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