (nread >= wc->reada_count)
7435                         break;
7436
7437                 cond_resched();
7438                 bytenr = btrfs_node_blockptr(eb, slot);
7439                 generation = btrfs_node_ptr_generation(eb, slot);
7440
7441                 if (slot == path->slots[wc->level])
7442                         goto reada;
7443
7444                 if (wc->stage == UPDATE_BACKREF &&
7445                     generation <= root->root_key.offset)
7446                         continue;
7447
7448                 /* We don't lock the tree block, it's OK to be racy here */
7449                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7450                                                wc->level - 1, 1, &refs,
7451                                                &flags);
7452                 /* We don't care about errors in readahead. */
7453                 if (ret < 0)
7454                         continue;
7455                 BUG_ON(refs == 0);
7456
7457                 if (wc->stage == DROP_REFERENCE) {
7458                         if (refs == 1)
7459                                 goto reada;
7460
7461                         if (wc->level == 1 &&
7462                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7463                                 continue;
7464                         if (!wc->update_ref ||
7465                             generation <= root->root_key.offset)
7466                                 continue;
7467                         btrfs_node_key_to_cpu(eb, &key, slot);
7468                         ret = btrfs_comp_cpu_keys(&key,
7469                                                   &wc->update_progress);
7470                         if (ret < 0)
7471                                 continue;
7472                 } else {
7473                         if (wc->level == 1 &&
7474                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7475                                 continue;
7476                 }
7477 reada:
7478                 readahead_tree_block(root, bytenr);
7479                 nread++;
7480         }
7481         wc->reada_slot = slot;
7482 }
7483
7484 static int account_leaf_items(struct btrfs_trans_handle *trans,
7485                               struct btrfs_root *root,
7486                               struct extent_buffer *eb)
7487 {
7488         int nr = btrfs_header_nritems(eb);
7489         int i, extent_type, ret;
7490         struct btrfs_key key;
7491         struct btrfs_file_extent_item *fi;
7492         u64 bytenr, num_bytes;
7493
7494         for (i = 0; i < nr; i++) {
7495                 btrfs_item_key_to_cpu(eb, &key, i);
7496
7497                 if (key.type != BTRFS_EXTENT_DATA_KEY)
7498                         continue;
7499
7500                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
7501                 /* filter out non qgroup-accountable extents  */
7502                 extent_type = btrfs_file_extent_type(eb, fi);
7503
7504                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
7505                         continue;
7506
7507                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
7508                 if (!bytenr)
7509                         continue;
7510
7511                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
7512
7513                 ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7514                                               root->objectid,
7515                                               bytenr, num_bytes,
7516                                               BTRFS_QGROUP_OPER_SUB_SUBTREE, 0);
7517                 if (ret)
7518                         return ret;
7519         }
7520         return 0;
7521 }
7522
7523 /*
7524  * Walk up the tree from the bottom, freeing leaves and any interior
7525  * nodes which have had all slots visited. If a node (leaf or
7526  * interior) is freed, the node above it will have it's slot
7527  * incremented. The root node will never be freed.
7528  *
7529  * At the end of this function, we should have a path which has all
7530  * slots incremented to the next position for a search. If we need to
7531  * read a new node it will be NULL and the node above it will have the
7532  * correct slot selected for a later read.
7533  *
7534  * If we increment the root nodes slot counter past the number of
7535  * elements, 1 is returned to signal completion of the search.
7536  */
7537 static int adjust_slots_upwards(struct btrfs_root *root,
7538                                 struct btrfs_path *path, int root_level)
7539 {
7540         int level = 0;
7541         int nr, slot;
7542         struct extent_buffer *eb;
7543
7544         if (root_level == 0)
7545                 return 1;
7546
7547         while (level <= root_level) {
7548                 eb = path->nodes[level];
7549                 nr = btrfs_header_nritems(eb);
7550                 path->slots[level]++;
7551                 slot = path->slots[level];
7552                 if (slot >= nr || level == 0) {
7553                         /*
7554                          * Don't free the root -  we will detect this
7555                          * condition after our loop and return a
7556                          * positive value for caller to stop walking the tree.
7557                          */
7558                         if (level != root_level) {
7559                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7560                                 path->locks[level] = 0;
7561
7562                                 free_extent_buffer(eb);
7563                                 path->nodes[level] = NULL;
7564                                 path->slots[level] = 0;
7565                         }
7566                 } else {
7567                         /*
7568                          * We have a valid slot to walk back down
7569                          * from. Stop here so caller can process these
7570                          * new nodes.
7571                          */
7572                         break;
7573                 }
7574
7575                 level++;
7576         }
7577
7578         eb = path->nodes[root_level];
7579         if (path->slots[root_level] >= btrfs_header_nritems(eb))
7580                 return 1;
7581
7582         return 0;
7583 }
7584
7585 /*
7586  * root_eb is the subtree root and is locked before this function is called.
7587  */
7588 static int account_shared_subtree(struct btrfs_trans_handle *trans,
7589                                   struct btrfs_root *root,
7590                                   struct extent_buffer *root_eb,
7591                                   u64 root_gen,
7592                                   int root_level)
7593 {
7594         int ret = 0;
7595         int level;
7596         struct extent_buffer *eb = root_eb;
7597         struct btrfs_path *path = NULL;
7598
7599         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
7600         BUG_ON(root_eb == NULL);
7601
7602         if (!root->fs_info->quota_enabled)
7603                 return 0;
7604
7605         if (!extent_buffer_uptodate(root_eb)) {
7606                 ret = btrfs_read_buffer(root_eb, root_gen);
7607                 if (ret)
7608                         goto out;
7609         }
7610
7611         if (root_level == 0) {
7612                 ret = account_leaf_items(trans, root, root_eb);
7613                 goto out;
7614         }
7615
7616         path = btrfs_alloc_path();
7617         if (!path)
7618                 return -ENOMEM;
7619
7620         /*
7621          * Walk down the tree.  Missing extent blocks are filled in as
7622          * we go. Metadata is accounted every time we read a new
7623          * extent block.
7624          *
7625          * When we reach a leaf, we account for file extent items in it,
7626          * walk back up the tree (adjusting slot pointers as we go)
7627          * and restart the search process.
7628          */
7629         extent_buffer_get(root_eb); /* For path */
7630         path->nodes[root_level] = root_eb;
7631         path->slots[root_level] = 0;
7632         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
7633 walk_down:
7634         level = root_level;
7635         while (level >= 0) {
7636                 if (path->nodes[level] == NULL) {
7637                         int parent_slot;
7638                         u64 child_gen;
7639                         u64 child_bytenr;
7640
7641                         /* We need to get child blockptr/gen from
7642                          * parent before we can read it. */
7643                         eb = path->nodes[level + 1];
7644                         parent_slot = path->slots[level + 1];
7645                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
7646                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
7647
7648                         eb = read_tree_block(root, child_bytenr, child_gen);
7649                         if (!eb || !extent_buffer_uptodate(eb)) {
7650                                 ret = -EIO;
7651                                 goto out;
7652                         }
7653
7654                         path->nodes[level] = eb;
7655                         path->slots[level] = 0;
7656
7657                         btrfs_tree_read_lock(eb);
7658                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
7659                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
7660
7661                         ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7662                                                 root->objectid,
7663                                                 child_bytenr,
7664                                                 root->nodesize,
7665                                                 BTRFS_QGROUP_OPER_SUB_SUBTREE,
7666                                                 0);
7667                         if (ret)
7668                                 goto out;
7669
7670                 }
7671
7672                 if (level == 0) {
7673                         ret = account_leaf_items(trans, root, path->nodes[level]);
7674                         if (ret)
7675                                 goto out;
7676
7677                         /* Nonzero return here means we completed our search */
7678                         ret = adjust_slots_upwards(root, path, root_level);
7679                         if (ret)
7680                                 break;
7681
7682                         /* Restart search with new slots */
7683                         goto walk_down;
7684                 }
7685
7686                 level--;
7687         }
7688
7689         ret = 0;
7690 out:
7691         btrfs_free_path(path);
7692
7693         return ret;
7694 }
7695
7696 /*
7697  * helper to process tree block while walking down the tree.
7698  *
7699  * when wc->stage == UPDATE_BACKREF, this function updates
7700  * back refs for pointers in the block.
7701  *
7702  * NOTE: return value 1 means we should stop walking down.
7703  */
7704 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7705                                    struct btrfs_root *root,
7706                                    struct btrfs_path *path,
7707                                    struct walk_control *wc, int lookup_info)
7708 {
7709         int level = wc->level;
7710         struct extent_buffer *eb = path->nodes[level];
7711         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7712         int ret;
7713
7714         if (wc->stage == UPDATE_BACKREF &&
7715             btrfs_header_owner(eb) != root->root_key.objectid)
7716                 return 1;
7717
7718         /*
7719          * when reference count of tree block is 1, it won't increase
7720          * again. once full backref flag is set, we never clear it.
7721          */
7722         if (lookup_info &&
7723             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7724              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7725                 BUG_ON(!path->locks[level]);
7726                 ret = btrfs_lookup_extent_info(trans, root,
7727                                                eb->start, level, 1,
7728                                                &wc->refs[level],
7729                                                &wc->flags[level]);
7730                 BUG_ON(ret == -ENOMEM);
7731                 if (ret)
7732                         return ret;
7733                 BUG_ON(wc->refs[level] == 0);
7734         }
7735
7736         if (wc->stage == DROP_REFERENCE) {
7737                 if (wc->refs[level] > 1)
7738                         return 1;
7739
7740                 if (path->locks[level] && !wc->keep_locks) {
7741                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7742                         path->locks[level] = 0;
7743                 }
7744                 return 0;
7745         }
7746
7747         /* wc->stage == UPDATE_BACKREF */
7748         if (!(wc->flags[level] & flag)) {
7749                 BUG_ON(!path->locks[level]);
7750                 ret = btrfs_inc_ref(trans, root, eb, 1);
7751                 BUG_ON(ret); /* -ENOMEM */
7752                 ret = btrfs_dec_ref(trans, root, eb, 0);
7753                 BUG_ON(ret); /* -ENOMEM */
7754                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7755                                                   eb->len, flag,
7756                                                   btrfs_header_level(eb), 0);
7757                 BUG_ON(ret); /* -ENOMEM */
7758                 wc->flags[level] |= flag;
7759         }
7760
7761         /*
7762          * the block is shared by multiple trees, so it's not good to
7763          * keep the tree lock
7764          */
7765         if (path->locks[level] && level > 0) {
7766                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7767                 path->locks[level] = 0;
7768         }
7769         return 0;
7770 }
7771
7772 /*
7773  * helper to process tree block pointer.
7774  *
7775  * when wc->stage == DROP_REFERENCE, this function checks
7776  * reference count of the block pointed to. if the block
7777  * is shared and we need update back refs for the subtree
7778  * rooted at the block, this function changes wc->stage to
7779  * UPDATE_BACKREF. if the block is shared and there is no
7780  * need to update back, this function drops the reference
7781  * to the block.
7782  *
7783  * NOTE: return value 1 means we should stop walking down.
7784  */
7785 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7786                                  struct btrfs_root *root,
7787                                  struct btrfs_path *path,
7788                                  struct walk_control *wc, int *lookup_info)
7789 {
7790         u64 bytenr;
7791         u64 generation;
7792         u64 parent;
7793         u32 blocksize;
7794         struct btrfs_key key;
7795         struct extent_buffer *next;
7796         int level = wc->level;
7797         int reada = 0;
7798         int ret = 0;
7799         bool need_account = false;
7800
7801         generation = btrfs_node_ptr_generation(path->nodes[level],
7802                                                path->slots[level]);
7803         /*
7804          * if the lower level block was created before the snapshot
7805          * was created, we know there is no need to update back refs
7806          * for the subtree
7807          */
7808         if (wc->stage == UPDATE_BACKREF &&
7809             generation <= root->root_key.offset) {
7810                 *lookup_info = 1;
7811                 return 1;
7812         }
7813
7814         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7815         blocksize = root->nodesize;
7816
7817         next = btrfs_find_tree_block(root->fs_info, bytenr);
7818         if (!next) {
7819                 next = btrfs_find_create_tree_block(root, bytenr);
7820                 if (!next)
7821                         return -ENOMEM;
7822                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
7823                                                level - 1);
7824                 reada = 1;
7825         }
7826         btrfs_tree_lock(next);
7827         btrfs_set_lock_blocking(next);
7828
7829         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7830                                        &wc->refs[level - 1],
7831                                        &wc->flags[level - 1]);
7832         if (ret < 0) {
7833                 btrfs_tree_unlock(next);
7834                 return ret;
7835         }
7836
7837         if (unlikely(wc->refs[level - 1] == 0)) {
7838                 btrfs_err(root->fs_info, "Missing references.");
7839                 BUG();
7840         }
7841         *lookup_info = 0;
7842
7843         if (wc->stage == DROP_REFERENCE) {
7844                 if (wc->refs[level - 1] > 1) {
7845                         need_account = true;
7846                         if (level == 1 &&
7847                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7848                                 goto skip;
7849
7850                         if (!wc->update_ref ||
7851                             generation <= root->root_key.offset)
7852                                 goto skip;
7853
7854                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7855                                               path->slots[level]);
7856                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7857                         if (ret < 0)
7858                                 goto skip;
7859
7860                         wc->stage = UPDATE_BACKREF;
7861                         wc->shared_level = level - 1;
7862                 }
7863         } else {
7864                 if (level == 1 &&
7865                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7866                         goto skip;
7867         }
7868
7869         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7870                 btrfs_tree_unlock(next);
7871                 free_extent_buffer(next);
7872                 next = NULL;
7873                 *lookup_info = 1;
7874         }
7875
7876         if (!next) {
7877                 if (reada && level == 1)
7878                         reada_walk_down(trans, root, wc, path);
7879                 next = read_tree_block(root, bytenr, generation);
7880                 if (!next || !extent_buffer_uptodate(next)) {
7881                         free_extent_buffer(next);
7882                         return -EIO;
7883                 }
7884                 btrfs_tree_lock(next);
7885                 btrfs_set_lock_blocking(next);
7886         }
7887
7888         level--;
7889         BUG_ON(level != btrfs_header_level(next));
7890         path->nodes[level] = next;
7891         path->slots[level] = 0;
7892         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7893         wc->level = level;
7894         if (wc->level == 1)
7895                 wc->reada_slot = 0;
7896         return 0;
7897 skip:
7898         wc->refs[level - 1] = 0;
7899         wc->flags[level - 1] = 0;
7900         if (wc->stage == DROP_REFERENCE) {
7901                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7902                         parent = path->nodes[level]->start;
7903                 } else {
7904                         BUG_ON(root->root_key.objectid !=
7905                                btrfs_header_owner(path->nodes[level]));
7906                         parent = 0;
7907                 }
7908
7909                 if (need_account) {
7910                         ret = account_shared_subtree(trans, root, next,
7911                                                      generation, level - 1);
7912                         if (ret) {
7913                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
7914                                         "%d accounting shared subtree. Quota "
7915                                         "is out of sync, rescan required.\n",
7916                                         root->fs_info->sb->s_id, ret);
7917                         }
7918                 }
7919                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7920                                 root->root_key.objectid, level - 1, 0, 0);
7921                 BUG_ON(ret); /* -ENOMEM */
7922         }
7923         btrfs_tree_unlock(next);
7924         free_extent_buffer(next);
7925         *lookup_info = 1;
7926         return 1;
7927 }
7928
7929 /*
7930  * helper to process tree block while walking up the tree.
7931  *
7932  * when wc->stage == DROP_REFERENCE, this function drops
7933  * reference count on the block.
7934  *
7935  * when wc->stage == UPDATE_BACKREF, this function changes
7936  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7937  * to UPDATE_BACKREF previously while processing the block.
7938  *
7939  * NOTE: return value 1 means we should stop walking up.
7940  */
7941 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7942                                  struct btrfs_root *root,
7943                                  struct btrfs_path *path,
7944                                  struct walk_control *wc)
7945 {
7946         int ret;
7947         int level = wc->level;
7948         struct extent_buffer *eb = path->nodes[level];
7949         u64 parent = 0;
7950
7951         if (wc->stage == UPDATE_BACKREF) {
7952                 BUG_ON(wc->shared_level < level);
7953                 if (level < wc->shared_level)
7954                         goto out;
7955
7956                 ret = find_next_key(path, level + 1, &wc->update_progress);
7957                 if (ret > 0)
7958                         wc->update_ref = 0;
7959
7960                 wc->stage = DROP_REFERENCE;
7961                 wc->shared_level = -1;
7962                 path->slots[level] = 0;
7963
7964                 /*
7965                  * check reference count again if the block isn't locked.
7966                  * we should start walking down the tree again if reference
7967                  * count is one.
7968                  */
7969                 if (!path->locks[level]) {
7970                         BUG_ON(level == 0);
7971                         btrfs_tree_lock(eb);
7972                         btrfs_set_lock_blocking(eb);
7973                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7974
7975                         ret = btrfs_lookup_extent_info(trans, root,
7976                                                        eb->start, level, 1,
7977                                                        &wc->refs[level],
7978                                                        &wc->flags[level]);
7979                         if (ret < 0) {
7980                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7981                                 path->locks[level] = 0;
7982                                 return ret;
7983                         }
7984                         BUG_ON(wc->refs[level] == 0);
7985                         if (wc->refs[level] == 1) {
7986                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7987                                 path->locks[level] = 0;
7988                                 return 1;
7989                         }
7990                 }
7991         }
7992
7993         /* wc->stage == DROP_REFERENCE */
7994         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7995
7996         if (wc->refs[level] == 1) {
7997                 if (level == 0) {
7998                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7999                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8000                         else
8001                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8002                         BUG_ON(ret); /* -ENOMEM */
8003                         ret = account_leaf_items(trans, root, eb);
8004                         if (ret) {
8005                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8006                                         "%d accounting leaf items. Quota "
8007                                         "is out of sync, rescan required.\n",
8008                                         root->fs_info->sb->s_id, ret);
8009                         }
8010                 }
8011                 /* make block locked assertion in clean_tree_block happy */
8012                 if (!path->locks[level] &&
8013                     btrfs_header_generation(eb) == trans->transid) {
8014                         btrfs_tree_lock(eb);
8015                         btrfs_set_lock_blocking(eb);
8016                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8017                 }
8018                 clean_tree_block(trans, root->fs_info, eb);
8019         }
8020
8021         if (eb == root->node) {
8022                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8023                         parent = eb->start;
8024                 else
8025                         BUG_ON(root->root_key.objectid !=
8026                                btrfs_header_owner(eb));
8027         } else {
8028                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8029                         parent = path->nodes[level + 1]->start;
8030                 else
8031                         BUG_ON(root->root_key.objectid !=
8032                                btrfs_header_owner(path->nodes[level + 1]));
8033         }
8034
8035         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8036 out:
8037         wc->refs[level] = 0;
8038         wc->flags[level] = 0;
8039         return 0;
8040 }
8041
8042 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8043                                    struct btrfs_root *root,
8044                                    struct btrfs_path *path,
8045                                    struct walk_control *wc)
8046 {
8047         int level = wc->level;
8048         int lookup_info = 1;
8049         int ret;
8050
8051         while (level >= 0) {
8052                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8053                 if (ret > 0)
8054                         break;
8055
8056                 if (level == 0)
8057                         break;
8058
8059                 if (path->slots[level] >=
8060                     btrfs_header_nritems(path->nodes[level]))
8061                         break;
8062
8063                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8064                 if (ret > 0) {
8065                         path->slots[level]++;
8066                         continue;
8067                 } else if (ret < 0)
8068                         return ret;
8069                 level = wc->level;
8070         }
8071         return 0;
8072 }
8073
8074 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8075                                  struct btrfs_root *root,
8076                                  struct btrfs_path *path,
8077                                  struct walk_control *wc, int max_level)
8078 {
8079         int level = wc->level;
8080         int ret;
8081
8082         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8083         while (level < max_level && path->nodes[level]) {
8084                 wc->level = level;
8085                 if (path->slots[level] + 1 <
8086                     btrfs_header_nritems(path->nodes[level])) {
8087                         path->slots[level]++;
8088                         return 0;
8089                 } else {
8090                         ret = walk_up_proc(trans, root, path, wc);
8091                         if (ret > 0)
8092                                 return 0;
8093
8094                         if (path->locks[level]) {
8095                                 btrfs_tree_unlock_rw(path->nodes[level],
8096                                                      path->locks[level]);
8097                                 path->locks[level] = 0;
8098                         }
8099                         free_extent_buffer(path->nodes[level]);
8100                         path->nodes[level] = NULL;
8101                         level++;
8102                 }
8103         }
8104         return 1;
8105 }
8106
8107 /*
8108  * drop a subvolume tree.
8109  *
8110  * this function traverses the tree freeing any blocks that only
8111  * referenced by the tree.
8112  *
8113  * when a shared tree block is found. this function decreases its
8114  * reference count by one. if update_ref is true, this function
8115  * also make sure backrefs for the shared block and all lower level
8116  * blocks are properly updated.
8117  *
8118  * If called with for_reloc == 0, may exit early with -EAGAIN
8119  */
8120 int btrfs_drop_snapshot(struct btrfs_root *root,
8121                          struct btrfs_block_rsv *block_rsv, int update_ref,
8122                          int for_reloc)
8123 {
8124         struct btrfs_path *path;
8125         struct btrfs_trans_handle *trans;
8126         struct btrfs_root *tree_root = root->fs_info->tree_root;
8127         struct btrfs_root_item *root_item = &root->root_item;
8128         struct walk_control *wc;
8129         struct btrfs_key key;
8130         int err = 0;
8131         int ret;
8132         int level;
8133         bool root_dropped = false;
8134
8135         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8136
8137         path = btrfs_alloc_path();
8138         if (!path) {
8139                 err = -ENOMEM;
8140                 goto out;
8141         }
8142
8143         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8144         if (!wc) {
8145                 btrfs_free_path(path);
8146                 err = -ENOMEM;
8147                 goto out;
8148         }
8149
8150         trans = btrfs_start_transaction(tree_root, 0);
8151         if (IS_ERR(trans)) {
8152                 err = PTR_ERR(trans);
8153                 goto out_free;
8154         }
8155
8156         if (block_rsv)
8157                 trans->block_rsv = block_rsv;
8158
8159         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8160                 level = btrfs_header_level(root->node);
8161                 path->nodes[level] = btrfs_lock_root_node(root);
8162                 btrfs_set_lock_blocking(path->nodes[level]);
8163                 path->slots[level] = 0;
8164                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8165                 memset(&wc->update_progress, 0,
8166                        sizeof(wc->update_progress));
8167         } else {
8168                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8169                 memcpy(&wc->update_progress, &key,
8170                        sizeof(wc->update_progress));
8171
8172                 level = root_item->drop_level;
8173                 BUG_ON(level == 0);
8174                 path->lowest_level = level;
8175                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8176                 path->lowest_level = 0;
8177                 if (ret < 0) {
8178                         err = ret;
8179                         goto out_end_trans;
8180                 }
8181                 WARN_ON(ret > 0);
8182
8183                 /*
8184                  * unlock our path, this is safe because only this
8185                  * function is allowed to delete this snapshot
8186                  */
8187                 btrfs_unlock_up_safe(path, 0);
8188
8189                 level = btrfs_header_level(root->node);
8190                 while (1) {
8191                         btrfs_tree_lock(path->nodes[level]);
8192                         btrfs_set_lock_blocking(path->nodes[level]);
8193                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8194
8195                         ret = btrfs_lookup_extent_info(trans, root,
8196                                                 path->nodes[level]->start,
8197                                                 level, 1, &wc->refs[level],
8198                                                 &wc->flags[level]);
8199                         if (ret < 0) {
8200                                 err = ret;
8201                                 goto out_end_trans;
8202                         }
8203                         BUG_ON(wc->refs[level] == 0);
8204
8205                         if (level == root_item->drop_level)
8206                                 break;
8207
8208                         btrfs_tree_unlock(path->nodes[level]);
8209                         path->locks[level] = 0;
8210                         WARN_ON(wc->refs[level] != 1);
8211                         level--;
8212                 }
8213         }
8214
8215         wc->level = level;
8216         wc->shared_level = -1;
8217         wc->stage = DROP_REFERENCE;
8218         wc->update_ref = update_ref;
8219         wc->keep_locks = 0;
8220         wc->for_reloc = for_reloc;
8221         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8222
8223         while (1) {
8224
8225                 ret = walk_down_tree(trans, root, path, wc);
8226                 if (ret < 0) {
8227                         err = ret;
8228                         break;
8229                 }
8230
8231                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8232                 if (ret < 0) {
8233                         err = ret;
8234                         break;
8235                 }
8236
8237                 if (ret > 0) {
8238                         BUG_ON(wc->stage != DROP_REFERENCE);
8239                         break;
8240                 }
8241
8242                 if (wc->stage == DROP_REFERENCE) {
8243                         level = wc->level;
8244                         btrfs_node_key(path->nodes[level],
8245                                        &root_item->drop_progress,
8246                                        path->slots[level]);
8247                         root_item->drop_level = level;
8248                 }
8249
8250                 BUG_ON(wc->level == 0);
8251                 if (btrfs_should_end_transaction(trans, tree_root) ||
8252                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8253                         ret = btrfs_update_root(trans, tree_root,
8254                                                 &root->root_key,
8255                                                 root_item);
8256                         if (ret) {
8257                                 btrfs_abort_transaction(trans, tree_root, ret);
8258                                 err = ret;
8259                                 goto out_end_trans;
8260                         }
8261
8262                         /*
8263                          * Qgroup update accounting is run from
8264                          * delayed ref handling. This usually works
8265                          * out because delayed refs are normally the
8266                          * only way qgroup updates are added. However,
8267                          * we may have added updates during our tree
8268                          * walk so run qgroups here to make sure we
8269                          * don't lose any updates.
8270                          */
8271                         ret = btrfs_delayed_qgroup_accounting(trans,
8272                                                               root->fs_info);
8273                         if (ret)
8274                                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8275                                                    "running qgroup updates "
8276                                                    "during snapshot delete. "
8277                                                    "Quota is out of sync, "
8278                                                    "rescan required.\n", ret);
8279
8280                         btrfs_end_transaction_throttle(trans, tree_root);
8281                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8282                                 pr_debug("BTRFS: drop snapshot early exit\n");
8283                                 err = -EAGAIN;
8284                                 goto out_free;
8285                         }
8286
8287                         trans = btrfs_start_transaction(tree_root, 0);
8288                         if (IS_ERR(trans)) {
8289                                 err = PTR_ERR(trans);
8290                                 goto out_free;
8291                         }
8292                         if (block_rsv)
8293                                 trans->block_rsv = block_rsv;
8294                 }
8295         }
8296         btrfs_release_path(path);
8297         if (err)
8298                 goto out_end_trans;
8299
8300         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8301         if (ret) {
8302                 btrfs_abort_transaction(trans, tree_root, ret);
8303                 goto out_end_trans;
8304         }
8305
8306         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8307                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8308                                       NULL, NULL);
8309                 if (ret < 0) {
8310                         btrfs_abort_transaction(trans, tree_root, ret);
8311                         err = ret;
8312                         goto out_end_trans;
8313                 } else if (ret > 0) {
8314                         /* if we fail to delete the orphan item this time
8315                          * around, it'll get picked up the next time.
8316                          *
8317                          * The most common failure here is just -ENOENT.
8318                          */
8319                         btrfs_del_orphan_item(trans, tree_root,
8320                                               root->root_key.objectid);
8321                 }
8322         }
8323
8324         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8325                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
8326         } else {
8327                 free_extent_buffer(root->node);
8328                 free_extent_buffer(root->commit_root);
8329                 btrfs_put_fs_root(root);
8330         }
8331         root_dropped = true;
8332 out_end_trans:
8333         ret = btrfs_delayed_qgroup_accounting(trans, tree_root->fs_info);
8334         if (ret)
8335                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8336                                    "running qgroup updates "
8337                                    "during snapshot delete. "
8338                                    "Quota is out of sync, "
8339                                    "rescan required.\n", ret);
8340
8341         btrfs_end_transaction_throttle(trans, tree_root);
8342 out_free:
8343         kfree(wc);
8344         btrfs_free_path(path);
8345 out:
8346         /*
8347          * So if we need to stop dropping the snapshot for whatever reason we
8348          * need to make sure to add it back to the dead root list so that we
8349          * keep trying to do the work later.  This also cleans up roots if we
8350          * don't have it in the radix (like when we recover after a power fail
8351          * or unmount) so we don't leak memory.
8352          */
8353         if (!for_reloc && root_dropped == false)
8354                 btrfs_add_dead_root(root);
8355         if (err && err != -EAGAIN)
8356                 btrfs_std_error(root->fs_info, err);
8357         return err;
8358 }
8359
8360 /*
8361  * drop subtree rooted at tree block 'node'.
8362  *
8363  * NOTE: this function will unlock and release tree block 'node'
8364  * only used by relocation code
8365  */
8366 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8367                         struct btrfs_root *root,
8368                         struct extent_buffer *node,
8369                         struct extent_buffer *parent)
8370 {
8371         struct btrfs_path *path;
8372         struct walk_control *wc;
8373         int level;
8374         int parent_level;
8375         int ret = 0;
8376         int wret;
8377
8378         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8379
8380         path = btrfs_alloc_path();
8381         if (!path)
8382                 return -ENOMEM;
8383
8384         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8385         if (!wc) {
8386                 btrfs_free_path(path);
8387                 return -ENOMEM;
8388         }
8389
8390         btrfs_assert_tree_locked(parent);
8391         parent_level = btrfs_header_level(parent);
8392         extent_buffer_get(parent);
8393         path->nodes[parent_level] = parent;
8394         path->slots[parent_level] = btrfs_header_nritems(parent);
8395
8396         btrfs_assert_tree_locked(node);
8397         level = btrfs_header_level(node);
8398         path->nodes[level] = node;
8399         path->slots[level] = 0;
8400         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8401
8402         wc->refs[parent_level] = 1;
8403         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8404         wc->level = level;
8405         wc->shared_level = -1;
8406         wc->stage = DROP_REFERENCE;
8407         wc->update_ref = 0;
8408         wc->keep_locks = 1;
8409         wc->for_reloc = 1;
8410         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8411
8412         while (1) {
8413                 wret = walk_down_tree(trans, root, path, wc);
8414                 if (wret < 0) {
8415                         ret = wret;
8416                         break;
8417                 }
8418
8419                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8420                 if (wret < 0)
8421                         ret = wret;
8422                 if (wret != 0)
8423                         break;
8424         }
8425
8426         kfree(wc);
8427         btrfs_free_path(path);
8428         return ret;
8429 }
8430
8431 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8432 {
8433         u64 num_devices;
8434         u64 stripped;
8435
8436         /*
8437          * if restripe for this chunk_type is on pick target profile and
8438          * return, otherwise do the usual balance
8439          */
8440         stripped = get_restripe_target(root->fs_info, flags);
8441         if (stripped)
8442                 return extended_to_chunk(stripped);
8443
8444         num_devices = root->fs_info->fs_devices->rw_devices;
8445
8446         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8447                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8448                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8449
8450         if (num_devices == 1) {
8451                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8452                 stripped = flags & ~stripped;
8453
8454                 /* turn raid0 into single device chunks */
8455                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8456                         return stripped;
8457
8458                 /* turn mirroring into duplication */
8459                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8460                              BTRFS_BLOCK_GROUP_RAID10))
8461                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8462         } else {
8463                 /* they already had raid on here, just return */
8464                 if (flags & stripped)
8465                         return flags;
8466
8467                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8468                 stripped = flags & ~stripped;
8469
8470                 /* switch duplicated blocks with raid1 */
8471                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8472                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8473
8474                 /* this is drive concat, leave it alone */
8475         }
8476
8477         return flags;
8478 }
8479
8480 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8481 {
8482         struct btrfs_space_info *sinfo = cache->space_info;
8483         u64 num_bytes;
8484         u64 min_allocable_bytes;
8485         int ret = -ENOSPC;
8486
8487
8488         /*
8489          * We need some metadata space and system metadata space for
8490          * allocating chunks in some corner cases until we force to set
8491          * it to be readonly.
8492          */
8493         if ((sinfo->flags &
8494              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8495             !force)
8496                 min_allocable_bytes = 1 * 1024 * 1024;
8497         else
8498                 min_allocable_bytes = 0;
8499
8500         spin_lock(&sinfo->lock);
8501         spin_lock(&cache->lock);
8502
8503         if (cache->ro) {
8504                 ret = 0;
8505                 goto out;
8506         }
8507
8508         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8509                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8510
8511         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8512             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8513             min_allocable_bytes <= sinfo->total_bytes) {
8514                 sinfo->bytes_readonly += num_bytes;
8515                 cache->ro = 1;
8516                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
8517                 ret = 0;
8518         }
8519 out:
8520         spin_unlock(&cache->lock);
8521         spin_unlock(&sinfo->lock);
8522         return ret;
8523 }
8524
8525 int btrfs_set_block_group_ro(struct btrfs_root *root,
8526                              struct btrfs_block_group_cache *cache)
8527
8528 {
8529         struct btrfs_trans_handle *trans;
8530         u64 alloc_flags;
8531         int ret;
8532
8533         BUG_ON(cache->ro);
8534
8535         trans = btrfs_join_transaction(root);
8536         if (IS_ERR(trans))
8537                 return PTR_ERR(trans);
8538
8539         ret = set_block_group_ro(cache, 0);
8540         if (!ret)
8541                 goto out;
8542         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8543         ret = do_chunk_alloc(trans, root, alloc_flags,
8544                              CHUNK_ALLOC_FORCE);
8545         if (ret < 0)
8546                 goto out;
8547         ret = set_block_group_ro(cache, 0);
8548 out:
8549         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
8550                 alloc_flags = update_block_group_flags(root, cache->flags);
8551                 check_system_chunk(trans, root, alloc_flags);
8552         }
8553
8554         btrfs_end_transaction(trans, root);
8555         return ret;
8556 }
8557
8558 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8559                             struct btrfs_root *root, u64 type)
8560 {
8561         u64 alloc_flags = get_alloc_profile(root, type);
8562         return do_chunk_alloc(trans, root, alloc_flags,
8563                               CHUNK_ALLOC_FORCE);
8564 }
8565
8566 /*
8567  * helper to account the unused space of all the readonly block group in the
8568  * space_info. takes mirrors into account.
8569  */
8570 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8571 {
8572         struct btrfs_block_group_cache *block_group;
8573         u64 free_bytes = 0;
8574         int factor;
8575
8576         /* It's df, we don't care if it's racey */
8577         if (list_empty(&sinfo->ro_bgs))
8578                 return 0;
8579
8580         spin_lock(&sinfo->lock);
8581         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
8582                 spin_lock(&block_group->lock);
8583
8584                 if (!block_group->ro) {
8585                         spin_unlock(&block_group->lock);
8586                         continue;
8587                 }
8588
8589                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8590                                           BTRFS_BLOCK_GROUP_RAID10 |
8591                                           BTRFS_BLOCK_GROUP_DUP))
8592                         factor = 2;
8593                 else
8594                         factor = 1;
8595
8596                 free_bytes += (block_group->key.offset -
8597                                btrfs_block_group_used(&block_group->item)) *
8598                                factor;
8599
8600                 spin_unlock(&block_group->lock);
8601         }
8602         spin_unlock(&sinfo->lock);
8603
8604         return free_bytes;
8605 }
8606
8607 void btrfs_set_block_group_rw(struct btrfs_root *root,
8608                               struct btrfs_block_group_cache *cache)
8609 {
8610         struct btrfs_space_info *sinfo = cache->space_info;
8611         u64 num_bytes;
8612
8613         BUG_ON(!cache->ro);
8614
8615         spin_lock(&sinfo->lock);
8616         spin_lock(&cache->lock);
8617         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8618                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8619         sinfo->bytes_readonly -= num_bytes;
8620         cache->ro = 0;
8621         list_del_init(&cache->ro_list);
8622         spin_unlock(&cache->lock);
8623         spin_unlock(&sinfo->lock);
8624 }
8625
8626 /*
8627  * checks to see if its even possible to relocate this block group.
8628  *
8629  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8630  * ok to go ahead and try.
8631  */
8632 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8633 {
8634         struct btrfs_block_group_cache *block_group;
8635         struct btrfs_space_info *space_info;
8636         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8637         struct btrfs_device *device;
8638         struct btrfs_trans_handle *trans;
8639         u64 min_free;
8640         u64 dev_min = 1;
8641         u64 dev_nr = 0;
8642         u64 target;
8643         int index;
8644         int full = 0;
8645         int ret = 0;
8646
8647         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8648
8649         /* odd, couldn't find the block group, leave it alone */
8650         if (!block_group)
8651                 return -1;
8652
8653         min_free = btrfs_block_group_used(&block_group->item);
8654
8655         /* no bytes used, we're good */
8656         if (!min_free)
8657                 goto out;
8658
8659         space_info = block_group->space_info;
8660         spin_lock(&space_info->lock);
8661
8662         full = space_info->full;
8663
8664         /*
8665          * if this is the last block group we have in this space, we can't
8666          * relocate it unless we're able to allocate a new chunk below.
8667          *
8668          * Otherwise, we need to make sure we have room in the space to handle
8669          * all of the extents from this block group.  If we can, we're good
8670          */
8671         if ((space_info->total_bytes != block_group->key.offset) &&
8672             (space_info->bytes_used + space_info->bytes_reserved +
8673              space_info->bytes_pinned + space_info->bytes_readonly +
8674              min_free < space_info->total_bytes)) {
8675                 spin_unlock(&space_info->lock);
8676                 goto out;
8677         }
8678         spin_unlock(&space_info->lock);
8679
8680         /*
8681          * ok we don't have enough space, but maybe we have free space on our
8682          * devices to allocate new chunks for relocation, so loop through our
8683          * alloc devices and guess if we have enough space.  if this block
8684          * group is going to be restriped, run checks against the target
8685          * profile instead of the current one.
8686          */
8687         ret = -1;
8688
8689         /*
8690          * index:
8691          *      0: raid10
8692          *      1: raid1
8693          *      2: dup
8694          *      3: raid0
8695          *      4: single
8696          */
8697         target = get_restripe_target(root->fs_info, block_group->flags);
8698         if (target) {
8699                 index = __get_raid_index(extended_to_chunk(target));
8700         } else {
8701                 /*
8702                  * this is just a balance, so if we were marked as full
8703                  * we know there is no space for a new chunk
8704                  */
8705                 if (full)
8706                         goto out;
8707
8708                 index = get_block_group_index(block_group);
8709         }
8710
8711         if (index == BTRFS_RAID_RAID10) {
8712                 dev_min = 4;
8713                 /* Divide by 2 */
8714                 min_free >>= 1;
8715         } else if (index == BTRFS_RAID_RAID1) {
8716                 dev_min = 2;
8717         } else if (index == BTRFS_RAID_DUP) {
8718                 /* Multiply by 2 */
8719                 min_free <<= 1;
8720         } else if (index == BTRFS_RAID_RAID0) {
8721                 dev_min = fs_devices->rw_devices;
8722                 min_free = div64_u64(min_free, dev_min);
8723         }
8724
8725         /* We need to do this so that we can look at pending chunks */
8726         trans = btrfs_join_transaction(root);
8727         if (IS_ERR(trans)) {
8728                 ret = PTR_ERR(trans);
8729                 goto out;
8730         }
8731
8732         mutex_lock(&root->fs_info->chunk_mutex);
8733         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8734                 u64 dev_offset;
8735
8736                 /*
8737                  * check to make sure we can actually find a chunk with enough
8738                  * space to fit our block group in.
8739                  */
8740                 if (device->total_bytes > device->bytes_used + min_free &&
8741                     !device->is_tgtdev_for_dev_replace) {
8742                         ret = find_free_dev_extent(trans, device, min_free,
8743                                                    &dev_offset, NULL);
8744                         if (!ret)
8745                                 dev_nr++;
8746
8747                         if (dev_nr >= dev_min)
8748                                 break;
8749
8750                         ret = -1;
8751                 }
8752         }
8753         mutex_unlock(&root->fs_info->chunk_mutex);
8754         btrfs_end_transaction(trans, root);
8755 out:
8756         btrfs_put_block_group(block_group);
8757         return ret;
8758 }
8759
8760 static int find_first_block_group(struct btrfs_root *root,
8761                 struct btrfs_path *path, struct btrfs_key *key)
8762 {
8763         int ret = 0;
8764         struct btrfs_key found_key;
8765         struct extent_buffer *leaf;
8766         int slot;
8767
8768         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8769         if (ret < 0)
8770                 goto out;
8771
8772         while (1) {
8773                 slot = path->slots[0];
8774                 leaf = path->nodes[0];
8775                 if (slot >= btrfs_header_nritems(leaf)) {
8776                         ret = btrfs_next_leaf(root, path);
8777                         if (ret == 0)
8778                                 continue;
8779                         if (ret < 0)
8780                                 goto out;
8781                         break;
8782                 }
8783                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8784
8785                 if (found_key.objectid >= key->objectid &&
8786                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8787                         ret = 0;
8788                         goto out;
8789                 }
8790                 path->slots[0]++;
8791         }
8792 out:
8793         return ret;
8794 }
8795
8796 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8797 {
8798         struct btrfs_block_group_cache *block_group;
8799         u64 last = 0;
8800
8801         while (1) {
8802                 struct inode *inode;
8803
8804                 block_group = btrfs_lookup_first_block_group(info, last);
8805                 while (block_group) {
8806                         spin_lock(&block_group->lock);
8807                         if (block_group->iref)
8808                                 break;
8809                         spin_unlock(&block_group->lock);
8810                         block_group = next_block_group(info->tree_root,
8811                                                        block_group);
8812                 }
8813                 if (!block_group) {
8814                         if (last == 0)
8815                                 break;
8816                         last = 0;
8817                         continue;
8818                 }
8819
8820                 inode = block_group->inode;
8821                 block_group->iref = 0;
8822                 block_group->inode = NULL;
8823                 spin_unlock(&block_group->lock);
8824                 iput(inode);
8825                 last = block_group->key.objectid + block_group->key.offset;
8826                 btrfs_put_block_group(block_group);
8827         }
8828 }
8829
8830 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8831 {
8832         struct btrfs_block_group_cache *block_group;
8833         struct btrfs_space_info *space_info;
8834         struct btrfs_caching_control *caching_ctl;
8835         struct rb_node *n;
8836
8837         down_write(&info->commit_root_sem);
8838         while (!list_empty(&info->caching_block_groups)) {
8839                 caching_ctl = list_entry(info->caching_block_groups.next,
8840                                          struct btrfs_caching_control, list);
8841                 list_del(&caching_ctl->list);
8842                 put_caching_control(caching_ctl);
8843         }
8844         up_write(&info->commit_root_sem);
8845
8846         spin_lock(&info->unused_bgs_lock);
8847         while (!list_empty(&info->unused_bgs)) {
8848                 block_group = list_first_entry(&info->unused_bgs,
8849                                                struct btrfs_block_group_cache,
8850                                                bg_list);
8851                 list_del_init(&block_group->bg_list);
8852                 btrfs_put_block_group(block_group);
8853         }
8854         spin_unlock(&info->unused_bgs_lock);
8855
8856         spin_lock(&info->block_group_cache_lock);
8857         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8858                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8859                                        cache_node);
8860                 rb_erase(&block_group->cache_node,
8861                          &info->block_group_cache_tree);
8862                 RB_CLEAR_NODE(&block_group->cache_node);
8863                 spin_unlock(&info->block_group_cache_lock);
8864
8865                 down_write(&block_group->space_info->groups_sem);
8866                 list_del(&block_group->list);
8867                 up_write(&block_group->space_info->groups_sem);
8868
8869                 if (block_group->cached == BTRFS_CACHE_STARTED)
8870                         wait_block_group_cache_done(block_group);
8871
8872                 /*
8873                  * We haven't cached this block group, which means we could
8874                  * possibly have excluded extents on this block group.
8875                  */
8876                 if (block_group->cached == BTRFS_CACHE_NO ||
8877                     block_group->cached == BTRFS_CACHE_ERROR)
8878                         free_excluded_extents(info->extent_root, block_group);
8879
8880                 btrfs_remove_free_space_cache(block_group);
8881                 btrfs_put_block_group(block_group);
8882
8883                 spin_lock(&info->block_group_cache_lock);
8884         }
8885         spin_unlock(&info->block_group_cache_lock);
8886
8887         /* now that all the block groups are freed, go through and
8888          * free all the space_info structs.  This is only called during
8889          * the final stages of unmount, and so we know nobody is
8890          * using them.  We call synchronize_rcu() once before we start,
8891          * just to be on the safe side.
8892          */
8893         synchronize_rcu();
8894
8895         release_global_block_rsv(info);
8896
8897         while (!list_empty(&info->space_info)) {
8898                 int i;
8899
8900                 space_info = list_entry(info->space_info.next,
8901                                         struct btrfs_space_info,
8902                                         list);
8903                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8904                         if (WARN_ON(space_info->bytes_pinned > 0 ||
8905                             space_info->bytes_reserved > 0 ||
8906                             space_info->bytes_may_use > 0)) {
8907                                 dump_space_info(space_info, 0, 0);
8908                         }
8909                 }
8910                 list_del(&space_info->list);
8911                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
8912                         struct kobject *kobj;
8913                         kobj = space_info->block_group_kobjs[i];
8914                         space_info->block_group_kobjs[i] = NULL;
8915                         if (kobj) {
8916                                 kobject_del(kobj);
8917                                 kobject_put(kobj);
8918                         }
8919                 }
8920                 kobject_del(&space_info->kobj);
8921                 kobject_put(&space_info->kobj);
8922         }
8923         return 0;
8924 }
8925
8926 static void __link_block_group(struct btrfs_space_info *space_info,
8927                                struct btrfs_block_group_cache *cache)
8928 {
8929         int index = get_block_group_index(cache);
8930         bool first = false;
8931
8932         down_write(&space_info->groups_sem);
8933         if (list_empty(&space_info->block_groups[index]))
8934                 first = true;
8935         list_add_tail(&cache->list, &space_info->block_groups[index]);
8936         up_write(&space_info->groups_sem);
8937
8938         if (first) {
8939                 struct raid_kobject *rkobj;
8940                 int ret;
8941
8942                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
8943                 if (!rkobj)
8944                         goto out_err;
8945                 rkobj->raid_type = index;
8946                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
8947                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
8948                                   "%s", get_raid_name(index));
8949                 if (ret) {
8950                         kobject_put(&rkobj->kobj);
8951                         goto out_err;
8952                 }
8953                 space_info->block_group_kobjs[index] = &rkobj->kobj;
8954         }
8955
8956         return;
8957 out_err:
8958         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
8959 }
8960
8961 static struct btrfs_block_group_cache *
8962 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
8963 {
8964         struct btrfs_block_group_cache *cache;
8965
8966         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8967         if (!cache)
8968                 return NULL;
8969
8970         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8971                                         GFP_NOFS);
8972         if (!cache->free_space_ctl) {
8973                 kfree(cache);
8974                 return NULL;
8975         }
8976
8977         cache->key.objectid = start;
8978         cache->key.offset = size;
8979         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8980
8981         cache->sectorsize = root->sectorsize;
8982         cache->fs_info = root->fs_info;
8983         cache->full_stripe_len = btrfs_full_stripe_len(root,
8984                                                &root->fs_info->mapping_tree,
8985                                                start);
8986         atomic_set(&cache->count, 1);
8987         spin_lock_init(&cache->lock);
8988         init_rwsem(&cache->data_rwsem);
8989         INIT_LIST_HEAD(&cache->list);
8990         INIT_LIST_HEAD(&cache->cluster_list);
8991         INIT_LIST_HEAD(&cache->bg_list);
8992         INIT_LIST_HEAD(&cache->ro_list);
8993         INIT_LIST_HEAD(&cache->dirty_list);
8994         btrfs_init_free_space_ctl(cache);
8995         atomic_set(&cache->trimming, 0);
8996
8997         return cache;
8998 }
8999
9000 int btrfs_read_block_groups(struct btrfs_root *root)
9001 {
9002         struct btrfs_path *path;
9003         int ret;
9004         struct btrfs_block_group_cache *cache;
9005         struct btrfs_fs_info *info = root->fs_info;
9006         struct btrfs_space_info *space_info;
9007         struct btrfs_key key;
9008         struct btrfs_key found_key;
9009         struct extent_buffer *leaf;
9010         int need_clear = 0;
9011         u64 cache_gen;
9012
9013         root = info->extent_root;
9014         key.objectid = 0;
9015         key.offset = 0;
9016         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9017         path = btrfs_alloc_path();
9018         if (!path)
9019                 return -ENOMEM;
9020         path->reada = 1;
9021
9022         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9023         if (btrfs_test_opt(root, SPACE_CACHE) &&
9024             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9025                 need_clear = 1;
9026         if (btrfs_test_opt(root, CLEAR_CACHE))
9027                 need_clear = 1;
9028
9029         while (1) {
9030                 ret = find_first_block_group(root, path, &key);
9031                 if (ret > 0)
9032                         break;
9033                 if (ret != 0)
9034                         goto error;
9035
9036                 leaf = path->nodes[0];
9037                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9038
9039                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9040                                                        found_key.offset);
9041                 if (!cache) {
9042                         ret = -ENOMEM;
9043                         goto error;
9044                 }
9045
9046                 if (need_clear) {
9047                         /*
9048                          * When we mount with old space cache, we need to
9049                          * set BTRFS_DC_CLEAR and set dirty flag.
9050                          *
9051                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9052                          *    truncate the old free space cache inode and
9053                          *    setup a new one.
9054                          * b) Setting 'dirty flag' makes sure that we flush
9055                          *    the new space cache info onto disk.
9056                          */
9057                         if (btrfs_test_opt(root, SPACE_CACHE))
9058                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9059                 }
9060
9061                 read_extent_buffer(leaf, &cache->item,
9062                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9063                                    sizeof(cache->item));
9064                 cache->flags = btrfs_block_group_flags(&cache->item);
9065
9066                 key.objectid = found_key.objectid + found_key.offset;
9067                 btrfs_release_path(path);
9068
9069                 /*
9070                  * We need to exclude the super stripes now so that the space
9071                  * info has super bytes accounted for, otherwise we'll think
9072                  * we have more space than we actually do.
9073                  */
9074                 ret = exclude_super_stripes(root, cache);
9075                 if (ret) {
9076                         /*
9077                          * We may have excluded something, so call this just in
9078                          * case.
9079                          */
9080                         free_excluded_extents(root, cache);
9081                         btrfs_put_block_group(cache);
9082                         goto error;
9083                 }
9084
9085                 /*
9086                  * check for two cases, either we are full, and therefore
9087                  * don't need to bother with the caching work since we won't
9088                  * find any space, or we are empty, and we can just add all
9089                  * the space in and be done with it.  This saves us _alot_ of
9090                  * time, particularly in the full case.
9091                  */
9092                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9093                         cache->last_byte_to_unpin = (u64)-1;
9094                         cache->cached = BTRFS_CACHE_FINISHED;
9095                         free_excluded_extents(root, cache);
9096                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9097                         cache->last_byte_to_unpin = (u64)-1;
9098                         cache->cached = BTRFS_CACHE_FINISHED;
9099                         add_new_free_space(cache, root->fs_info,
9100                                            found_key.objectid,
9101                                            found_key.objectid +
9102                                            found_key.offset);
9103                         free_excluded_extents(root, cache);
9104                 }
9105
9106                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9107                 if (ret) {
9108                         btrfs_remove_free_space_cache(cache);
9109                         btrfs_put_block_group(cache);
9110                         goto error;
9111                 }
9112
9113                 ret = update_space_info(info, cache->flags, found_key.offset,
9114                                         btrfs_block_group_used(&cache->item),
9115                                         &space_info);
9116                 if (ret) {
9117                         btrfs_remove_free_space_cache(cache);
9118                         spin_lock(&info->block_group_cache_lock);
9119                         rb_erase(&cache->cache_node,
9120                                  &info->block_group_cache_tree);
9121                         RB_CLEAR_NODE(&cache->cache_node);
9122                         spin_unlock(&info->block_group_cache_lock);
9123                         btrfs_put_block_group(cache);
9124                         goto error;
9125                 }
9126
9127                 cache->space_info = space_info;
9128                 spin_lock(&cache->space_info->lock);
9129                 cache->space_info->bytes_readonly += cache->bytes_super;
9130                 spin_unlock(&cache->space_info->lock);
9131
9132                 __link_block_group(space_info, cache);
9133
9134                 set_avail_alloc_bits(root->fs_info, cache->flags);
9135                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9136                         set_block_group_ro(cache, 1);
9137                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9138                         spin_lock(&info->unused_bgs_lock);
9139                         /* Should always be true but just in case. */
9140                         if (list_empty(&cache->bg_list)) {
9141                                 btrfs_get_block_group(cache);
9142                                 list_add_tail(&cache->bg_list,
9143                                               &info->unused_bgs);
9144                         }
9145                         spin_unlock(&info->unused_bgs_lock);
9146                 }
9147         }
9148
9149         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9150                 if (!(get_alloc_profile(root, space_info->flags) &
9151                       (BTRFS_BLOCK_GROUP_RAID10 |
9152                        BTRFS_BLOCK_GROUP_RAID1 |
9153                        BTRFS_BLOCK_GROUP_RAID5 |
9154                        BTRFS_BLOCK_GROUP_RAID6 |
9155                        BTRFS_BLOCK_GROUP_DUP)))
9156                         continue;
9157                 /*
9158                  * avoid allocating from un-mirrored block group if there are
9159                  * mirrored block groups.
9160                  */
9161                 list_for_each_entry(cache,
9162                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9163                                 list)
9164                         set_block_group_ro(cache, 1);
9165                 list_for_each_entry(cache,
9166                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9167                                 list)
9168                         set_block_group_ro(cache, 1);
9169         }
9170
9171         init_global_block_rsv(info);
9172         ret = 0;
9173 error:
9174         btrfs_free_path(path);
9175         return ret;
9176 }
9177
9178 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9179                                        struct btrfs_root *root)
9180 {
9181         struct btrfs_block_group_cache *block_group, *tmp;
9182         struct btrfs_root *extent_root = root->fs_info->extent_root;
9183         struct btrfs_block_group_item item;
9184         struct btrfs_key key;
9185         int ret = 0;
9186
9187         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9188                 if (ret)
9189                         goto next;
9190
9191                 spin_lock(&block_group->lock);
9192                 memcpy(&item, &block_group->item, sizeof(item));
9193                 memcpy(&key, &block_group->key, sizeof(key));
9194                 spin_unlock(&block_group->lock);
9195
9196                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9197                                         sizeof(item));
9198                 if (ret)
9199                         btrfs_abort_transaction(trans, extent_root, ret);
9200                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9201                                                key.objectid, key.offset);
9202                 if (ret)
9203                         btrfs_abort_transaction(trans, extent_root, ret);
9204 next:
9205                 list_del_init(&block_group->bg_list);
9206         }
9207 }
9208
9209 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9210                            struct btrfs_root *root, u64 bytes_used,
9211                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9212                            u64 size)
9213 {
9214         int ret;
9215         struct btrfs_root *extent_root;
9216         struct btrfs_block_group_cache *cache;
9217
9218         extent_root = root->fs_info->extent_root;
9219
9220         btrfs_set_log_full_commit(root->fs_info, trans);
9221
9222         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9223         if (!cache)
9224                 return -ENOMEM;
9225
9226         btrfs_set_block_group_used(&cache->item, bytes_used);
9227         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9228         btrfs_set_block_group_flags(&cache->item, type);
9229
9230         cache->flags = type;
9231         cache->last_byte_to_unpin = (u64)-1;
9232         cache->cached = BTRFS_CACHE_FINISHED;
9233         ret = exclude_super_stripes(root, cache);
9234         if (ret) {
9235                 /*
9236                  * We may have excluded something, so call this just in
9237                  * case.
9238                  */
9239                 free_excluded_extents(root, cache);
9240                 btrfs_put_block_group(cache);
9241                 return ret;
9242         }
9243
9244         add_new_free_space(cache, root->fs_info, chunk_offset,
9245                            chunk_offset + size);
9246
9247         free_excluded_extents(root, cache);
9248
9249         ret = btrfs_add_block_group_cache(root->fs_info, cache);
9250         if (ret) {
9251                 btrfs_remove_free_space_cache(cache);
9252                 btrfs_put_block_group(cache);
9253                 return ret;
9254         }
9255
9256         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
9257                                 &cache->space_info);
9258         if (ret) {
9259                 btrfs_remove_free_space_cache(cache);
9260                 spin_lock(&root->fs_info->block_group_cache_lock);
9261                 rb_erase(&cache->cache_node,
9262                          &root->fs_info->block_group_cache_tree);
9263                 RB_CLEAR_NODE(&cache->cache_node);
9264                 spin_unlock(&root->fs_info->block_group_cache_lock);
9265                 btrfs_put_block_group(cache);
9266                 return ret;
9267         }
9268         update_global_block_rsv(root->fs_info);
9269
9270         spin_lock(&cache->space_info->lock);
9271         cache->space_info->bytes_readonly += cache->bytes_super;
9272         spin_unlock(&cache->space_info->lock);
9273
9274         __link_block_group(cache->space_info, cache);
9275
9276         list_add_tail(&cache->bg_list, &trans->new_bgs);
9277
9278         set_avail_alloc_bits(extent_root->fs_info, type);
9279
9280         return 0;
9281 }
9282
9283 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9284 {
9285         u64 extra_flags = chunk_to_extended(flags) &
9286                                 BTRFS_EXTENDED_PROFILE_MASK;
9287
9288         write_seqlock(&fs_info->profiles_lock);
9289         if (flags & BTRFS_BLOCK_GROUP_DATA)
9290                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9291         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9292                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9293         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9294                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9295         write_sequnlock(&fs_info->profiles_lock);
9296 }
9297
9298 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9299                              struct btrfs_root *root, u64 group_start,
9300                              struct extent_map *em)
9301 {
9302         struct btrfs_path *path;
9303         struct btrfs_block_group_cache *block_group;
9304         struct btrfs_free_cluster *cluster;
9305         struct btrfs_root *tree_root = root->fs_info->tree_root;
9306         struct btrfs_key key;
9307         struct inode *inode;
9308         struct kobject *kobj = NULL;
9309         int ret;
9310         int index;
9311         int factor;
9312         struct btrfs_caching_control *caching_ctl = NULL;
9313         bool remove_em;
9314
9315         root = root->fs_info->extent_root;
9316
9317         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9318         BUG_ON(!block_group);
9319         BUG_ON(!block_group->ro);
9320
9321         /*
9322          * Free the reserved super bytes from this block group before
9323          * remove it.
9324          */
9325         free_excluded_extents(root, block_group);
9326
9327         memcpy(&key, &block_group->key, sizeof(key));
9328         index = get_block_group_index(block_group);
9329         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9330                                   BTRFS_BLOCK_GROUP_RAID1 |
9331                                   BTRFS_BLOCK_GROUP_RAID10))
9332                 factor = 2;
9333         else
9334                 factor = 1;
9335
9336         /* make sure this block group isn't part of an allocation cluster */
9337         cluster = &root->fs_info->data_alloc_cluster;
9338         spin_lock(&cluster->refill_lock);
9339         btrfs_return_cluster_to_free_space(block_group, cluster);
9340         spin_unlock(&cluster->refill_lock);
9341
9342         /*
9343          * make sure this block group isn't part of a metadata
9344          * allocation cluster
9345          */
9346         cluster = &root->fs_info->meta_alloc_cluster;
9347         spin_lock(&cluster->refill_lock);
9348         btrfs_return_cluster_to_free_space(block_group, cluster);
9349         spin_unlock(&cluster->refill_lock);
9350
9351         path = btrfs_alloc_path();
9352         if (!path) {
9353                 ret = -ENOMEM;
9354                 goto out;
9355         }
9356
9357         inode = lookup_free_space_inode(tree_root, block_group, path);
9358         if (!IS_ERR(inode)) {
9359                 ret = btrfs_orphan_add(trans, inode);
9360                 if (ret) {
9361                         btrfs_add_delayed_iput(inode);
9362                         goto out;
9363                 }
9364                 clear_nlink(inode);
9365                 /* One for the block groups ref */
9366                 spin_lock(&block_group->lock);
9367                 if (block_group->iref) {
9368                         block_group->iref = 0;
9369                         block_group->inode = NULL;
9370                         spin_unlock(&block_group->lock);
9371                         iput(inode);
9372                 } else {
9373                         spin_unlock(&block_group->lock);
9374                 }
9375                 /* One for our lookup ref */
9376                 btrfs_add_delayed_iput(inode);
9377         }
9378
9379         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9380         key.offset = block_group->key.objectid;
9381         key.type = 0;
9382
9383         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9384         if (ret < 0)
9385                 goto out;
9386         if (ret > 0)
9387                 btrfs_release_path(path);
9388         if (ret == 0) {
9389                 ret = btrfs_del_item(trans, tree_root, path);
9390                 if (ret)
9391                         goto out;
9392                 btrfs_release_path(path);
9393         }
9394
9395         spin_lock(&root->fs_info->block_group_cache_lock);
9396         rb_erase(&block_group->cache_node,
9397                  &root->fs_info->block_group_cache_tree);
9398         RB_CLEAR_NODE(&block_group->cache_node);
9399
9400         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9401                 root->fs_info->first_logical_byte = (u64)-1;
9402         spin_unlock(&root->fs_info->block_group_cache_lock);
9403
9404         down_write(&block_group->space_info->groups_sem);
9405         /*
9406          * we must use list_del_init so people can check to see if they
9407          * are still on the list after taking the semaphore
9408          */
9409         list_del_init(&block_group->list);
9410         if (list_empty(&block_group->space_info->block_groups[index])) {
9411                 kobj = block_group->space_info->block_group_kobjs[index];
9412                 block_group->space_info->block_group_kobjs[index] = NULL;
9413                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9414         }
9415         up_write(&block_group->space_info->groups_sem);
9416         if (kobj) {
9417                 kobject_del(kobj);
9418                 kobject_put(kobj);
9419         }
9420
9421         if (block_group->has_caching_ctl)
9422                 caching_ctl = get_caching_control(block_group);
9423         if (block_group->cached == BTRFS_CACHE_STARTED)
9424                 wait_block_group_cache_done(block_group);
9425         if (block_group->has_caching_ctl) {
9426                 down_write(&root->fs_info->commit_root_sem);
9427                 if (!caching_ctl) {
9428                         struct btrfs_caching_control *ctl;
9429
9430                         list_for_each_entry(ctl,
9431                                     &root->fs_info->caching_block_groups, list)
9432                                 if (ctl->block_group == block_group) {
9433                                         caching_ctl = ctl;
9434                                         atomic_inc(&caching_ctl->count);
9435                                         break;
9436                                 }
9437                 }
9438                 if (caching_ctl)
9439                         list_del_init(&caching_ctl->list);
9440                 up_write(&root->fs_info->commit_root_sem);
9441                 if (caching_ctl) {
9442                         /* Once for the caching bgs list and once for us. */
9443                         put_caching_control(caching_ctl);
9444                         put_caching_control(caching_ctl);
9445                 }
9446         }
9447
9448         spin_lock(&trans->transaction->dirty_bgs_lock);
9449         if (!list_empty(&block_group->dirty_list)) {
9450                 list_del_init(&block_group->dirty_list);
9451                 btrfs_put_block_group(block_group);
9452         }
9453         spin_unlock(&trans->transaction->dirty_bgs_lock);
9454
9455         btrfs_remove_free_space_cache(block_group);
9456
9457         spin_lock(&block_group->space_info->lock);
9458         list_del_init(&block_group->ro_list);
9459         block_group->space_info->total_bytes -= block_group->key.offset;
9460         block_group->space_info->bytes_readonly -= block_group->key.offset;
9461         block_group->space_info->disk_total -= block_group->key.offset * factor;
9462         spin_unlock(&block_group->space_info->lock);
9463
9464         memcpy(&key, &block_group->key, sizeof(key));
9465
9466         lock_chunks(root);
9467         if (!list_empty(&em->list)) {
9468                 /* We're in the transaction->pending_chunks list. */
9469                 free_extent_map(em);
9470         }
9471         spin_lock(&block_group->lock);
9472         block_group->removed = 1;
9473         /*
9474          * At this point trimming can't start on this block group, because we
9475          * removed the block group from the tree fs_info->block_group_cache_tree
9476          * so no one can't find it anymore and even if someone already got this
9477          * block group before we removed it from the rbtree, they have already
9478          * incremented block_group->trimming - if they didn't, they won't find
9479          * any free space entries because we already removed them all when we
9480          * called btrfs_remove_free_space_cache().
9481          *
9482          * And we must not remove the extent map from the fs_info->mapping_tree
9483          * to prevent the same logical address range and physical device space
9484          * ranges from being reused for a new block group. This is because our
9485          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
9486          * completely transactionless, so while it is trimming a range the
9487          * currently running transaction might finish and a new one start,
9488          * allowing for new block groups to be created that can reuse the same
9489          * physical device locations unless we take this special care.
9490          */
9491         remove_em = (atomic_read(&block_group->trimming) == 0);
9492         /*
9493          * Make sure a trimmer task always sees the em in the pinned_chunks list
9494          * if it sees block_group->removed == 1 (needs to lock block_group->lock
9495          * before checking block_group->removed).
9496          */
9497         if (!remove_em) {
9498                 /*
9499                  * Our em might be in trans->transaction->pending_chunks which
9500                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
9501                  * and so is the fs_info->pinned_chunks list.
9502                  *
9503                  * So at this point we must be holding the chunk_mutex to avoid
9504                  * any races with chunk allocation (more specifically at
9505                  * volumes.c:contains_pending_extent()), to ensure it always
9506                  * sees the em, either in the pending_chunks list or in the
9507                  * pinned_chunks list.
9508                  */
9509                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
9510         }
9511         spin_unlock(&block_group->lock);
9512
9513         if (remove_em) {
9514                 struct extent_map_tree *em_tree;
9515
9516                 em_tree = &root->fs_info->mapping_tree.map_tree;
9517                 write_lock(&em_tree->lock);
9518                 /*
9519                  * The em might be in the pending_chunks list, so make sure the
9520                  * chunk mutex is locked, since remove_extent_mapping() will
9521                  * delete us from that list.
9522                  */
9523                 remove_extent_mapping(em_tree, em);
9524                 write_unlock(&em_tree->lock);
9525                 /* once for the tree */
9526                 free_extent_map(em);
9527         }
9528
9529         unlock_chunks(root);
9530
9531         btrfs_put_block_group(block_group);
9532         btrfs_put_block_group(block_group);
9533
9534         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9535         if (ret > 0)
9536                 ret = -EIO;
9537         if (ret < 0)
9538                 goto out;
9539
9540         ret = btrfs_del_item(trans, root, path);
9541 out:
9542         btrfs_free_path(path);
9543         return ret;
9544 }
9545
9546 /*
9547  * Process the unused_bgs list and remove any that don't have any allocated
9548  * space inside of them.
9549  */
9550 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
9551 {
9552         struct btrfs_block_group_cache *block_group;
9553         struct btrfs_space_info *space_info;
9554         struct btrfs_root *root = fs_info->extent_root;
9555         struct btrfs_trans_handle *trans;
9556         int ret = 0;
9557
9558         if (!fs_info->open)
9559                 return;
9560
9561         spin_lock(&fs_info->unused_bgs_lock);
9562         while (!list_empty(&fs_info->unused_bgs)) {
9563                 u64 start, end;
9564
9565                 block_group = list_first_entry(&fs_info->unused_bgs,
9566                                                struct btrfs_block_group_cache,
9567                                                bg_list);
9568                 space_info = block_group->space_info;
9569                 list_del_init(&block_group->bg_list);
9570                 if (ret || btrfs_mixed_space_info(space_info)) {
9571                         btrfs_put_block_group(block_group);
9572                         continue;
9573                 }
9574                 spin_unlock(&fs_info->unused_bgs_lock);
9575
9576                 /* Don't want to race with allocators so take the groups_sem */
9577                 down_write(&space_info->groups_sem);
9578                 spin_lock(&block_group->lock);
9579                 if (block_group->reserved ||
9580                     btrfs_block_group_used(&block_group->item) ||
9581                     block_group->ro) {
9582                         /*
9583                          * We want to bail if we made new allocations or have
9584                          * outstanding allocations in this block group.  We do
9585                          * the ro check in case balance is currently acting on
9586                          * this block group.
9587                          */
9588                         spin_unlock(&block_group->lock);
9589                         up_write(&space_info->groups_sem);
9590                         goto next;
9591                 }
9592                 spin_unlock(&block_group->lock);
9593
9594                 /* We don't want to force the issue, only flip if it's ok. */
9595                 ret = set_block_group_ro(block_group, 0);
9596                 up_write(&space_info->groups_sem);
9597                 if (ret < 0) {
9598                         ret = 0;
9599                         goto next;
9600                 }
9601
9602                 /*
9603                  * Want to do this before we do anything else so we can recover
9604                  * properly if we fail to join the transaction.
9605                  */
9606                 /* 1 for btrfs_orphan_reserve_metadata() */
9607                 trans = btrfs_start_transaction(root, 1);
9608                 if (IS_ERR(trans)) {
9609                         btrfs_set_block_group_rw(root, block_group);
9610                         ret = PTR_ERR(trans);
9611                         goto next;
9612                 }
9613
9614                 /*
9615                  * We could have pending pinned extents for this block group,
9616                  * just delete them, we don't care about them anymore.
9617                  */
9618                 start = block_group->key.objectid;
9619                 end = start + block_group->key.offset - 1;
9620                 /*
9621                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
9622                  * btrfs_finish_extent_commit(). If we are at transaction N,
9623                  * another task might be running finish_extent_commit() for the
9624                  * previous transaction N - 1, and have seen a range belonging
9625                  * to the block group in freed_extents[] before we were able to
9626                  * clear the whole block group range from freed_extents[]. This
9627                  * means that task can lookup for the block group after we
9628                  * unpinned it from freed_extents[] and removed it, leading to
9629                  * a BUG_ON() at btrfs_unpin_extent_range().
9630                  */
9631                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
9632                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
9633                                   EXTENT_DIRTY, GFP_NOFS);
9634                 if (ret) {
9635                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9636                         btrfs_set_block_group_rw(root, block_group);
9637                         goto end_trans;
9638                 }
9639                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
9640                                   EXTENT_DIRTY, GFP_NOFS);
9641                 if (ret) {
9642                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9643                         btrfs_set_block_group_rw(root, block_group);
9644                         goto end_trans;
9645                 }
9646                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9647
9648                 /* Reset pinned so btrfs_put_block_group doesn't complain */
9649                 block_group->pinned = 0;
9650
9651                 /*
9652                  * Btrfs_remove_chunk will abort the transaction if things go
9653                  * horribly wrong.
9654                  */
9655                 ret = btrfs_remove_chunk(trans, root,
9656                                          block_group->key.objectid);
9657 end_trans:
9658                 btrfs_end_transaction(trans, root);
9659 next:
9660                 btrfs_put_block_group(block_group);
9661                 spin_lock(&fs_info->unused_bgs_lock);
9662         }
9663         spin_unlock(&fs_info->unused_bgs_lock);
9664 }
9665
9666 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
9667 {
9668         struct btrfs_space_info *space_info;
9669         struct btrfs_super_block *disk_super;
9670         u64 features;
9671         u64 flags;
9672         int mixed = 0;
9673         int ret;
9674
9675         disk_super = fs_info->super_copy;
9676         if (!btrfs_super_root(disk_super))
9677                 return 1;
9678
9679         features = btrfs_super_incompat_flags(disk_super);
9680         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
9681                 mixed = 1;
9682
9683         flags = BTRFS_BLOCK_GROUP_SYSTEM;
9684         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9685         if (ret)
9686                 goto out;
9687
9688         if (mixed) {
9689                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
9690                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9691         } else {
9692                 flags = BTRFS_BLOCK_GROUP_METADATA;
9693                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9694                 if (ret)
9695                         goto out;
9696
9697                 flags = BTRFS_BLOCK_GROUP_DATA;
9698                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9699         }
9700 out:
9701         return ret;
9702 }
9703
9704 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
9705 {
9706         return unpin_extent_range(root, start, end, false);
9707 }
9708
9709 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
9710 {
9711         struct btrfs_fs_info *fs_info = root->fs_info;
9712         struct btrfs_block_group_cache *cache = NULL;
9713         u64 group_trimmed;
9714         u64 start;
9715         u64 end;
9716         u64 trimmed = 0;
9717         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
9718         int ret = 0;
9719
9720         /*
9721          * try to trim all FS space, our block group may start from non-zero.
9722          */
9723         if (range->len == total_bytes)
9724                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
9725         else
9726                 cache = btrfs_lookup_block_group(fs_info, range->start);
9727
9728         while (cache) {
9729                 if (cache->key.objectid >= (range->start + range->len)) {
9730                         btrfs_put_block_group(cache);
9731                         break;
9732                 }
9733
9734                 start = max(range->start, cache->key.objectid);
9735                 end = min(range->start + range->len,
9736                                 cache->key.objectid + cache->key.offset);
9737
9738                 if (end - start >= range->minlen) {
9739                         if (!block_group_cache_done(cache)) {
9740                                 ret = cache_block_group(cache, 0);
9741                                 if (ret) {
9742                                         btrfs_put_block_group(cache);
9743                                         break;
9744                                 }
9745                                 ret = wait_block_group_cache_done(cache);
9746                                 if (ret) {
9747                                         btrfs_put_block_group(cache);
9748                                         break;
9749                                 }
9750                         }
9751                         ret = btrfs_trim_block_group(cache,
9752                                                      &group_trimmed,
9753                                                      start,
9754                                                      end,
9755                                                      range->minlen);
9756
9757                         trimmed += group_trimmed;
9758                         if (ret) {
9759                                 btrfs_put_block_group(cache);
9760                                 break;
9761                         }
9762                 }
9763
9764                 cache = next_block_group(fs_info->tree_root, cache);
9765         }
9766
9767         range->len = trimmed;
9768         return ret;
9769 }
9770
9771 /*
9772  * btrfs_{start,end}_write_no_snapshoting() are similar to
9773  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
9774  * data into the page cache through nocow before the subvolume is snapshoted,
9775  * but flush the data into disk after the snapshot creation, or to prevent
9776  * operations while snapshoting is ongoing and that cause the snapshot to be
9777  * inconsistent (writes followed by expanding truncates for example).
9778  */
9779 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
9780 {
9781         percpu_counter_dec(&root->subv_writers->counter);
9782         /*
9783          * Make sure counter is updated before we wake up
9784          * waiters.
9785          */
9786         smp_mb();
9787         if (waitqueue_active(&root->subv_writers->wait))
9788                 wake_up(&root->subv_writers->wait);
9789 }
9790
9791 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
9792 {
9793         if (atomic_read(&root->will_be_snapshoted))
9794                 return 0;
9795
9796         percpu_counter_inc(&root->subv_writers->counter);
9797         /*
9798          * Make sure counter is updated before we check for snapshot creation.
9799          */
9800         smp_mb();
9801         if (atomic_read(&root->will_be_snapshoted)) {
9802                 btrfs_end_write_no_snapshoting(root);
9803                 return 0;
9804         }
9805         return 1;
9806 }