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                                 struct btrfs_delayed_ref_node *node, 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 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  87                                     struct extent_buffer *leaf,
  88                                     struct btrfs_extent_item *ei);
  89 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  90                                       struct btrfs_root *root,
  91                                       u64 parent, u64 root_objectid,
  92                                       u64 flags, u64 owner, u64 offset,
  93                                       struct btrfs_key *ins, int ref_mod);
  94 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  95                                      struct btrfs_root *root,
  96                                      u64 parent, u64 root_objectid,
  97                                      u64 flags, struct btrfs_disk_key *key,
  98                                      int level, struct btrfs_key *ins,
  99                                      int no_quota);
 100 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
 101                           struct btrfs_root *extent_root, u64 flags,
 102                           int force);
 103 static int find_next_key(struct btrfs_path *path, int level,
 104                          struct btrfs_key *key);
 105 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 106                             int dump_block_groups);
 107 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 108                                        u64 num_bytes, int reserve,
 109                                        int delalloc);
 110 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 111                                u64 num_bytes);
 112 int btrfs_pin_extent(struct btrfs_root *root,
 113                      u64 bytenr, u64 num_bytes, int reserved);
 114
 115 static noinline int
 116 block_group_cache_done(struct btrfs_block_group_cache *cache)
 117 {
 118         smp_mb();
 119         return cache->cached == BTRFS_CACHE_FINISHED ||
 120                 cache->cached == BTRFS_CACHE_ERROR;
 121 }
 122
 123 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 124 {
 125         return (cache->flags & bits) == bits;
 126 }
 127
 128 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 129 {
 130         atomic_inc(&cache->count);
 131 }
 132
 133 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 134 {
 135         if (atomic_dec_and_test(&cache->count)) {
 136                 WARN_ON(cache->pinned > 0);
 137                 WARN_ON(cache->reserved > 0);
 138                 kfree(cache->free_space_ctl);
 139                 kfree(cache);
 140         }
 141 }
 142
 143 /*
 144  * this adds the block group to the fs_info rb tree for the block group
 145  * cache
 146  */
 147 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 148                                 struct btrfs_block_group_cache *block_group)
 149 {
 150         struct rb_node **p;
 151         struct rb_node *parent = NULL;
 152         struct btrfs_block_group_cache *cache;
 153
 154         spin_lock(&info->block_group_cache_lock);
 155         p = &info->block_group_cache_tree.rb_node;
 156
 157         while (*p) {
 158                 parent = *p;
 159                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 160                                  cache_node);
 161                 if (block_group->key.objectid < cache->key.objectid) {
 162                         p = &(*p)->rb_left;
 163                 } else if (block_group->key.objectid > cache->key.objectid) {
 164                         p = &(*p)->rb_right;
 165                 } else {
 166                         spin_unlock(&info->block_group_cache_lock);
 167                         return -EEXIST;
 168                 }
 169         }
 170
 171         rb_link_node(&block_group->cache_node, parent, p);
 172         rb_insert_color(&block_group->cache_node,
 173                         &info->block_group_cache_tree);
 174
 175         if (info->first_logical_byte > block_group->key.objectid)
 176                 info->first_logical_byte = block_group->key.objectid;
 177
 178         spin_unlock(&info->block_group_cache_lock);
 179
 180         return 0;
 181 }
 182
 183 /*
 184  * This will return the block group at or after bytenr if contains is 0, else
 185  * it will return the block group that contains the bytenr
 186  */
 187 static struct btrfs_block_group_cache *
 188 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 189                               int contains)
 190 {
 191         struct btrfs_block_group_cache *cache, *ret = NULL;
 192         struct rb_node *n;
 193         u64 end, start;
 194
 195         spin_lock(&info->block_group_cache_lock);
 196         n = info->block_group_cache_tree.rb_node;
 197
 198         while (n) {
 199                 cache = rb_entry(n, struct btrfs_block_group_cache,
 200                                  cache_node);
 201                 end = cache->key.objectid + cache->key.offset - 1;
 202                 start = cache->key.objectid;
 203
 204                 if (bytenr < start) {
 205                         if (!contains && (!ret || start < ret->key.objectid))
 206                                 ret = cache;
 207                         n = n->rb_left;
 208                 } else if (bytenr > start) {
 209                         if (contains && bytenr <= end) {
 210                                 ret = cache;
 211                                 break;
 212                         }
 213                         n = n->rb_right;
 214                 } else {
 215                         ret = cache;
 216                         break;
 217                 }
 218         }
 219         if (ret) {
 220                 btrfs_get_block_group(ret);
 221                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 222                         info->first_logical_byte = ret->key.objectid;
 223         }
 224         spin_unlock(&info->block_group_cache_lock);
 225
 226         return ret;
 227 }
 228
 229 static int add_excluded_extent(struct btrfs_root *root,
 230                                u64 start, u64 num_bytes)
 231 {
 232         u64 end = start + num_bytes - 1;
 233         set_extent_bits(&root->fs_info->freed_extents[0],
 234                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 235         set_extent_bits(&root->fs_info->freed_extents[1],
 236                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 237         return 0;
 238 }
 239
 240 static void free_excluded_extents(struct btrfs_root *root,
 241                                   struct btrfs_block_group_cache *cache)
 242 {
 243         u64 start, end;
 244
 245         start = cache->key.objectid;
 246         end = start + cache->key.offset - 1;
 247
 248         clear_extent_bits(&root->fs_info->freed_extents[0],
 249                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 250         clear_extent_bits(&root->fs_info->freed_extents[1],
 251                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 252 }
 253
 254 static int exclude_super_stripes(struct btrfs_root *root,
 255                                  struct btrfs_block_group_cache *cache)
 256 {
 257         u64 bytenr;
 258         u64 *logical;
 259         int stripe_len;
 260         int i, nr, ret;
 261
 262         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 263                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 264                 cache->bytes_super += stripe_len;
 265                 ret = add_excluded_extent(root, cache->key.objectid,
 266                                           stripe_len);
 267                 if (ret)
 268                         return ret;
 269         }
 270
 271         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 272                 bytenr = btrfs_sb_offset(i);
 273                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 274                                        cache->key.objectid, bytenr,
 275                                        0, &logical, &nr, &stripe_len);
 276                 if (ret)
 277                         return ret;
 278
 279                 while (nr--) {
 280                         u64 start, len;
 281
 282                         if (logical[nr] > cache->key.objectid +
 283                             cache->key.offset)
 284                                 continue;
 285
 286                         if (logical[nr] + stripe_len <= cache->key.objectid)
 287                                 continue;
 288
 289                         start = logical[nr];
 290                         if (start < cache->key.objectid) {
 291                                 start = cache->key.objectid;
 292                                 len = (logical[nr] + stripe_len) - start;
 293                         } else {
 294                                 len = min_t(u64, stripe_len,
 295                                             cache->key.objectid +
 296                                             cache->key.offset - start);
 297                         }
 298
 299                         cache->bytes_super += len;
 300                         ret = add_excluded_extent(root, start, len);
 301                         if (ret) {
 302                                 kfree(logical);
 303                                 return ret;
 304                         }
 305                 }
 306
 307                 kfree(logical);
 308         }
 309         return 0;
 310 }
 311
 312 static struct btrfs_caching_control *
 313 get_caching_control(struct btrfs_block_group_cache *cache)
 314 {
 315         struct btrfs_caching_control *ctl;
 316
 317         spin_lock(&cache->lock);
 318         if (!cache->caching_ctl) {
 319                 spin_unlock(&cache->lock);
 320                 return NULL;
 321         }
 322
 323         ctl = cache->caching_ctl;
 324         atomic_inc(&ctl->count);
 325         spin_unlock(&cache->lock);
 326         return ctl;
 327 }
 328
 329 static void put_caching_control(struct btrfs_caching_control *ctl)
 330 {
 331         if (atomic_dec_and_test(&ctl->count))
 332                 kfree(ctl);
 333 }
 334
 335 /*
 336  * this is only called by cache_block_group, since we could have freed extents
 337  * we need to check the pinned_extents for any extents that can't be used yet
 338  * since their free space will be released as soon as the transaction commits.
 339  */
 340 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 341                               struct btrfs_fs_info *info, u64 start, u64 end)
 342 {
 343         u64 extent_start, extent_end, size, total_added = 0;
 344         int ret;
 345
 346         while (start < end) {
 347                 ret = find_first_extent_bit(info->pinned_extents, start,
 348                                             &extent_start, &extent_end,
 349                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 350                                             NULL);
 351                 if (ret)
 352                         break;
 353
 354                 if (extent_start <= start) {
 355                         start = extent_end + 1;
 356                 } else if (extent_start > start && extent_start < end) {
 357                         size = extent_start - start;
 358                         total_added += size;
 359                         ret = btrfs_add_free_space(block_group, start,
 360                                                    size);
 361                         BUG_ON(ret); /* -ENOMEM or logic error */
 362                         start = extent_end + 1;
 363                 } else {
 364                         break;
 365                 }
 366         }
 367
 368         if (start < end) {
 369                 size = end - start;
 370                 total_added += size;
 371                 ret = btrfs_add_free_space(block_group, start, size);
 372                 BUG_ON(ret); /* -ENOMEM or logic error */
 373         }
 374
 375         return total_added;
 376 }
 377
 378 static noinline void caching_thread(struct btrfs_work *work)
 379 {
 380         struct btrfs_block_group_cache *block_group;
 381         struct btrfs_fs_info *fs_info;
 382         struct btrfs_caching_control *caching_ctl;
 383         struct btrfs_root *extent_root;
 384         struct btrfs_path *path;
 385         struct extent_buffer *leaf;
 386         struct btrfs_key key;
 387         u64 total_found = 0;
 388         u64 last = 0;
 389         u32 nritems;
 390         int ret = -ENOMEM;
 391
 392         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 393         block_group = caching_ctl->block_group;
 394         fs_info = block_group->fs_info;
 395         extent_root = fs_info->extent_root;
 396
 397         path = btrfs_alloc_path();
 398         if (!path)
 399                 goto out;
 400
 401         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 402
 403         /*
 404          * We don't want to deadlock with somebody trying to allocate a new
 405          * extent for the extent root while also trying to search the extent
 406          * root to add free space.  So we skip locking and search the commit
 407          * root, since its read-only
 408          */
 409         path->skip_locking = 1;
 410         path->search_commit_root = 1;
 411         path->reada = 1;
 412
 413         key.objectid = last;
 414         key.offset = 0;
 415         key.type = BTRFS_EXTENT_ITEM_KEY;
 416 again:
 417         mutex_lock(&caching_ctl->mutex);
 418         /* need to make sure the commit_root doesn't disappear */
 419         down_read(&fs_info->commit_root_sem);
 420
 421 next:
 422         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 423         if (ret < 0)
 424                 goto err;
 425
 426         leaf = path->nodes[0];
 427         nritems = btrfs_header_nritems(leaf);
 428
 429         while (1) {
 430                 if (btrfs_fs_closing(fs_info) > 1) {
 431                         last = (u64)-1;
 432                         break;
 433                 }
 434
 435                 if (path->slots[0] < nritems) {
 436                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 437                 } else {
 438                         ret = find_next_key(path, 0, &key);
 439                         if (ret)
 440                                 break;
 441
 442                         if (need_resched() ||
 443                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 444                                 caching_ctl->progress = last;
 445                                 btrfs_release_path(path);
 446                                 up_read(&fs_info->commit_root_sem);
 447                                 mutex_unlock(&caching_ctl->mutex);
 448                                 cond_resched();
 449                                 goto again;
 450                         }
 451
 452                         ret = btrfs_next_leaf(extent_root, path);
 453                         if (ret < 0)
 454                                 goto err;
 455                         if (ret)
 456                                 break;
 457                         leaf = path->nodes[0];
 458                         nritems = btrfs_header_nritems(leaf);
 459                         continue;
 460                 }
 461
 462                 if (key.objectid < last) {
 463                         key.objectid = last;
 464                         key.offset = 0;
 465                         key.type = BTRFS_EXTENT_ITEM_KEY;
 466
 467                         caching_ctl->progress = last;
 468                         btrfs_release_path(path);
 469                         goto next;
 470                 }
 471
 472                 if (key.objectid < block_group->key.objectid) {
 473                         path->slots[0]++;
 474                         continue;
 475                 }
 476
 477                 if (key.objectid >= block_group->key.objectid +
 478                     block_group->key.offset)
 479                         break;
 480
 481                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 482                     key.type == BTRFS_METADATA_ITEM_KEY) {
 483                         total_found += add_new_free_space(block_group,
 484                                                           fs_info, last,
 485                                                           key.objectid);
 486                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 487                                 last = key.objectid +
 488                                         fs_info->tree_root->nodesize;
 489                         else
 490                                 last = key.objectid + key.offset;
 491
 492                         if (total_found > (1024 * 1024 * 2)) {
 493                                 total_found = 0;
 494                                 wake_up(&caching_ctl->wait);
 495                         }
 496                 }
 497                 path->slots[0]++;
 498         }
 499         ret = 0;
 500
 501         total_found += add_new_free_space(block_group, fs_info, last,
 502                                           block_group->key.objectid +
 503                                           block_group->key.offset);
 504         caching_ctl->progress = (u64)-1;
 505
 506         spin_lock(&block_group->lock);
 507         block_group->caching_ctl = NULL;
 508         block_group->cached = BTRFS_CACHE_FINISHED;
 509         spin_unlock(&block_group->lock);
 510
 511 err:
 512         btrfs_free_path(path);
 513         up_read(&fs_info->commit_root_sem);
 514
 515         free_excluded_extents(extent_root, block_group);
 516
 517         mutex_unlock(&caching_ctl->mutex);
 518 out:
 519         if (ret) {
 520                 spin_lock(&block_group->lock);
 521                 block_group->caching_ctl = NULL;
 522                 block_group->cached = BTRFS_CACHE_ERROR;
 523                 spin_unlock(&block_group->lock);
 524         }
 525         wake_up(&caching_ctl->wait);
 526
 527         put_caching_control(caching_ctl);
 528         btrfs_put_block_group(block_group);
 529 }
 530
 531 static int cache_block_group(struct btrfs_block_group_cache *cache,
 532                              int load_cache_only)
 533 {
 534         DEFINE_WAIT(wait);
 535         struct btrfs_fs_info *fs_info = cache->fs_info;
 536         struct btrfs_caching_control *caching_ctl;
 537         int ret = 0;
 538
 539         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 540         if (!caching_ctl)
 541                 return -ENOMEM;
 542
 543         INIT_LIST_HEAD(&caching_ctl->list);
 544         mutex_init(&caching_ctl->mutex);
 545         init_waitqueue_head(&caching_ctl->wait);
 546         caching_ctl->block_group = cache;
 547         caching_ctl->progress = cache->key.objectid;
 548         atomic_set(&caching_ctl->count, 1);
 549         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 550                         caching_thread, NULL, NULL);
 551
 552         spin_lock(&cache->lock);
 553         /*
 554          * This should be a rare occasion, but this could happen I think in the
 555          * case where one thread starts to load the space cache info, and then
 556          * some other thread starts a transaction commit which tries to do an
 557          * allocation while the other thread is still loading the space cache
 558          * info.  The previous loop should have kept us from choosing this block
 559          * group, but if we've moved to the state where we will wait on caching
 560          * block groups we need to first check if we're doing a fast load here,
 561          * so we can wait for it to finish, otherwise we could end up allocating
 562          * from a block group who's cache gets evicted for one reason or
 563          * another.
 564          */
 565         while (cache->cached == BTRFS_CACHE_FAST) {
 566                 struct btrfs_caching_control *ctl;
 567
 568                 ctl = cache->caching_ctl;
 569                 atomic_inc(&ctl->count);
 570                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 571                 spin_unlock(&cache->lock);
 572
 573                 schedule();
 574
 575                 finish_wait(&ctl->wait, &wait);
 576                 put_caching_control(ctl);
 577                 spin_lock(&cache->lock);
 578         }
 579
 580         if (cache->cached != BTRFS_CACHE_NO) {
 581                 spin_unlock(&cache->lock);
 582                 kfree(caching_ctl);
 583                 return 0;
 584         }
 585         WARN_ON(cache->caching_ctl);
 586         cache->caching_ctl = caching_ctl;
 587         cache->cached = BTRFS_CACHE_FAST;
 588         spin_unlock(&cache->lock);
 589
 590         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 591                 mutex_lock(&caching_ctl->mutex);
 592                 ret = load_free_space_cache(fs_info, cache);
 593
 594                 spin_lock(&cache->lock);
 595                 if (ret == 1) {
 596                         cache->caching_ctl = NULL;
 597                         cache->cached = BTRFS_CACHE_FINISHED;
 598                         cache->last_byte_to_unpin = (u64)-1;
 599                         caching_ctl->progress = (u64)-1;
 600                 } else {
 601                         if (load_cache_only) {
 602                                 cache->caching_ctl = NULL;
 603                                 cache->cached = BTRFS_CACHE_NO;
 604                         } else {
 605                                 cache->cached = BTRFS_CACHE_STARTED;
 606                                 cache->has_caching_ctl = 1;
 607                         }
 608                 }
 609                 spin_unlock(&cache->lock);
 610                 mutex_unlock(&caching_ctl->mutex);
 611
 612                 wake_up(&caching_ctl->wait);
 613                 if (ret == 1) {
 614                         put_caching_control(caching_ctl);
 615                         free_excluded_extents(fs_info->extent_root, cache);
 616                         return 0;
 617                 }
 618         } else {
 619                 /*
 620                  * We are not going to do the fast caching, set cached to the
 621                  * appropriate value and wakeup any waiters.
 622                  */
 623                 spin_lock(&cache->lock);
 624                 if (load_cache_only) {
 625                         cache->caching_ctl = NULL;
 626                         cache->cached = BTRFS_CACHE_NO;
 627                 } else {
 628                         cache->cached = BTRFS_CACHE_STARTED;
 629                         cache->has_caching_ctl = 1;
 630                 }
 631                 spin_unlock(&cache->lock);
 632                 wake_up(&caching_ctl->wait);
 633         }
 634
 635         if (load_cache_only) {
 636                 put_caching_control(caching_ctl);
 637                 return 0;
 638         }
 639
 640         down_write(&fs_info->commit_root_sem);
 641         atomic_inc(&caching_ctl->count);
 642         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 643         up_write(&fs_info->commit_root_sem);
 644
 645         btrfs_get_block_group(cache);
 646
 647         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 648
 649         return ret;
 650 }
 651
 652 /*
 653  * return the block group that starts at or after bytenr
 654  */
 655 static struct btrfs_block_group_cache *
 656 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 657 {
 658         struct btrfs_block_group_cache *cache;
 659
 660         cache = block_group_cache_tree_search(info, bytenr, 0);
 661
 662         return cache;
 663 }
 664
 665 /*
 666  * return the block group that contains the given bytenr
 667  */
 668 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 669                                                  struct btrfs_fs_info *info,
 670                                                  u64 bytenr)
 671 {
 672         struct btrfs_block_group_cache *cache;
 673
 674         cache = block_group_cache_tree_search(info, bytenr, 1);
 675
 676         return cache;
 677 }
 678
 679 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 680                                                   u64 flags)
 681 {
 682         struct list_head *head = &info->space_info;
 683         struct btrfs_space_info *found;
 684
 685         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 686
 687         rcu_read_lock();
 688         list_for_each_entry_rcu(found, head, list) {
 689                 if (found->flags & flags) {
 690                         rcu_read_unlock();
 691                         return found;
 692                 }
 693         }
 694         rcu_read_unlock();
 695         return NULL;
 696 }
 697
 698 /*
 699  * after adding space to the filesystem, we need to clear the full flags
 700  * on all the space infos.
 701  */
 702 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 703 {
 704         struct list_head *head = &info->space_info;
 705         struct btrfs_space_info *found;
 706
 707         rcu_read_lock();
 708         list_for_each_entry_rcu(found, head, list)
 709                 found->full = 0;
 710         rcu_read_unlock();
 711 }
 712
 713 /* simple helper to search for an existing data extent at a given offset */
 714 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 715 {
 716         int ret;
 717         struct btrfs_key key;
 718         struct btrfs_path *path;
 719
 720         path = btrfs_alloc_path();
 721         if (!path)
 722                 return -ENOMEM;
 723
 724         key.objectid = start;
 725         key.offset = len;
 726         key.type = BTRFS_EXTENT_ITEM_KEY;
 727         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 728                                 0, 0);
 729         btrfs_free_path(path);
 730         return ret;
 731 }
 732
 733 /*
 734  * helper function to lookup reference count and flags of a tree block.
 735  *
 736  * the head node for delayed ref is used to store the sum of all the
 737  * reference count modifications queued up in the rbtree. the head
 738  * node may also store the extent flags to set. This way you can check
 739  * to see what the reference count and extent flags would be if all of
 740  * the delayed refs are not processed.
 741  */
 742 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 743                              struct btrfs_root *root, u64 bytenr,
 744                              u64 offset, int metadata, u64 *refs, u64 *flags)
 745 {
 746         struct btrfs_delayed_ref_head *head;
 747         struct btrfs_delayed_ref_root *delayed_refs;
 748         struct btrfs_path *path;
 749         struct btrfs_extent_item *ei;
 750         struct extent_buffer *leaf;
 751         struct btrfs_key key;
 752         u32 item_size;
 753         u64 num_refs;
 754         u64 extent_flags;
 755         int ret;
 756
 757         /*
 758          * If we don't have skinny metadata, don't bother doing anything
 759          * different
 760          */
 761         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 762                 offset = root->nodesize;
 763                 metadata = 0;
 764         }
 765
 766         path = btrfs_alloc_path();
 767         if (!path)
 768                 return -ENOMEM;
 769
 770         if (!trans) {
 771                 path->skip_locking = 1;
 772                 path->search_commit_root = 1;
 773         }
 774
 775 search_again:
 776         key.objectid = bytenr;
 777         key.offset = offset;
 778         if (metadata)
 779                 key.type = BTRFS_METADATA_ITEM_KEY;
 780         else
 781                 key.type = BTRFS_EXTENT_ITEM_KEY;
 782
 783         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 784                                 &key, path, 0, 0);
 785         if (ret < 0)
 786                 goto out_free;
 787
 788         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 789                 if (path->slots[0]) {
 790                         path->slots[0]--;
 791                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 792                                               path->slots[0]);
 793                         if (key.objectid == bytenr &&
 794                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 795                             key.offset == root->nodesize)
 796                                 ret = 0;
 797                 }
 798         }
 799
 800         if (ret == 0) {
 801                 leaf = path->nodes[0];
 802                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 803                 if (item_size >= sizeof(*ei)) {
 804                         ei = btrfs_item_ptr(leaf, path->slots[0],
 805                                             struct btrfs_extent_item);
 806                         num_refs = btrfs_extent_refs(leaf, ei);
 807                         extent_flags = btrfs_extent_flags(leaf, ei);
 808                 } else {
 809 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 810                         struct btrfs_extent_item_v0 *ei0;
 811                         BUG_ON(item_size != sizeof(*ei0));
 812                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 813                                              struct btrfs_extent_item_v0);
 814                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 815                         /* FIXME: this isn't correct for data */
 816                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 817 #else
 818                         BUG();
 819 #endif
 820                 }
 821                 BUG_ON(num_refs == 0);
 822         } else {
 823                 num_refs = 0;
 824                 extent_flags = 0;
 825                 ret = 0;
 826         }
 827
 828         if (!trans)
 829                 goto out;
 830
 831         delayed_refs = &trans->transaction->delayed_refs;
 832         spin_lock(&delayed_refs->lock);
 833         head = btrfs_find_delayed_ref_head(trans, bytenr);
 834         if (head) {
 835                 if (!mutex_trylock(&head->mutex)) {
 836                         atomic_inc(&head->node.refs);
 837                         spin_unlock(&delayed_refs->lock);
 838
 839                         btrfs_release_path(path);
 840
 841                         /*
 842                          * Mutex was contended, block until it's released and try
 843                          * again
 844                          */
 845                         mutex_lock(&head->mutex);
 846                         mutex_unlock(&head->mutex);
 847                         btrfs_put_delayed_ref(&head->node);
 848                         goto search_again;
 849                 }
 850                 spin_lock(&head->lock);
 851                 if (head->extent_op && head->extent_op->update_flags)
 852                         extent_flags |= head->extent_op->flags_to_set;
 853                 else
 854                         BUG_ON(num_refs == 0);
 855
 856                 num_refs += head->node.ref_mod;
 857                 spin_unlock(&head->lock);
 858                 mutex_unlock(&head->mutex);
 859         }
 860         spin_unlock(&delayed_refs->lock);
 861 out:
 862         WARN_ON(num_refs == 0);
 863         if (refs)
 864                 *refs = num_refs;
 865         if (flags)
 866                 *flags = extent_flags;
 867 out_free:
 868         btrfs_free_path(path);
 869         return ret;
 870 }
 871
 872 /*
 873  * Back reference rules.  Back refs have three main goals:
 874  *
 875  * 1) differentiate between all holders of references to an extent so that
 876  *    when a reference is dropped we can make sure it was a valid reference
 877  *    before freeing the extent.
 878  *
 879  * 2) Provide enough information to quickly find the holders of an extent
 880  *    if we notice a given block is corrupted or bad.
 881  *
 882  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 883  *    maintenance.  This is actually the same as #2, but with a slightly
 884  *    different use case.
 885  *
 886  * There are two kinds of back refs. The implicit back refs is optimized
 887  * for pointers in non-shared tree blocks. For a given pointer in a block,
 888  * back refs of this kind provide information about the block's owner tree
 889  * and the pointer's key. These information allow us to find the block by
 890  * b-tree searching. The full back refs is for pointers in tree blocks not
 891  * referenced by their owner trees. The location of tree block is recorded
 892  * in the back refs. Actually the full back refs is generic, and can be
 893  * used in all cases the implicit back refs is used. The major shortcoming
 894  * of the full back refs is its overhead. Every time a tree block gets
 895  * COWed, we have to update back refs entry for all pointers in it.
 896  *
 897  * For a newly allocated tree block, we use implicit back refs for
 898  * pointers in it. This means most tree related operations only involve
 899  * implicit back refs. For a tree block created in old transaction, the
 900  * only way to drop a reference to it is COW it. So we can detect the
 901  * event that tree block loses its owner tree's reference and do the
 902  * back refs conversion.
 903  *
 904  * When a tree block is COW'd through a tree, there are four cases:
 905  *
 906  * The reference count of the block is one and the tree is the block's
 907  * owner tree. Nothing to do in this case.
 908  *
 909  * The reference count of the block is one and the tree is not the
 910  * block's owner tree. In this case, full back refs is used for pointers
 911  * in the block. Remove these full back refs, add implicit back refs for
 912  * every pointers in the new block.
 913  *
 914  * The reference count of the block is greater than one and the tree is
 915  * the block's owner tree. In this case, implicit back refs is used for
 916  * pointers in the block. Add full back refs for every pointers in the
 917  * block, increase lower level extents' reference counts. The original
 918  * implicit back refs are entailed to the new block.
 919  *
 920  * The reference count of the block is greater than one and the tree is
 921  * not the block's owner tree. Add implicit back refs for every pointer in
 922  * the new block, increase lower level extents' reference count.
 923  *
 924  * Back Reference Key composing:
 925  *
 926  * The key objectid corresponds to the first byte in the extent,
 927  * The key type is used to differentiate between types of back refs.
 928  * There are different meanings of the key offset for different types
 929  * of back refs.
 930  *
 931  * File extents can be referenced by:
 932  *
 933  * - multiple snapshots, subvolumes, or different generations in one subvol
 934  * - different files inside a single subvolume
 935  * - different offsets inside a file (bookend extents in file.c)
 936  *
 937  * The extent ref structure for the implicit back refs has fields for:
 938  *
 939  * - Objectid of the subvolume root
 940  * - objectid of the file holding the reference
 941  * - original offset in the file
 942  * - how many bookend extents
 943  *
 944  * The key offset for the implicit back refs is hash of the first
 945  * three fields.
 946  *
 947  * The extent ref structure for the full back refs has field for:
 948  *
 949  * - number of pointers in the tree leaf
 950  *
 951  * The key offset for the implicit back refs is the first byte of
 952  * the tree leaf
 953  *
 954  * When a file extent is allocated, The implicit back refs is used.
 955  * the fields are filled in:
 956  *
 957  *     (root_key.objectid, inode objectid, offset in file, 1)
 958  *
 959  * When a file extent is removed file truncation, we find the
 960  * corresponding implicit back refs and check the following fields:
 961  *
 962  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 963  *
 964  * Btree extents can be referenced by:
 965  *
 966  * - Different subvolumes
 967  *
 968  * Both the implicit back refs and the full back refs for tree blocks
 969  * only consist of key. The key offset for the implicit back refs is
 970  * objectid of block's owner tree. The key offset for the full back refs
 971  * is the first byte of parent block.
 972  *
 973  * When implicit back refs is used, information about the lowest key and
 974  * level of the tree block are required. These information are stored in
 975  * tree block info structure.
 976  */
 977
 978 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 979 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 980                                   struct btrfs_root *root,
 981                                   struct btrfs_path *path,
 982                                   u64 owner, u32 extra_size)
 983 {
 984         struct btrfs_extent_item *item;
 985         struct btrfs_extent_item_v0 *ei0;
 986         struct btrfs_extent_ref_v0 *ref0;
 987         struct btrfs_tree_block_info *bi;
 988         struct extent_buffer *leaf;
 989         struct btrfs_key key;
 990         struct btrfs_key found_key;
 991         u32 new_size = sizeof(*item);
 992         u64 refs;
 993         int ret;
 994
 995         leaf = path->nodes[0];
 996         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 997
 998         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 999         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1000                              struct btrfs_extent_item_v0);
1001         refs = btrfs_extent_refs_v0(leaf, ei0);
1002
1003         if (owner == (u64)-1) {
1004                 while (1) {
1005                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1006                                 ret = btrfs_next_leaf(root, path);
1007                                 if (ret < 0)
1008                                         return ret;
1009                                 BUG_ON(ret > 0); /* Corruption */
1010                                 leaf = path->nodes[0];
1011                         }
1012                         btrfs_item_key_to_cpu(leaf, &found_key,
1013                                               path->slots[0]);
1014                         BUG_ON(key.objectid != found_key.objectid);
1015                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1016                                 path->slots[0]++;
1017                                 continue;
1018                         }
1019                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1020                                               struct btrfs_extent_ref_v0);
1021                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1022                         break;
1023                 }
1024         }
1025         btrfs_release_path(path);
1026
1027         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1028                 new_size += sizeof(*bi);
1029
1030         new_size -= sizeof(*ei0);
1031         ret = btrfs_search_slot(trans, root, &key, path,
1032                                 new_size + extra_size, 1);
1033         if (ret < 0)
1034                 return ret;
1035         BUG_ON(ret); /* Corruption */
1036
1037         btrfs_extend_item(root, path, new_size);
1038
1039         leaf = path->nodes[0];
1040         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1041         btrfs_set_extent_refs(leaf, item, refs);
1042         /* FIXME: get real generation */
1043         btrfs_set_extent_generation(leaf, item, 0);
1044         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1045                 btrfs_set_extent_flags(leaf, item,
1046                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1047                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1048                 bi = (struct btrfs_tree_block_info *)(item + 1);
1049                 /* FIXME: get first key of the block */
1050                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1051                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1052         } else {
1053                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1054         }
1055         btrfs_mark_buffer_dirty(leaf);
1056         return 0;
1057 }
1058 #endif
1059
1060 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1061 {
1062         u32 high_crc = ~(u32)0;
1063         u32 low_crc = ~(u32)0;
1064         __le64 lenum;
1065
1066         lenum = cpu_to_le64(root_objectid);
1067         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1068         lenum = cpu_to_le64(owner);
1069         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1070         lenum = cpu_to_le64(offset);
1071         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1072
1073         return ((u64)high_crc << 31) ^ (u64)low_crc;
1074 }
1075
1076 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1077                                      struct btrfs_extent_data_ref *ref)
1078 {
1079         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1080                                     btrfs_extent_data_ref_objectid(leaf, ref),
1081                                     btrfs_extent_data_ref_offset(leaf, ref));
1082 }
1083
1084 static int match_extent_data_ref(struct extent_buffer *leaf,
1085                                  struct btrfs_extent_data_ref *ref,
1086                                  u64 root_objectid, u64 owner, u64 offset)
1087 {
1088         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1089             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1090             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1091                 return 0;
1092         return 1;
1093 }
1094
1095 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1096                                            struct btrfs_root *root,
1097                                            struct btrfs_path *path,
1098                                            u64 bytenr, u64 parent,
1099                                            u64 root_objectid,
1100                                            u64 owner, u64 offset)
1101 {
1102         struct btrfs_key key;
1103         struct btrfs_extent_data_ref *ref;
1104         struct extent_buffer *leaf;
1105         u32 nritems;
1106         int ret;
1107         int recow;
1108         int err = -ENOENT;
1109
1110         key.objectid = bytenr;
1111         if (parent) {
1112                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1113                 key.offset = parent;
1114         } else {
1115                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1116                 key.offset = hash_extent_data_ref(root_objectid,
1117                                                   owner, offset);
1118         }
1119 again:
1120         recow = 0;
1121         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1122         if (ret < 0) {
1123                 err = ret;
1124                 goto fail;
1125         }
1126
1127         if (parent) {
1128                 if (!ret)
1129                         return 0;
1130 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1131                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1132                 btrfs_release_path(path);
1133                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1134                 if (ret < 0) {
1135                         err = ret;
1136                         goto fail;
1137                 }
1138                 if (!ret)
1139                         return 0;
1140 #endif
1141                 goto fail;
1142         }
1143
1144         leaf = path->nodes[0];
1145         nritems = btrfs_header_nritems(leaf);
1146         while (1) {
1147                 if (path->slots[0] >= nritems) {
1148                         ret = btrfs_next_leaf(root, path);
1149                         if (ret < 0)
1150                                 err = ret;
1151                         if (ret)
1152                                 goto fail;
1153
1154                         leaf = path->nodes[0];
1155                         nritems = btrfs_header_nritems(leaf);
1156                         recow = 1;
1157                 }
1158
1159                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1160                 if (key.objectid != bytenr ||
1161                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1162                         goto fail;
1163
1164                 ref = btrfs_item_ptr(leaf, path->slots[0],
1165                                      struct btrfs_extent_data_ref);
1166
1167                 if (match_extent_data_ref(leaf, ref, root_objectid,
1168                                           owner, offset)) {
1169                         if (recow) {
1170                                 btrfs_release_path(path);
1171                                 goto again;
1172                         }
1173                         err = 0;
1174                         break;
1175                 }
1176                 path->slots[0]++;
1177         }
1178 fail:
1179         return err;
1180 }
1181
1182 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1183                                            struct btrfs_root *root,
1184                                            struct btrfs_path *path,
1185                                            u64 bytenr, u64 parent,
1186                                            u64 root_objectid, u64 owner,
1187                                            u64 offset, int refs_to_add)
1188 {
1189         struct btrfs_key key;
1190         struct extent_buffer *leaf;
1191         u32 size;
1192         u32 num_refs;
1193         int ret;
1194
1195         key.objectid = bytenr;
1196         if (parent) {
1197                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1198                 key.offset = parent;
1199                 size = sizeof(struct btrfs_shared_data_ref);
1200         } else {
1201                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1202                 key.offset = hash_extent_data_ref(root_objectid,
1203                                                   owner, offset);
1204                 size = sizeof(struct btrfs_extent_data_ref);
1205         }
1206
1207         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1208         if (ret && ret != -EEXIST)
1209                 goto fail;
1210
1211         leaf = path->nodes[0];
1212         if (parent) {
1213                 struct btrfs_shared_data_ref *ref;
1214                 ref = btrfs_item_ptr(leaf, path->slots[0],
1215                                      struct btrfs_shared_data_ref);
1216                 if (ret == 0) {
1217                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1218                 } else {
1219                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1220                         num_refs += refs_to_add;
1221                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1222                 }
1223         } else {
1224                 struct btrfs_extent_data_ref *ref;
1225                 while (ret == -EEXIST) {
1226                         ref = btrfs_item_ptr(leaf, path->slots[0],
1227                                              struct btrfs_extent_data_ref);
1228                         if (match_extent_data_ref(leaf, ref, root_objectid,
1229                                                   owner, offset))
1230                                 break;
1231                         btrfs_release_path(path);
1232                         key.offset++;
1233                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1234                                                       size);
1235                         if (ret && ret != -EEXIST)
1236                                 goto fail;
1237
1238                         leaf = path->nodes[0];
1239                 }
1240                 ref = btrfs_item_ptr(leaf, path->slots[0],
1241                                      struct btrfs_extent_data_ref);
1242                 if (ret == 0) {
1243                         btrfs_set_extent_data_ref_root(leaf, ref,
1244                                                        root_objectid);
1245                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1246                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1247                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1248                 } else {
1249                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1250                         num_refs += refs_to_add;
1251                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1252                 }
1253         }
1254         btrfs_mark_buffer_dirty(leaf);
1255         ret = 0;
1256 fail:
1257         btrfs_release_path(path);
1258         return ret;
1259 }
1260
1261 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1262                                            struct btrfs_root *root,
1263                                            struct btrfs_path *path,
1264                                            int refs_to_drop, int *last_ref)
1265 {
1266         struct btrfs_key key;
1267         struct btrfs_extent_data_ref *ref1 = NULL;
1268         struct btrfs_shared_data_ref *ref2 = NULL;
1269         struct extent_buffer *leaf;
1270         u32 num_refs = 0;
1271         int ret = 0;
1272
1273         leaf = path->nodes[0];
1274         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1275
1276         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1277                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1278                                       struct btrfs_extent_data_ref);
1279                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1280         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1281                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1282                                       struct btrfs_shared_data_ref);
1283                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1284 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1285         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1286                 struct btrfs_extent_ref_v0 *ref0;
1287                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1288                                       struct btrfs_extent_ref_v0);
1289                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1290 #endif
1291         } else {
1292                 BUG();
1293         }
1294
1295         BUG_ON(num_refs < refs_to_drop);
1296         num_refs -= refs_to_drop;
1297
1298         if (num_refs == 0) {
1299                 ret = btrfs_del_item(trans, root, path);
1300                 *last_ref = 1;
1301         } else {
1302                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1303                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1304                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1305                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1306 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1307                 else {
1308                         struct btrfs_extent_ref_v0 *ref0;
1309                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1310                                         struct btrfs_extent_ref_v0);
1311                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1312                 }
1313 #endif
1314                 btrfs_mark_buffer_dirty(leaf);
1315         }
1316         return ret;
1317 }
1318
1319 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1320                                           struct btrfs_extent_inline_ref *iref)
1321 {
1322         struct btrfs_key key;
1323         struct extent_buffer *leaf;
1324         struct btrfs_extent_data_ref *ref1;
1325         struct btrfs_shared_data_ref *ref2;
1326         u32 num_refs = 0;
1327
1328         leaf = path->nodes[0];
1329         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1330         if (iref) {
1331                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1332                     BTRFS_EXTENT_DATA_REF_KEY) {
1333                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1334                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1335                 } else {
1336                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1337                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1338                 }
1339         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1340                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1341                                       struct btrfs_extent_data_ref);
1342                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1343         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1344                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1345                                       struct btrfs_shared_data_ref);
1346                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1347 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1348         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1349                 struct btrfs_extent_ref_v0 *ref0;
1350                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1351                                       struct btrfs_extent_ref_v0);
1352                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1353 #endif
1354         } else {
1355                 WARN_ON(1);
1356         }
1357         return num_refs;
1358 }
1359
1360 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1361                                           struct btrfs_root *root,
1362                                           struct btrfs_path *path,
1363                                           u64 bytenr, u64 parent,
1364                                           u64 root_objectid)
1365 {
1366         struct btrfs_key key;
1367         int ret;
1368
1369         key.objectid = bytenr;
1370         if (parent) {
1371                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1372                 key.offset = parent;
1373         } else {
1374                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1375                 key.offset = root_objectid;
1376         }
1377
1378         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1379         if (ret > 0)
1380                 ret = -ENOENT;
1381 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1382         if (ret == -ENOENT && parent) {
1383                 btrfs_release_path(path);
1384                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1385                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1386                 if (ret > 0)
1387                         ret = -ENOENT;
1388         }
1389 #endif
1390         return ret;
1391 }
1392
1393 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1394                                           struct btrfs_root *root,
1395                                           struct btrfs_path *path,
1396                                           u64 bytenr, u64 parent,
1397                                           u64 root_objectid)
1398 {
1399         struct btrfs_key key;
1400         int ret;
1401
1402         key.objectid = bytenr;
1403         if (parent) {
1404                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1405                 key.offset = parent;
1406         } else {
1407                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1408                 key.offset = root_objectid;
1409         }
1410
1411         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1412         btrfs_release_path(path);
1413         return ret;
1414 }
1415
1416 static inline int extent_ref_type(u64 parent, u64 owner)
1417 {
1418         int type;
1419         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1420                 if (parent > 0)
1421                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1422                 else
1423                         type = BTRFS_TREE_BLOCK_REF_KEY;
1424         } else {
1425                 if (parent > 0)
1426                         type = BTRFS_SHARED_DATA_REF_KEY;
1427                 else
1428                         type = BTRFS_EXTENT_DATA_REF_KEY;
1429         }
1430         return type;
1431 }
1432
1433 static int find_next_key(struct btrfs_path *path, int level,
1434                          struct btrfs_key *key)
1435
1436 {
1437         for (; level < BTRFS_MAX_LEVEL; level++) {
1438                 if (!path->nodes[level])
1439                         break;
1440                 if (path->slots[level] + 1 >=
1441                     btrfs_header_nritems(path->nodes[level]))
1442                         continue;
1443                 if (level == 0)
1444                         btrfs_item_key_to_cpu(path->nodes[level], key,
1445                                               path->slots[level] + 1);
1446                 else
1447                         btrfs_node_key_to_cpu(path->nodes[level], key,
1448                                               path->slots[level] + 1);
1449                 return 0;
1450         }
1451         return 1;
1452 }
1453
1454 /*
1455  * look for inline back ref. if back ref is found, *ref_ret is set
1456  * to the address of inline back ref, and 0 is returned.
1457  *
1458  * if back ref isn't found, *ref_ret is set to the address where it
1459  * should be inserted, and -ENOENT is returned.
1460  *
1461  * if insert is true and there are too many inline back refs, the path
1462  * points to the extent item, and -EAGAIN is returned.
1463  *
1464  * NOTE: inline back refs are ordered in the same way that back ref
1465  *       items in the tree are ordered.
1466  */
1467 static noinline_for_stack
1468 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1469                                  struct btrfs_root *root,
1470                                  struct btrfs_path *path,
1471                                  struct btrfs_extent_inline_ref **ref_ret,
1472                                  u64 bytenr, u64 num_bytes,
1473                                  u64 parent, u64 root_objectid,
1474                                  u64 owner, u64 offset, int insert)
1475 {
1476         struct btrfs_key key;
1477         struct extent_buffer *leaf;
1478         struct btrfs_extent_item *ei;
1479         struct btrfs_extent_inline_ref *iref;
1480         u64 flags;
1481         u64 item_size;
1482         unsigned long ptr;
1483         unsigned long end;
1484         int extra_size;
1485         int type;
1486         int want;
1487         int ret;
1488         int err = 0;
1489         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1490                                                  SKINNY_METADATA);
1491
1492         key.objectid = bytenr;
1493         key.type = BTRFS_EXTENT_ITEM_KEY;
1494         key.offset = num_bytes;
1495
1496         want = extent_ref_type(parent, owner);
1497         if (insert) {
1498                 extra_size = btrfs_extent_inline_ref_size(want);
1499                 path->keep_locks = 1;
1500         } else
1501                 extra_size = -1;
1502
1503         /*
1504          * Owner is our parent level, so we can just add one to get the level
1505          * for the block we are interested in.
1506          */
1507         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1508                 key.type = BTRFS_METADATA_ITEM_KEY;
1509                 key.offset = owner;
1510         }
1511
1512 again:
1513         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1514         if (ret < 0) {
1515                 err = ret;
1516                 goto out;
1517         }
1518
1519         /*
1520          * We may be a newly converted file system which still has the old fat
1521          * extent entries for metadata, so try and see if we have one of those.
1522          */
1523         if (ret > 0 && skinny_metadata) {
1524                 skinny_metadata = false;
1525                 if (path->slots[0]) {
1526                         path->slots[0]--;
1527                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1528                                               path->slots[0]);
1529                         if (key.objectid == bytenr &&
1530                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1531                             key.offset == num_bytes)
1532                                 ret = 0;
1533                 }
1534                 if (ret) {
1535                         key.objectid = bytenr;
1536                         key.type = BTRFS_EXTENT_ITEM_KEY;
1537                         key.offset = num_bytes;
1538                         btrfs_release_path(path);
1539                         goto again;
1540                 }
1541         }
1542
1543         if (ret && !insert) {
1544                 err = -ENOENT;
1545                 goto out;
1546         } else if (WARN_ON(ret)) {
1547                 err = -EIO;
1548                 goto out;
1549         }
1550
1551         leaf = path->nodes[0];
1552         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1553 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1554         if (item_size < sizeof(*ei)) {
1555                 if (!insert) {
1556                         err = -ENOENT;
1557                         goto out;
1558                 }
1559                 ret = convert_extent_item_v0(trans, root, path, owner,
1560                                              extra_size);
1561                 if (ret < 0) {
1562                         err = ret;
1563                         goto out;
1564                 }
1565                 leaf = path->nodes[0];
1566                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1567         }
1568 #endif
1569         BUG_ON(item_size < sizeof(*ei));
1570
1571         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1572         flags = btrfs_extent_flags(leaf, ei);
1573
1574         ptr = (unsigned long)(ei + 1);
1575         end = (unsigned long)ei + item_size;
1576
1577         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1578                 ptr += sizeof(struct btrfs_tree_block_info);
1579                 BUG_ON(ptr > end);
1580         }
1581
1582         err = -ENOENT;
1583         while (1) {
1584                 if (ptr >= end) {
1585                         WARN_ON(ptr > end);
1586                         break;
1587                 }
1588                 iref = (struct btrfs_extent_inline_ref *)ptr;
1589                 type = btrfs_extent_inline_ref_type(leaf, iref);
1590                 if (want < type)
1591                         break;
1592                 if (want > type) {
1593                         ptr += btrfs_extent_inline_ref_size(type);
1594                         continue;
1595                 }
1596
1597                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1598                         struct btrfs_extent_data_ref *dref;
1599                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1600                         if (match_extent_data_ref(leaf, dref, root_objectid,
1601                                                   owner, offset)) {
1602                                 err = 0;
1603                                 break;
1604                         }
1605                         if (hash_extent_data_ref_item(leaf, dref) <
1606                             hash_extent_data_ref(root_objectid, owner, offset))
1607                                 break;
1608                 } else {
1609                         u64 ref_offset;
1610                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1611                         if (parent > 0) {
1612                                 if (parent == ref_offset) {
1613                                         err = 0;
1614                                         break;
1615                                 }
1616                                 if (ref_offset < parent)
1617                                         break;
1618                         } else {
1619                                 if (root_objectid == ref_offset) {
1620                                         err = 0;
1621                                         break;
1622                                 }
1623                                 if (ref_offset < root_objectid)
1624                                         break;
1625                         }
1626                 }
1627                 ptr += btrfs_extent_inline_ref_size(type);
1628         }
1629         if (err == -ENOENT && insert) {
1630                 if (item_size + extra_size >=
1631                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1632                         err = -EAGAIN;
1633                         goto out;
1634                 }
1635                 /*
1636                  * To add new inline back ref, we have to make sure
1637                  * there is no corresponding back ref item.
1638                  * For simplicity, we just do not add new inline back
1639                  * ref if there is any kind of item for this block
1640                  */
1641                 if (find_next_key(path, 0, &key) == 0 &&
1642                     key.objectid == bytenr &&
1643                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1644                         err = -EAGAIN;
1645                         goto out;
1646                 }
1647         }
1648         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1649 out:
1650         if (insert) {
1651                 path->keep_locks = 0;
1652                 btrfs_unlock_up_safe(path, 1);
1653         }
1654         return err;
1655 }
1656
1657 /*
1658  * helper to add new inline back ref
1659  */
1660 static noinline_for_stack
1661 void setup_inline_extent_backref(struct btrfs_root *root,
1662                                  struct btrfs_path *path,
1663                                  struct btrfs_extent_inline_ref *iref,
1664                                  u64 parent, u64 root_objectid,
1665                                  u64 owner, u64 offset, int refs_to_add,
1666                                  struct btrfs_delayed_extent_op *extent_op)
1667 {
1668         struct extent_buffer *leaf;
1669         struct btrfs_extent_item *ei;
1670         unsigned long ptr;
1671         unsigned long end;
1672         unsigned long item_offset;
1673         u64 refs;
1674         int size;
1675         int type;
1676
1677         leaf = path->nodes[0];
1678         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1679         item_offset = (unsigned long)iref - (unsigned long)ei;
1680
1681         type = extent_ref_type(parent, owner);
1682         size = btrfs_extent_inline_ref_size(type);
1683
1684         btrfs_extend_item(root, path, size);
1685
1686         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1687         refs = btrfs_extent_refs(leaf, ei);
1688         refs += refs_to_add;
1689         btrfs_set_extent_refs(leaf, ei, refs);
1690         if (extent_op)
1691                 __run_delayed_extent_op(extent_op, leaf, ei);
1692
1693         ptr = (unsigned long)ei + item_offset;
1694         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1695         if (ptr < end - size)
1696                 memmove_extent_buffer(leaf, ptr + size, ptr,
1697                                       end - size - ptr);
1698
1699         iref = (struct btrfs_extent_inline_ref *)ptr;
1700         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1701         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1702                 struct btrfs_extent_data_ref *dref;
1703                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1704                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1705                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1706                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1707                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1708         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1709                 struct btrfs_shared_data_ref *sref;
1710                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1711                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1712                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1713         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1714                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1715         } else {
1716                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1717         }
1718         btrfs_mark_buffer_dirty(leaf);
1719 }
1720
1721 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1722                                  struct btrfs_root *root,
1723                                  struct btrfs_path *path,
1724                                  struct btrfs_extent_inline_ref **ref_ret,
1725                                  u64 bytenr, u64 num_bytes, u64 parent,
1726                                  u64 root_objectid, u64 owner, u64 offset)
1727 {
1728         int ret;
1729
1730         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1731                                            bytenr, num_bytes, parent,
1732                                            root_objectid, owner, offset, 0);
1733         if (ret != -ENOENT)
1734                 return ret;
1735
1736         btrfs_release_path(path);
1737         *ref_ret = NULL;
1738
1739         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1740                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1741                                             root_objectid);
1742         } else {
1743                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1744                                              root_objectid, owner, offset);
1745         }
1746         return ret;
1747 }
1748
1749 /*
1750  * helper to update/remove inline back ref
1751  */
1752 static noinline_for_stack
1753 void update_inline_extent_backref(struct btrfs_root *root,
1754                                   struct btrfs_path *path,
1755                                   struct btrfs_extent_inline_ref *iref,
1756                                   int refs_to_mod,
1757                                   struct btrfs_delayed_extent_op *extent_op,
1758                                   int *last_ref)
1759 {
1760         struct extent_buffer *leaf;
1761         struct btrfs_extent_item *ei;
1762         struct btrfs_extent_data_ref *dref = NULL;
1763         struct btrfs_shared_data_ref *sref = NULL;
1764         unsigned long ptr;
1765         unsigned long end;
1766         u32 item_size;
1767         int size;
1768         int type;
1769         u64 refs;
1770
1771         leaf = path->nodes[0];
1772         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1773         refs = btrfs_extent_refs(leaf, ei);
1774         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1775         refs += refs_to_mod;
1776         btrfs_set_extent_refs(leaf, ei, refs);
1777         if (extent_op)
1778                 __run_delayed_extent_op(extent_op, leaf, ei);
1779
1780         type = btrfs_extent_inline_ref_type(leaf, iref);
1781
1782         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1783                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1784                 refs = btrfs_extent_data_ref_count(leaf, dref);
1785         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1786                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1787                 refs = btrfs_shared_data_ref_count(leaf, sref);
1788         } else {
1789                 refs = 1;
1790                 BUG_ON(refs_to_mod != -1);
1791         }
1792
1793         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1794         refs += refs_to_mod;
1795
1796         if (refs > 0) {
1797                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1798                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1799                 else
1800                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1801         } else {
1802                 *last_ref = 1;
1803                 size =  btrfs_extent_inline_ref_size(type);
1804                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1805                 ptr = (unsigned long)iref;
1806                 end = (unsigned long)ei + item_size;
1807                 if (ptr + size < end)
1808                         memmove_extent_buffer(leaf, ptr, ptr + size,
1809                                               end - ptr - size);
1810                 item_size -= size;
1811                 btrfs_truncate_item(root, path, item_size, 1);
1812         }
1813         btrfs_mark_buffer_dirty(leaf);
1814 }
1815
1816 static noinline_for_stack
1817 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1818                                  struct btrfs_root *root,
1819                                  struct btrfs_path *path,
1820                                  u64 bytenr, u64 num_bytes, u64 parent,
1821                                  u64 root_objectid, u64 owner,
1822                                  u64 offset, int refs_to_add,
1823                                  struct btrfs_delayed_extent_op *extent_op)
1824 {
1825         struct btrfs_extent_inline_ref *iref;
1826         int ret;
1827
1828         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1829                                            bytenr, num_bytes, parent,
1830                                            root_objectid, owner, offset, 1);
1831         if (ret == 0) {
1832                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1833                 update_inline_extent_backref(root, path, iref,
1834                                              refs_to_add, extent_op, NULL);
1835         } else if (ret == -ENOENT) {
1836                 setup_inline_extent_backref(root, path, iref, parent,
1837                                             root_objectid, owner, offset,
1838                                             refs_to_add, extent_op);
1839                 ret = 0;
1840         }
1841         return ret;
1842 }
1843
1844 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1845                                  struct btrfs_root *root,
1846                                  struct btrfs_path *path,
1847                                  u64 bytenr, u64 parent, u64 root_objectid,
1848                                  u64 owner, u64 offset, int refs_to_add)
1849 {
1850         int ret;
1851         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1852                 BUG_ON(refs_to_add != 1);
1853                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1854                                             parent, root_objectid);
1855         } else {
1856                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1857                                              parent, root_objectid,
1858                                              owner, offset, refs_to_add);
1859         }
1860         return ret;
1861 }
1862
1863 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1864                                  struct btrfs_root *root,
1865                                  struct btrfs_path *path,
1866                                  struct btrfs_extent_inline_ref *iref,
1867                                  int refs_to_drop, int is_data, int *last_ref)
1868 {
1869         int ret = 0;
1870
1871         BUG_ON(!is_data && refs_to_drop != 1);
1872         if (iref) {
1873                 update_inline_extent_backref(root, path, iref,
1874                                              -refs_to_drop, NULL, last_ref);
1875         } else if (is_data) {
1876                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1877                                              last_ref);
1878         } else {
1879                 *last_ref = 1;
1880                 ret = btrfs_del_item(trans, root, path);
1881         }
1882         return ret;
1883 }
1884
1885 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1886 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1887                                u64 *discarded_bytes)
1888 {
1889         int j, ret = 0;
1890         u64 bytes_left, end;
1891         u64 aligned_start = ALIGN(start, 1 << 9);
1892
1893         if (WARN_ON(start != aligned_start)) {
1894                 len -= aligned_start - start;
1895                 len = round_down(len, 1 << 9);
1896                 start = aligned_start;
1897         }
1898
1899         *discarded_bytes = 0;
1900
1901         if (!len)
1902                 return 0;
1903
1904         end = start + len;
1905         bytes_left = len;
1906
1907         /* Skip any superblocks on this device. */
1908         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1909                 u64 sb_start = btrfs_sb_offset(j);
1910                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1911                 u64 size = sb_start - start;
1912
1913                 if (!in_range(sb_start, start, bytes_left) &&
1914                     !in_range(sb_end, start, bytes_left) &&
1915                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1916                         continue;
1917
1918                 /*
1919                  * Superblock spans beginning of range.  Adjust start and
1920                  * try again.
1921                  */
1922                 if (sb_start <= start) {
1923                         start += sb_end - start;
1924                         if (start > end) {
1925                                 bytes_left = 0;
1926                                 break;
1927                         }
1928                         bytes_left = end - start;
1929                         continue;
1930                 }
1931
1932                 if (size) {
1933                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1934                                                    GFP_NOFS, 0);
1935                         if (!ret)
1936                                 *discarded_bytes += size;
1937                         else if (ret != -EOPNOTSUPP)
1938                                 return ret;
1939                 }
1940
1941                 start = sb_end;
1942                 if (start > end) {
1943                         bytes_left = 0;
1944                         break;
1945                 }
1946                 bytes_left = end - start;
1947         }
1948
1949         if (bytes_left) {
1950                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1951                                            GFP_NOFS, 0);
1952                 if (!ret)
1953                         *discarded_bytes += bytes_left;
1954         }
1955         return ret;
1956 }
1957
1958 int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1959                          u64 num_bytes, u64 *actual_bytes)
1960 {
1961         int ret;
1962         u64 discarded_bytes = 0;
1963         struct btrfs_bio *bbio = NULL;
1964
1965
1966         /* Tell the block device(s) that the sectors can be discarded */
1967         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1968                               bytenr, &num_bytes, &bbio, 0);
1969         /* Error condition is -ENOMEM */
1970         if (!ret) {
1971                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1972                 int i;
1973
1974
1975                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1976                         u64 bytes;
1977                         if (!stripe->dev->can_discard)
1978                                 continue;
1979
1980                         ret = btrfs_issue_discard(stripe->dev->bdev,
1981                                                   stripe->physical,
1982                                                   stripe->length,
1983                                                   &bytes);
1984                         if (!ret)
1985                                 discarded_bytes += bytes;
1986                         else if (ret != -EOPNOTSUPP)
1987                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1988
1989                         /*
1990                          * Just in case we get back EOPNOTSUPP for some reason,
1991                          * just ignore the return value so we don't screw up
1992                          * people calling discard_extent.
1993                          */
1994                         ret = 0;
1995                 }
1996                 btrfs_put_bbio(bbio);
1997         }
1998
1999         if (actual_bytes)
2000                 *actual_bytes = discarded_bytes;
2001
2002
2003         if (ret == -EOPNOTSUPP)
2004                 ret = 0;
2005         return ret;
2006 }
2007
2008 /* Can return -ENOMEM */
2009 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2010                          struct btrfs_root *root,
2011                          u64 bytenr, u64 num_bytes, u64 parent,
2012                          u64 root_objectid, u64 owner, u64 offset,
2013                          int no_quota)
2014 {
2015         int ret;
2016         struct btrfs_fs_info *fs_info = root->fs_info;
2017
2018         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2019                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2020
2021         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2022                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2023                                         num_bytes,
2024                                         parent, root_objectid, (int)owner,
2025                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
2026         } else {
2027                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2028                                         num_bytes,
2029                                         parent, root_objectid, owner, offset,
2030                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
2031         }
2032         return ret;
2033 }
2034
2035 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2036                                   struct btrfs_root *root,
2037                                   struct btrfs_delayed_ref_node *node,
2038                                   u64 parent, u64 root_objectid,
2039                                   u64 owner, u64 offset, int refs_to_add,
2040                                   struct btrfs_delayed_extent_op *extent_op)
2041 {
2042         struct btrfs_fs_info *fs_info = root->fs_info;
2043         struct btrfs_path *path;
2044         struct extent_buffer *leaf;
2045         struct btrfs_extent_item *item;
2046         struct btrfs_key key;
2047         u64 bytenr = node->bytenr;
2048         u64 num_bytes = node->num_bytes;
2049         u64 refs;
2050         int ret;
2051         int no_quota = node->no_quota;
2052
2053         path = btrfs_alloc_path();
2054         if (!path)
2055                 return -ENOMEM;
2056
2057         if (!is_fstree(root_objectid) || !root->fs_info->quota_enabled)
2058                 no_quota = 1;
2059
2060         path->reada = 1;
2061         path->leave_spinning = 1;
2062         /* this will setup the path even if it fails to insert the back ref */
2063         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2064                                            bytenr, num_bytes, parent,
2065                                            root_objectid, owner, offset,
2066                                            refs_to_add, extent_op);
2067         if ((ret < 0 && ret != -EAGAIN) || !ret)
2068                 goto out;
2069
2070         /*
2071          * Ok we had -EAGAIN which means we didn't have space to insert and
2072          * inline extent ref, so just update the reference count and add a
2073          * normal backref.
2074          */
2075         leaf = path->nodes[0];
2076         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2077         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2078         refs = btrfs_extent_refs(leaf, item);
2079         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2080         if (extent_op)
2081                 __run_delayed_extent_op(extent_op, leaf, item);
2082
2083         btrfs_mark_buffer_dirty(leaf);
2084         btrfs_release_path(path);
2085
2086         path->reada = 1;
2087         path->leave_spinning = 1;
2088         /* now insert the actual backref */
2089         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2090                                     path, bytenr, parent, root_objectid,
2091                                     owner, offset, refs_to_add);
2092         if (ret)
2093                 btrfs_abort_transaction(trans, root, ret);
2094 out:
2095         btrfs_free_path(path);
2096         return ret;
2097 }
2098
2099 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2100                                 struct btrfs_root *root,
2101                                 struct btrfs_delayed_ref_node *node,
2102                                 struct btrfs_delayed_extent_op *extent_op,
2103                                 int insert_reserved)
2104 {
2105         int ret = 0;
2106         struct btrfs_delayed_data_ref *ref;
2107         struct btrfs_key ins;
2108         u64 parent = 0;
2109         u64 ref_root = 0;
2110         u64 flags = 0;
2111
2112         ins.objectid = node->bytenr;
2113         ins.offset = node->num_bytes;
2114         ins.type = BTRFS_EXTENT_ITEM_KEY;
2115
2116         ref = btrfs_delayed_node_to_data_ref(node);
2117         trace_run_delayed_data_ref(node, ref, node->action);
2118
2119         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2120                 parent = ref->parent;
2121         ref_root = ref->root;
2122
2123         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2124                 if (extent_op)
2125                         flags |= extent_op->flags_to_set;
2126                 ret = alloc_reserved_file_extent(trans, root,
2127                                                  parent, ref_root, flags,
2128                                                  ref->objectid, ref->offset,
2129                                                  &ins, node->ref_mod);
2130         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2131                 ret = __btrfs_inc_extent_ref(trans, root, node, parent,
2132                                              ref_root, ref->objectid,
2133                                              ref->offset, node->ref_mod,
2134                                              extent_op);
2135         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2136                 ret = __btrfs_free_extent(trans, root, node, parent,
2137                                           ref_root, ref->objectid,
2138                                           ref->offset, node->ref_mod,
2139                                           extent_op);
2140         } else {
2141                 BUG();
2142         }
2143         return ret;
2144 }
2145
2146 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2147                                     struct extent_buffer *leaf,
2148                                     struct btrfs_extent_item *ei)
2149 {
2150         u64 flags = btrfs_extent_flags(leaf, ei);
2151         if (extent_op->update_flags) {
2152                 flags |= extent_op->flags_to_set;
2153                 btrfs_set_extent_flags(leaf, ei, flags);
2154         }
2155
2156         if (extent_op->update_key) {
2157                 struct btrfs_tree_block_info *bi;
2158                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2159                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2160                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2161         }
2162 }
2163
2164 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2165                                  struct btrfs_root *root,
2166                                  struct btrfs_delayed_ref_node *node,
2167                                  struct btrfs_delayed_extent_op *extent_op)
2168 {
2169         struct btrfs_key key;
2170         struct btrfs_path *path;
2171         struct btrfs_extent_item *ei;
2172         struct extent_buffer *leaf;
2173         u32 item_size;
2174         int ret;
2175         int err = 0;
2176         int metadata = !extent_op->is_data;
2177
2178         if (trans->aborted)
2179                 return 0;
2180
2181         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2182                 metadata = 0;
2183
2184         path = btrfs_alloc_path();
2185         if (!path)
2186                 return -ENOMEM;
2187
2188         key.objectid = node->bytenr;
2189
2190         if (metadata) {
2191                 key.type = BTRFS_METADATA_ITEM_KEY;
2192                 key.offset = extent_op->level;
2193         } else {
2194                 key.type = BTRFS_EXTENT_ITEM_KEY;
2195                 key.offset = node->num_bytes;
2196         }
2197
2198 again:
2199         path->reada = 1;
2200         path->leave_spinning = 1;
2201         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2202                                 path, 0, 1);
2203         if (ret < 0) {
2204                 err = ret;
2205                 goto out;
2206         }
2207         if (ret > 0) {
2208                 if (metadata) {
2209                         if (path->slots[0] > 0) {
2210                                 path->slots[0]--;
2211                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2212                                                       path->slots[0]);
2213                                 if (key.objectid == node->bytenr &&
2214                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2215                                     key.offset == node->num_bytes)
2216                                         ret = 0;
2217                         }
2218                         if (ret > 0) {
2219                                 btrfs_release_path(path);
2220                                 metadata = 0;
2221
2222                                 key.objectid = node->bytenr;
2223                                 key.offset = node->num_bytes;
2224                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2225                                 goto again;
2226                         }
2227                 } else {
2228                         err = -EIO;
2229                         goto out;
2230                 }
2231         }
2232
2233         leaf = path->nodes[0];
2234         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2235 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2236         if (item_size < sizeof(*ei)) {
2237                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2238                                              path, (u64)-1, 0);
2239                 if (ret < 0) {
2240                         err = ret;
2241                         goto out;
2242                 }
2243                 leaf = path->nodes[0];
2244                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2245         }
2246 #endif
2247         BUG_ON(item_size < sizeof(*ei));
2248         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2249         __run_delayed_extent_op(extent_op, leaf, ei);
2250
2251         btrfs_mark_buffer_dirty(leaf);
2252 out:
2253         btrfs_free_path(path);
2254         return err;
2255 }
2256
2257 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2258                                 struct btrfs_root *root,
2259                                 struct btrfs_delayed_ref_node *node,
2260                                 struct btrfs_delayed_extent_op *extent_op,
2261                                 int insert_reserved)
2262 {
2263         int ret = 0;
2264         struct btrfs_delayed_tree_ref *ref;
2265         struct btrfs_key ins;
2266         u64 parent = 0;
2267         u64 ref_root = 0;
2268         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2269                                                  SKINNY_METADATA);
2270
2271         ref = btrfs_delayed_node_to_tree_ref(node);
2272         trace_run_delayed_tree_ref(node, ref, node->action);
2273
2274         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2275                 parent = ref->parent;
2276         ref_root = ref->root;
2277
2278         ins.objectid = node->bytenr;
2279         if (skinny_metadata) {
2280                 ins.offset = ref->level;
2281                 ins.type = BTRFS_METADATA_ITEM_KEY;
2282         } else {
2283                 ins.offset = node->num_bytes;
2284                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2285         }
2286
2287         BUG_ON(node->ref_mod != 1);
2288         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2289                 BUG_ON(!extent_op || !extent_op->update_flags);
2290                 ret = alloc_reserved_tree_block(trans, root,
2291                                                 parent, ref_root,
2292                                                 extent_op->flags_to_set,
2293                                                 &extent_op->key,
2294                                                 ref->level, &ins,
2295                                                 node->no_quota);
2296         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2297                 ret = __btrfs_inc_extent_ref(trans, root, node,
2298                                              parent, ref_root,
2299                                              ref->level, 0, 1,
2300                                              extent_op);
2301         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2302                 ret = __btrfs_free_extent(trans, root, node,
2303                                           parent, ref_root,
2304                                           ref->level, 0, 1, extent_op);
2305         } else {
2306                 BUG();
2307         }
2308         return ret;
2309 }
2310
2311 /* helper function to actually process a single delayed ref entry */
2312 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2313                                struct btrfs_root *root,
2314                                struct btrfs_delayed_ref_node *node,
2315                                struct btrfs_delayed_extent_op *extent_op,
2316                                int insert_reserved)
2317 {
2318         int ret = 0;
2319
2320         if (trans->aborted) {
2321                 if (insert_reserved)
2322                         btrfs_pin_extent(root, node->bytenr,
2323                                          node->num_bytes, 1);
2324                 return 0;
2325         }
2326
2327         if (btrfs_delayed_ref_is_head(node)) {
2328                 struct btrfs_delayed_ref_head *head;
2329                 /*
2330                  * we've hit the end of the chain and we were supposed
2331                  * to insert this extent into the tree.  But, it got
2332                  * deleted before we ever needed to insert it, so all
2333                  * we have to do is clean up the accounting
2334                  */
2335                 BUG_ON(extent_op);
2336                 head = btrfs_delayed_node_to_head(node);
2337                 trace_run_delayed_ref_head(node, head, node->action);
2338
2339                 if (insert_reserved) {
2340                         btrfs_pin_extent(root, node->bytenr,
2341                                          node->num_bytes, 1);
2342                         if (head->is_data) {
2343                                 ret = btrfs_del_csums(trans, root,
2344                                                       node->bytenr,
2345                                                       node->num_bytes);
2346                         }
2347                 }
2348
2349                 /* Also free its reserved qgroup space */
2350                 btrfs_qgroup_free_delayed_ref(root->fs_info,
2351                                               head->qgroup_ref_root,
2352                                               head->qgroup_reserved);
2353                 return ret;
2354         }
2355
2356         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2357             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2358                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2359                                            insert_reserved);
2360         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2361                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2362                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2363                                            insert_reserved);
2364         else
2365                 BUG();
2366         return ret;
2367 }
2368
2369 static inline struct btrfs_delayed_ref_node *
2370 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2371 {
2372         struct btrfs_delayed_ref_node *ref;
2373
2374         if (list_empty(&head->ref_list))
2375                 return NULL;
2376
2377         /*
2378          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2379          * This is to prevent a ref count from going down to zero, which deletes
2380          * the extent item from the extent tree, when there still are references
2381          * to add, which would fail because they would not find the extent item.
2382          */
2383         list_for_each_entry(ref, &head->ref_list, list) {
2384                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2385                         return ref;
2386         }
2387
2388         return list_entry(head->ref_list.next, struct btrfs_delayed_ref_node,
2389                           list);
2390 }
2391
2392 /*
2393  * Returns 0 on success or if called with an already aborted transaction.
2394  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2395  */
2396 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2397                                              struct btrfs_root *root,
2398                                              unsigned long nr)
2399 {
2400         struct btrfs_delayed_ref_root *delayed_refs;
2401         struct btrfs_delayed_ref_node *ref;
2402         struct btrfs_delayed_ref_head *locked_ref = NULL;
2403         struct btrfs_delayed_extent_op *extent_op;
2404         struct btrfs_fs_info *fs_info = root->fs_info;
2405         ktime_t start = ktime_get();
2406         int ret;
2407         unsigned long count = 0;
2408         unsigned long actual_count = 0;
2409         int must_insert_reserved = 0;
2410
2411         delayed_refs = &trans->transaction->delayed_refs;
2412         while (1) {
2413                 if (!locked_ref) {
2414                         if (count >= nr)
2415                                 break;
2416
2417                         spin_lock(&delayed_refs->lock);
2418                         locked_ref = btrfs_select_ref_head(trans);
2419                         if (!locked_ref) {
2420                                 spin_unlock(&delayed_refs->lock);
2421                                 break;
2422                         }
2423
2424                         /* grab the lock that says we are going to process
2425                          * all the refs for this head */
2426                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2427                         spin_unlock(&delayed_refs->lock);
2428                         /*
2429                          * we may have dropped the spin lock to get the head
2430                          * mutex lock, and that might have given someone else
2431                          * time to free the head.  If that's true, it has been
2432                          * removed from our list and we can move on.
2433                          */
2434                         if (ret == -EAGAIN) {
2435                                 locked_ref = NULL;
2436                                 count++;
2437                                 continue;
2438                         }
2439                 }
2440
2441                 spin_lock(&locked_ref->lock);
2442
2443                 /*
2444                  * locked_ref is the head node, so we have to go one
2445                  * node back for any delayed ref updates
2446                  */
2447                 ref = select_delayed_ref(locked_ref);
2448
2449                 if (ref && ref->seq &&
2450                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2451                         spin_unlock(&locked_ref->lock);
2452                         btrfs_delayed_ref_unlock(locked_ref);
2453                         spin_lock(&delayed_refs->lock);
2454                         locked_ref->processing = 0;
2455                         delayed_refs->num_heads_ready++;
2456                         spin_unlock(&delayed_refs->lock);
2457                         locked_ref = NULL;
2458                         cond_resched();
2459                         count++;
2460                         continue;
2461                 }
2462
2463                 /*
2464                  * record the must insert reserved flag before we
2465                  * drop the spin lock.
2466                  */
2467                 must_insert_reserved = locked_ref->must_insert_reserved;
2468                 locked_ref->must_insert_reserved = 0;
2469
2470                 extent_op = locked_ref->extent_op;
2471                 locked_ref->extent_op = NULL;
2472
2473                 if (!ref) {
2474
2475
2476                         /* All delayed refs have been processed, Go ahead
2477                          * and send the head node to run_one_delayed_ref,
2478                          * so that any accounting fixes can happen
2479                          */
2480                         ref = &locked_ref->node;
2481
2482                         if (extent_op && must_insert_reserved) {
2483                                 btrfs_free_delayed_extent_op(extent_op);
2484                                 extent_op = NULL;
2485                         }
2486
2487                         if (extent_op) {
2488                                 spin_unlock(&locked_ref->lock);
2489                                 ret = run_delayed_extent_op(trans, root,
2490                                                             ref, extent_op);
2491                                 btrfs_free_delayed_extent_op(extent_op);
2492
2493                                 if (ret) {
2494                                         /*
2495                                          * Need to reset must_insert_reserved if
2496                                          * there was an error so the abort stuff
2497                                          * can cleanup the reserved space
2498                                          * properly.
2499                                          */
2500                                         if (must_insert_reserved)
2501                                                 locked_ref->must_insert_reserved = 1;
2502                                         locked_ref->processing = 0;
2503                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2504                                         btrfs_delayed_ref_unlock(locked_ref);
2505                                         return ret;
2506                                 }
2507                                 continue;
2508                         }
2509
2510                         /*
2511                          * Need to drop our head ref lock and re-aqcuire the
2512                          * delayed ref lock and then re-check to make sure
2513                          * nobody got added.
2514                          */
2515                         spin_unlock(&locked_ref->lock);
2516                         spin_lock(&delayed_refs->lock);
2517                         spin_lock(&locked_ref->lock);
2518                         if (!list_empty(&locked_ref->ref_list) ||
2519                             locked_ref->extent_op) {
2520                                 spin_unlock(&locked_ref->lock);
2521                                 spin_unlock(&delayed_refs->lock);
2522                                 continue;
2523                         }
2524                         ref->in_tree = 0;
2525                         delayed_refs->num_heads--;
2526                         rb_erase(&locked_ref->href_node,
2527                                  &delayed_refs->href_root);
2528                         spin_unlock(&delayed_refs->lock);
2529                 } else {
2530                         actual_count++;
2531                         ref->in_tree = 0;
2532                         list_del(&ref->list);
2533                 }
2534                 atomic_dec(&delayed_refs->num_entries);
2535
2536                 if (!btrfs_delayed_ref_is_head(ref)) {
2537                         /*
2538                          * when we play the delayed ref, also correct the
2539                          * ref_mod on head
2540                          */
2541                         switch (ref->action) {
2542                         case BTRFS_ADD_DELAYED_REF:
2543                         case BTRFS_ADD_DELAYED_EXTENT:
2544                                 locked_ref->node.ref_mod -= ref->ref_mod;
2545                                 break;
2546                         case BTRFS_DROP_DELAYED_REF:
2547                                 locked_ref->node.ref_mod += ref->ref_mod;
2548                                 break;
2549                         default:
2550                                 WARN_ON(1);
2551                         }
2552                 }
2553                 spin_unlock(&locked_ref->lock);
2554
2555                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2556                                           must_insert_reserved);
2557
2558                 btrfs_free_delayed_extent_op(extent_op);
2559                 if (ret) {
2560                         locked_ref->processing = 0;
2561                         btrfs_delayed_ref_unlock(locked_ref);
2562                         btrfs_put_delayed_ref(ref);
2563                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2564                         return ret;
2565                 }
2566
2567                 /*
2568                  * If this node is a head, that means all the refs in this head
2569                  * have been dealt with, and we will pick the next head to deal
2570                  * with, so we must unlock the head and drop it from the cluster
2571                  * list before we release it.
2572                  */
2573                 if (btrfs_delayed_ref_is_head(ref)) {
2574                         if (locked_ref->is_data &&
2575                             locked_ref->total_ref_mod < 0) {
2576                                 spin_lock(&delayed_refs->lock);
2577                                 delayed_refs->pending_csums -= ref->num_bytes;
2578                                 spin_unlock(&delayed_refs->lock);
2579                         }
2580                         btrfs_delayed_ref_unlock(locked_ref);
2581                         locked_ref = NULL;
2582                 }
2583                 btrfs_put_delayed_ref(ref);
2584                 count++;
2585                 cond_resched();
2586         }
2587
2588         /*
2589          * We don't want to include ref heads since we can have empty ref heads
2590          * and those will drastically skew our runtime down since we just do
2591          * accounting, no actual extent tree updates.
2592          */
2593         if (actual_count > 0) {
2594                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2595                 u64 avg;
2596
2597                 /*
2598                  * We weigh the current average higher than our current runtime
2599                  * to avoid large swings in the average.
2600                  */
2601                 spin_lock(&delayed_refs->lock);
2602                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2603                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2604                 spin_unlock(&delayed_refs->lock);
2605         }
2606         return 0;
2607 }
2608
2609 #ifdef SCRAMBLE_DELAYED_REFS
2610 /*
2611  * Normally delayed refs get processed in ascending bytenr order. This
2612  * correlates in most cases to the order added. To expose dependencies on this
2613  * order, we start to process the tree in the middle instead of the beginning
2614  */
2615 static u64 find_middle(struct rb_root *root)
2616 {
2617         struct rb_node *n = root->rb_node;
2618         struct btrfs_delayed_ref_node *entry;
2619         int alt = 1;
2620         u64 middle;
2621         u64 first = 0, last = 0;
2622
2623         n = rb_first(root);
2624         if (n) {
2625                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2626                 first = entry->bytenr;
2627         }
2628         n = rb_last(root);
2629         if (n) {
2630                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2631                 last = entry->bytenr;
2632         }
2633         n = root->rb_node;
2634
2635         while (n) {
2636                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2637                 WARN_ON(!entry->in_tree);
2638
2639                 middle = entry->bytenr;
2640
2641                 if (alt)
2642                         n = n->rb_left;
2643                 else
2644                         n = n->rb_right;
2645
2646                 alt = 1 - alt;
2647         }
2648         return middle;
2649 }
2650 #endif
2651
2652 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2653 {
2654         u64 num_bytes;
2655
2656         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2657                              sizeof(struct btrfs_extent_inline_ref));
2658         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2659                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2660
2661         /*
2662          * We don't ever fill up leaves all the way so multiply by 2 just to be
2663          * closer to what we're really going to want to ouse.
2664          */
2665         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2666 }
2667
2668 /*
2669  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2670  * would require to store the csums for that many bytes.
2671  */
2672 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2673 {
2674         u64 csum_size;
2675         u64 num_csums_per_leaf;
2676         u64 num_csums;
2677
2678         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
2679         num_csums_per_leaf = div64_u64(csum_size,
2680                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2681         num_csums = div64_u64(csum_bytes, root->sectorsize);
2682         num_csums += num_csums_per_leaf - 1;
2683         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2684         return num_csums;
2685 }
2686
2687 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2688                                        struct btrfs_root *root)
2689 {
2690         struct btrfs_block_rsv *global_rsv;
2691         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2692         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2693         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2694         u64 num_bytes, num_dirty_bgs_bytes;
2695         int ret = 0;
2696
2697         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2698         num_heads = heads_to_leaves(root, num_heads);
2699         if (num_heads > 1)
2700                 num_bytes += (num_heads - 1) * root->nodesize;
2701         num_bytes <<= 1;
2702         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2703         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2704                                                              num_dirty_bgs);
2705         global_rsv = &root->fs_info->global_block_rsv;
2706
2707         /*
2708          * If we can't allocate any more chunks lets make sure we have _lots_ of
2709          * wiggle room since running delayed refs can create more delayed refs.
2710          */
2711         if (global_rsv->space_info->full) {
2712                 num_dirty_bgs_bytes <<= 1;
2713                 num_bytes <<= 1;
2714         }
2715
2716         spin_lock(&global_rsv->lock);
2717         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2718                 ret = 1;
2719         spin_unlock(&global_rsv->lock);
2720         return ret;
2721 }
2722
2723 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2724                                        struct btrfs_root *root)
2725 {
2726         struct btrfs_fs_info *fs_info = root->fs_info;
2727         u64 num_entries =
2728                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2729         u64 avg_runtime;
2730         u64 val;
2731
2732         smp_mb();
2733         avg_runtime = fs_info->avg_delayed_ref_runtime;
2734         val = num_entries * avg_runtime;
2735         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2736                 return 1;
2737         if (val >= NSEC_PER_SEC / 2)
2738                 return 2;
2739
2740         return btrfs_check_space_for_delayed_refs(trans, root);
2741 }
2742
2743 struct async_delayed_refs {
2744         struct btrfs_root *root;
2745         int count;
2746         int error;
2747         int sync;
2748         struct completion wait;
2749         struct btrfs_work work;
2750 };
2751
2752 static void delayed_ref_async_start(struct btrfs_work *work)
2753 {
2754         struct async_delayed_refs *async;
2755         struct btrfs_trans_handle *trans;
2756         int ret;
2757
2758         async = container_of(work, struct async_delayed_refs, work);
2759
2760         trans = btrfs_join_transaction(async->root);
2761         if (IS_ERR(trans)) {
2762                 async->error = PTR_ERR(trans);
2763                 goto done;
2764         }
2765
2766         /*
2767          * trans->sync means that when we call end_transaciton, we won't
2768          * wait on delayed refs
2769          */
2770         trans->sync = true;
2771         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2772         if (ret)
2773                 async->error = ret;
2774
2775         ret = btrfs_end_transaction(trans, async->root);
2776         if (ret && !async->error)
2777                 async->error = ret;
2778 done:
2779         if (async->sync)
2780                 complete(&async->wait);
2781         else
2782                 kfree(async);
2783 }
2784
2785 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2786                                  unsigned long count, int wait)
2787 {
2788         struct async_delayed_refs *async;
2789         int ret;
2790
2791         async = kmalloc(sizeof(*async), GFP_NOFS);
2792         if (!async)
2793                 return -ENOMEM;
2794
2795         async->root = root->fs_info->tree_root;
2796         async->count = count;
2797         async->error = 0;
2798         if (wait)
2799                 async->sync = 1;
2800         else
2801                 async->sync = 0;
2802         init_completion(&async->wait);
2803
2804         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2805                         delayed_ref_async_start, NULL, NULL);
2806
2807         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2808
2809         if (wait) {
2810                 wait_for_completion(&async->wait);
2811                 ret = async->error;
2812                 kfree(async);
2813                 return ret;
2814         }
2815         return 0;
2816 }
2817
2818 /*
2819  * this starts processing the delayed reference count updates and
2820  * extent insertions we have queued up so far.  count can be
2821  * 0, which means to process everything in the tree at the start
2822  * of the run (but not newly added entries), or it can be some target
2823  * number you'd like to process.
2824  *
2825  * Returns 0 on success or if called with an aborted transaction
2826  * Returns <0 on error and aborts the transaction
2827  */
2828 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2829                            struct btrfs_root *root, unsigned long count)
2830 {
2831         struct rb_node *node;
2832         struct btrfs_delayed_ref_root *delayed_refs;
2833         struct btrfs_delayed_ref_head *head;
2834         int ret;
2835         int run_all = count == (unsigned long)-1;
2836         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2837
2838         /* We'll clean this up in btrfs_cleanup_transaction */
2839         if (trans->aborted)
2840                 return 0;
2841
2842         if (root == root->fs_info->extent_root)
2843                 root = root->fs_info->tree_root;
2844
2845         delayed_refs = &trans->transaction->delayed_refs;
2846         if (count == 0)
2847                 count = atomic_read(&delayed_refs->num_entries) * 2;
2848
2849 again:
2850 #ifdef SCRAMBLE_DELAYED_REFS
2851         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2852 #endif
2853         trans->can_flush_pending_bgs = false;
2854         ret = __btrfs_run_delayed_refs(trans, root, count);
2855         if (ret < 0) {
2856                 btrfs_abort_transaction(trans, root, ret);
2857                 return ret;
2858         }
2859
2860         if (run_all) {
2861                 if (!list_empty(&trans->new_bgs))
2862                         btrfs_create_pending_block_groups(trans, root);
2863
2864                 spin_lock(&delayed_refs->lock);
2865                 node = rb_first(&delayed_refs->href_root);
2866                 if (!node) {
2867                         spin_unlock(&delayed_refs->lock);
2868                         goto out;
2869                 }
2870                 count = (unsigned long)-1;
2871
2872                 while (node) {
2873                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2874                                         href_node);
2875                         if (btrfs_delayed_ref_is_head(&head->node)) {
2876                                 struct btrfs_delayed_ref_node *ref;
2877
2878                                 ref = &head->node;
2879                                 atomic_inc(&ref->refs);
2880
2881                                 spin_unlock(&delayed_refs->lock);
2882                                 /*
2883                                  * Mutex was contended, block until it's
2884                                  * released and try again
2885                                  */
2886                                 mutex_lock(&head->mutex);
2887                                 mutex_unlock(&head->mutex);
2888
2889                                 btrfs_put_delayed_ref(ref);
2890                                 cond_resched();
2891                                 goto again;
2892                         } else {
2893                                 WARN_ON(1);
2894                         }
2895                         node = rb_next(node);
2896                 }
2897                 spin_unlock(&delayed_refs->lock);
2898                 cond_resched();
2899                 goto again;
2900         }
2901 out:
2902         assert_qgroups_uptodate(trans);
2903         trans->can_flush_pending_bgs = can_flush_pending_bgs;
2904         return 0;
2905 }
2906
2907 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2908                                 struct btrfs_root *root,
2909                                 u64 bytenr, u64 num_bytes, u64 flags,
2910                                 int level, int is_data)
2911 {
2912         struct btrfs_delayed_extent_op *extent_op;
2913         int ret;
2914
2915         extent_op = btrfs_alloc_delayed_extent_op();
2916         if (!extent_op)
2917                 return -ENOMEM;
2918
2919         extent_op->flags_to_set = flags;
2920         extent_op->update_flags = 1;
2921         extent_op->update_key = 0;
2922         extent_op->is_data = is_data ? 1 : 0;
2923         extent_op->level = level;
2924
2925         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2926                                           num_bytes, extent_op);
2927         if (ret)
2928                 btrfs_free_delayed_extent_op(extent_op);
2929         return ret;
2930 }
2931
2932 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2933                                       struct btrfs_root *root,
2934                                       struct btrfs_path *path,
2935                                       u64 objectid, u64 offset, u64 bytenr)
2936 {
2937         struct btrfs_delayed_ref_head *head;
2938         struct btrfs_delayed_ref_node *ref;
2939         struct btrfs_delayed_data_ref *data_ref;
2940         struct btrfs_delayed_ref_root *delayed_refs;
2941         int ret = 0;
2942
2943         delayed_refs = &trans->transaction->delayed_refs;
2944         spin_lock(&delayed_refs->lock);
2945         head = btrfs_find_delayed_ref_head(trans, bytenr);
2946         if (!head) {
2947                 spin_unlock(&delayed_refs->lock);
2948                 return 0;
2949         }
2950
2951         if (!mutex_trylock(&head->mutex)) {
2952                 atomic_inc(&head->node.refs);
2953                 spin_unlock(&delayed_refs->lock);
2954
2955                 btrfs_release_path(path);
2956
2957                 /*
2958                  * Mutex was contended, block until it's released and let
2959                  * caller try again
2960                  */
2961                 mutex_lock(&head->mutex);
2962                 mutex_unlock(&head->mutex);
2963                 btrfs_put_delayed_ref(&head->node);
2964                 return -EAGAIN;
2965         }
2966         spin_unlock(&delayed_refs->lock);
2967
2968         spin_lock(&head->lock);
2969         list_for_each_entry(ref, &head->ref_list, list) {
2970                 /* If it's a shared ref we know a cross reference exists */
2971                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2972                         ret = 1;
2973                         break;
2974                 }
2975
2976                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2977
2978                 /*
2979                  * If our ref doesn't match the one we're currently looking at
2980                  * then we have a cross reference.
2981                  */
2982                 if (data_ref->root != root->root_key.objectid ||
2983                     data_ref->objectid != objectid ||
2984                     data_ref->offset != offset) {
2985                         ret = 1;
2986                         break;
2987                 }
2988         }
2989         spin_unlock(&head->lock);
2990         mutex_unlock(&head->mutex);
2991         return ret;
2992 }
2993
2994 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2995                                         struct btrfs_root *root,
2996                                         struct btrfs_path *path,
2997                                         u64 objectid, u64 offset, u64 bytenr)
2998 {
2999         struct btrfs_root *extent_root = root->fs_info->extent_root;
3000         struct extent_buffer *leaf;
3001         struct btrfs_extent_data_ref *ref;
3002         struct btrfs_extent_inline_ref *iref;
3003         struct btrfs_extent_item *ei;
3004         struct btrfs_key key;
3005         u32 item_size;
3006         int ret;
3007
3008         key.objectid = bytenr;
3009         key.offset = (u64)-1;
3010         key.type = BTRFS_EXTENT_ITEM_KEY;
3011
3012         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3013         if (ret < 0)
3014                 goto out;
3015         BUG_ON(ret == 0); /* Corruption */
3016
3017         ret = -ENOENT;
3018         if (path->slots[0] == 0)
3019                 goto out;
3020
3021         path->slots[0]--;
3022         leaf = path->nodes[0];
3023         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3024
3025         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3026                 goto out;
3027
3028         ret = 1;
3029         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3030 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3031         if (item_size < sizeof(*ei)) {
3032                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3033                 goto out;
3034         }
3035 #endif
3036         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3037
3038         if (item_size != sizeof(*ei) +
3039             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3040                 goto out;
3041
3042         if (btrfs_extent_generation(leaf, ei) <=
3043             btrfs_root_last_snapshot(&root->root_item))
3044                 goto out;
3045
3046         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3047         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3048             BTRFS_EXTENT_DATA_REF_KEY)
3049                 goto out;
3050
3051         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3052         if (btrfs_extent_refs(leaf, ei) !=
3053             btrfs_extent_data_ref_count(leaf, ref) ||
3054             btrfs_extent_data_ref_root(leaf, ref) !=
3055             root->root_key.objectid ||
3056             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3057             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3058                 goto out;
3059
3060         ret = 0;
3061 out:
3062         return ret;
3063 }
3064
3065 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3066                           struct btrfs_root *root,
3067                           u64 objectid, u64 offset, u64 bytenr)
3068 {
3069         struct btrfs_path *path;
3070         int ret;
3071         int ret2;
3072
3073         path = btrfs_alloc_path();
3074         if (!path)
3075                 return -ENOENT;
3076
3077         do {
3078                 ret = check_committed_ref(trans, root, path, objectid,
3079                                           offset, bytenr);
3080                 if (ret && ret != -ENOENT)
3081                         goto out;
3082
3083                 ret2 = check_delayed_ref(trans, root, path, objectid,
3084                                          offset, bytenr);
3085         } while (ret2 == -EAGAIN);
3086
3087         if (ret2 && ret2 != -ENOENT) {
3088                 ret = ret2;
3089                 goto out;
3090         }
3091
3092         if (ret != -ENOENT || ret2 != -ENOENT)
3093                 ret = 0;
3094 out:
3095         btrfs_free_path(path);
3096         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3097                 WARN_ON(ret > 0);
3098         return ret;
3099 }
3100
3101 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3102                            struct btrfs_root *root,
3103                            struct extent_buffer *buf,
3104                            int full_backref, int inc)
3105 {
3106         u64 bytenr;
3107         u64 num_bytes;
3108         u64 parent;
3109         u64 ref_root;
3110         u32 nritems;
3111         struct btrfs_key key;
3112         struct btrfs_file_extent_item *fi;
3113         int i;
3114         int level;
3115         int ret = 0;
3116         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3117                             u64, u64, u64, u64, u64, u64, int);
3118
3119
3120         if (btrfs_test_is_dummy_root(root))
3121                 return 0;
3122
3123         ref_root = btrfs_header_owner(buf);
3124         nritems = btrfs_header_nritems(buf);
3125         level = btrfs_header_level(buf);
3126
3127         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3128                 return 0;
3129
3130         if (inc)
3131                 process_func = btrfs_inc_extent_ref;
3132         else
3133                 process_func = btrfs_free_extent;
3134
3135         if (full_backref)
3136                 parent = buf->start;
3137         else
3138                 parent = 0;
3139
3140         for (i = 0; i < nritems; i++) {
3141                 if (level == 0) {
3142                         btrfs_item_key_to_cpu(buf, &key, i);
3143                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3144                                 continue;
3145                         fi = btrfs_item_ptr(buf, i,
3146                                             struct btrfs_file_extent_item);
3147                         if (btrfs_file_extent_type(buf, fi) ==
3148                             BTRFS_FILE_EXTENT_INLINE)
3149                                 continue;
3150                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3151                         if (bytenr == 0)
3152                                 continue;
3153
3154                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3155                         key.offset -= btrfs_file_extent_offset(buf, fi);
3156                         ret = process_func(trans, root, bytenr, num_bytes,
3157                                            parent, ref_root, key.objectid,
3158                                            key.offset, 1);
3159                         if (ret)
3160                                 goto fail;
3161                 } else {
3162                         bytenr = btrfs_node_blockptr(buf, i);
3163                         num_bytes = root->nodesize;
3164                         ret = process_func(trans, root, bytenr, num_bytes,
3165                                            parent, ref_root, level - 1, 0,
3166                                            1);
3167                         if (ret)
3168                                 goto fail;
3169                 }
3170         }
3171         return 0;
3172 fail:
3173         return ret;
3174 }
3175
3176 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3177                   struct extent_buffer *buf, int full_backref)
3178 {
3179         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3180 }
3181
3182 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3183                   struct extent_buffer *buf, int full_backref)
3184 {
3185         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3186 }
3187
3188 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3189                                  struct btrfs_root *root,
3190                                  struct btrfs_path *path,
3191                                  struct btrfs_block_group_cache *cache)
3192 {
3193         int ret;
3194         struct btrfs_root *extent_root = root->fs_info->extent_root;
3195         unsigned long bi;
3196         struct extent_buffer *leaf;
3197
3198         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3199         if (ret) {
3200                 if (ret > 0)
3201                         ret = -ENOENT;
3202                 goto fail;
3203         }
3204
3205         leaf = path->nodes[0];
3206         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3207         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3208         btrfs_mark_buffer_dirty(leaf);
3209 fail:
3210         btrfs_release_path(path);
3211         return ret;
3212
3213 }
3214
3215 static struct btrfs_block_group_cache *
3216 next_block_group(struct btrfs_root *root,
3217                  struct btrfs_block_group_cache *cache)
3218 {
3219         struct rb_node *node;
3220
3221         spin_lock(&root->fs_info->block_group_cache_lock);
3222
3223         /* If our block group was removed, we need a full search. */
3224         if (RB_EMPTY_NODE(&cache->cache_node)) {
3225                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3226
3227                 spin_unlock(&root->fs_info->block_group_cache_lock);
3228                 btrfs_put_block_group(cache);
3229                 cache = btrfs_lookup_first_block_group(root->fs_info,
3230                                                        next_bytenr);
3231                 return cache;
3232         }
3233         node = rb_next(&cache->cache_node);
3234         btrfs_put_block_group(cache);
3235         if (node) {
3236                 cache = rb_entry(node, struct btrfs_block_group_cache,
3237                                  cache_node);
3238                 btrfs_get_block_group(cache);
3239         } else
3240                 cache = NULL;
3241         spin_unlock(&root->fs_info->block_group_cache_lock);
3242         return cache;
3243 }
3244
3245 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3246                             struct btrfs_trans_handle *trans,
3247                             struct btrfs_path *path)
3248 {
3249         struct btrfs_root *root = block_group->fs_info->tree_root;
3250         struct inode *inode = NULL;
3251         u64 alloc_hint = 0;
3252         int dcs = BTRFS_DC_ERROR;
3253         u64 num_pages = 0;
3254         int retries = 0;
3255         int ret = 0;
3256
3257         /*
3258          * If this block group is smaller than 100 megs don't bother caching the
3259          * block group.
3260          */
3261         if (block_group->key.offset < (100 * 1024 * 1024)) {
3262                 spin_lock(&block_group->lock);
3263                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3264                 spin_unlock(&block_group->lock);
3265                 return 0;
3266         }
3267
3268         if (trans->aborted)
3269                 return 0;
3270 again:
3271         inode = lookup_free_space_inode(root, block_group, path);
3272         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3273                 ret = PTR_ERR(inode);
3274                 btrfs_release_path(path);
3275                 goto out;
3276         }
3277
3278         if (IS_ERR(inode)) {
3279                 BUG_ON(retries);
3280                 retries++;
3281
3282                 if (block_group->ro)
3283                         goto out_free;
3284
3285                 ret = create_free_space_inode(root, trans, block_group, path);
3286                 if (ret)
3287                         goto out_free;
3288                 goto again;
3289         }
3290
3291         /* We've already setup this transaction, go ahead and exit */
3292         if (block_group->cache_generation == trans->transid &&
3293             i_size_read(inode)) {
3294                 dcs = BTRFS_DC_SETUP;
3295                 goto out_put;
3296         }
3297
3298         /*
3299          * We want to set the generation to 0, that way if anything goes wrong
3300          * from here on out we know not to trust this cache when we load up next
3301          * time.
3302          */
3303         BTRFS_I(inode)->generation = 0;
3304         ret = btrfs_update_inode(trans, root, inode);
3305         if (ret) {
3306                 /*
3307                  * So theoretically we could recover from this, simply set the
3308                  * super cache generation to 0 so we know to invalidate the
3309                  * cache, but then we'd have to keep track of the block groups
3310                  * that fail this way so we know we _have_ to reset this cache
3311                  * before the next commit or risk reading stale cache.  So to
3312                  * limit our exposure to horrible edge cases lets just abort the
3313                  * transaction, this only happens in really bad situations
3314                  * anyway.
3315                  */
3316                 btrfs_abort_transaction(trans, root, ret);
3317                 goto out_put;
3318         }
3319         WARN_ON(ret);
3320
3321         if (i_size_read(inode) > 0) {
3322                 ret = btrfs_check_trunc_cache_free_space(root,
3323                                         &root->fs_info->global_block_rsv);
3324                 if (ret)
3325                         goto out_put;
3326
3327                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3328                 if (ret)
3329                         goto out_put;
3330         }
3331
3332         spin_lock(&block_group->lock);
3333         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3334             !btrfs_test_opt(root, SPACE_CACHE)) {
3335                 /*
3336                  * don't bother trying to write stuff out _if_
3337                  * a) we're not cached,
3338                  * b) we're with nospace_cache mount option.
3339                  */
3340                 dcs = BTRFS_DC_WRITTEN;
3341                 spin_unlock(&block_group->lock);
3342                 goto out_put;
3343         }
3344         spin_unlock(&block_group->lock);
3345
3346         /*
3347          * Try to preallocate enough space based on how big the block group is.
3348          * Keep in mind this has to include any pinned space which could end up
3349          * taking up quite a bit since it's not folded into the other space
3350          * cache.
3351          */
3352         num_pages = div_u64(block_group->key.offset, 256 * 1024 * 1024);
3353         if (!num_pages)
3354                 num_pages = 1;
3355
3356         num_pages *= 16;
3357         num_pages *= PAGE_CACHE_SIZE;
3358
3359         ret = __btrfs_check_data_free_space(inode, 0, num_pages);
3360         if (ret)
3361                 goto out_put;
3362
3363         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3364                                               num_pages, num_pages,
3365                                               &alloc_hint);
3366         if (!ret)
3367                 dcs = BTRFS_DC_SETUP;
3368         __btrfs_free_reserved_data_space(inode, 0, num_pages);
3369
3370 out_put:
3371         iput(inode);
3372 out_free:
3373         btrfs_release_path(path);
3374 out:
3375         spin_lock(&block_group->lock);
3376         if (!ret && dcs == BTRFS_DC_SETUP)
3377                 block_group->cache_generation = trans->transid;
3378         block_group->disk_cache_state = dcs;
3379         spin_unlock(&block_group->lock);
3380
3381         return ret;
3382 }
3383
3384 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3385                             struct btrfs_root *root)
3386 {
3387         struct btrfs_block_group_cache *cache, *tmp;
3388         struct btrfs_transaction *cur_trans = trans->transaction;
3389         struct btrfs_path *path;
3390
3391         if (list_empty(&cur_trans->dirty_bgs) ||
3392             !btrfs_test_opt(root, SPACE_CACHE))
3393                 return 0;
3394
3395         path = btrfs_alloc_path();
3396         if (!path)
3397                 return -ENOMEM;
3398
3399         /* Could add new block groups, use _safe just in case */
3400         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3401                                  dirty_list) {
3402                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3403                         cache_save_setup(cache, trans, path);
3404         }
3405
3406         btrfs_free_path(path);
3407         return 0;
3408 }
3409
3410 /*
3411  * transaction commit does final block group cache writeback during a
3412  * critical section where nothing is allowed to change the FS.  This is
3413  * required in order for the cache to actually match the block group,
3414  * but can introduce a lot of latency into the commit.
3415  *
3416  * So, btrfs_start_dirty_block_groups is here to kick off block group
3417  * cache IO.  There's a chance we'll have to redo some of it if the
3418  * block group changes again during the commit, but it greatly reduces
3419  * the commit latency by getting rid of the easy block groups while
3420  * we're still allowing others to join the commit.
3421  */
3422 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3423                                    struct btrfs_root *root)
3424 {
3425         struct btrfs_block_group_cache *cache;
3426         struct btrfs_transaction *cur_trans = trans->transaction;
3427         int ret = 0;
3428         int should_put;
3429         struct btrfs_path *path = NULL;
3430         LIST_HEAD(dirty);
3431         struct list_head *io = &cur_trans->io_bgs;
3432         int num_started = 0;
3433         int loops = 0;
3434
3435         spin_lock(&cur_trans->dirty_bgs_lock);
3436         if (list_empty(&cur_trans->dirty_bgs)) {
3437                 spin_unlock(&cur_trans->dirty_bgs_lock);
3438                 return 0;
3439         }
3440         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3441         spin_unlock(&cur_trans->dirty_bgs_lock);
3442
3443 again:
3444         /*
3445          * make sure all the block groups on our dirty list actually
3446          * exist
3447          */
3448         btrfs_create_pending_block_groups(trans, root);
3449
3450         if (!path) {
3451                 path = btrfs_alloc_path();
3452                 if (!path)
3453                         return -ENOMEM;
3454         }
3455
3456         /*
3457          * cache_write_mutex is here only to save us from balance or automatic
3458          * removal of empty block groups deleting this block group while we are
3459          * writing out the cache
3460          */
3461         mutex_lock(&trans->transaction->cache_write_mutex);
3462         while (!list_empty(&dirty)) {
3463                 cache = list_first_entry(&dirty,
3464                                          struct btrfs_block_group_cache,
3465                                          dirty_list);
3466                 /*
3467                  * this can happen if something re-dirties a block
3468                  * group that is already under IO.  Just wait for it to
3469                  * finish and then do it all again
3470                  */
3471                 if (!list_empty(&cache->io_list)) {
3472                         list_del_init(&cache->io_list);
3473                         btrfs_wait_cache_io(root, trans, cache,
3474                                             &cache->io_ctl, path,
3475                                             cache->key.objectid);
3476                         btrfs_put_block_group(cache);
3477                 }
3478
3479
3480                 /*
3481                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3482                  * if it should update the cache_state.  Don't delete
3483                  * until after we wait.
3484                  *
3485                  * Since we're not running in the commit critical section
3486                  * we need the dirty_bgs_lock to protect from update_block_group
3487                  */
3488                 spin_lock(&cur_trans->dirty_bgs_lock);
3489                 list_del_init(&cache->dirty_list);
3490                 spin_unlock(&cur_trans->dirty_bgs_lock);
3491
3492                 should_put = 1;
3493
3494                 cache_save_setup(cache, trans, path);
3495
3496                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3497                         cache->io_ctl.inode = NULL;
3498                         ret = btrfs_write_out_cache(root, trans, cache, path);
3499                         if (ret == 0 && cache->io_ctl.inode) {
3500                                 num_started++;
3501                                 should_put = 0;
3502
3503                                 /*
3504                                  * the cache_write_mutex is protecting
3505                                  * the io_list
3506                                  */
3507                                 list_add_tail(&cache->io_list, io);
3508                         } else {
3509                                 /*
3510                                  * if we failed to write the cache, the
3511                                  * generation will be bad and life goes on
3512                                  */
3513                                 ret = 0;
3514                         }
3515                 }
3516                 if (!ret) {
3517                         ret = write_one_cache_group(trans, root, path, cache);
3518                         /*
3519                          * Our block group might still be attached to the list
3520                          * of new block groups in the transaction handle of some
3521                          * other task (struct btrfs_trans_handle->new_bgs). This
3522                          * means its block group item isn't yet in the extent
3523                          * tree. If this happens ignore the error, as we will
3524                          * try again later in the critical section of the
3525                          * transaction commit.
3526                          */
3527                         if (ret == -ENOENT) {
3528                                 ret = 0;
3529                                 spin_lock(&cur_trans->dirty_bgs_lock);
3530                                 if (list_empty(&cache->dirty_list)) {
3531                                         list_add_tail(&cache->dirty_list,
3532                                                       &cur_trans->dirty_bgs);
3533                                         btrfs_get_block_group(cache);
3534                                 }
3535                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3536                         } else if (ret) {
3537                                 btrfs_abort_transaction(trans, root, ret);
3538                         }
3539                 }
3540
3541                 /* if its not on the io list, we need to put the block group */
3542                 if (should_put)
3543                         btrfs_put_block_group(cache);
3544
3545                 if (ret)
3546                         break;
3547
3548                 /*
3549                  * Avoid blocking other tasks for too long. It might even save
3550                  * us from writing caches for block groups that are going to be
3551                  * removed.
3552                  */
3553                 mutex_unlock(&trans->transaction->cache_write_mutex);
3554                 mutex_lock(&trans->transaction->cache_write_mutex);
3555         }
3556         mutex_unlock(&trans->transaction->cache_write_mutex);
3557
3558         /*
3559          * go through delayed refs for all the stuff we've just kicked off
3560          * and then loop back (just once)
3561          */
3562         ret = btrfs_run_delayed_refs(trans, root, 0);
3563         if (!ret && loops == 0) {
3564                 loops++;
3565                 spin_lock(&cur_trans->dirty_bgs_lock);
3566                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3567                 /*
3568                  * dirty_bgs_lock protects us from concurrent block group
3569                  * deletes too (not just cache_write_mutex).
3570                  */
3571                 if (!list_empty(&dirty)) {
3572                         spin_unlock(&cur_trans->dirty_bgs_lock);
3573                         goto again;
3574                 }
3575                 spin_unlock(&cur_trans->dirty_bgs_lock);
3576         }
3577
3578         btrfs_free_path(path);
3579         return ret;
3580 }
3581
3582 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3583                                    struct btrfs_root *root)
3584 {
3585         struct btrfs_block_group_cache *cache;
3586         struct btrfs_transaction *cur_trans = trans->transaction;
3587         int ret = 0;
3588         int should_put;
3589         struct btrfs_path *path;
3590         struct list_head *io = &cur_trans->io_bgs;
3591         int num_started = 0;
3592
3593         path = btrfs_alloc_path();
3594         if (!path)
3595                 return -ENOMEM;
3596
3597         /*
3598          * We don't need the lock here since we are protected by the transaction
3599          * commit.  We want to do the cache_save_setup first and then run the
3600          * delayed refs to make sure we have the best chance at doing this all
3601          * in one shot.
3602          */
3603         while (!list_empty(&cur_trans->dirty_bgs)) {
3604                 cache = list_first_entry(&cur_trans->dirty_bgs,
3605                                          struct btrfs_block_group_cache,
3606                                          dirty_list);
3607
3608                 /*
3609                  * this can happen if cache_save_setup re-dirties a block
3610                  * group that is already under IO.  Just wait for it to
3611                  * finish and then do it all again
3612                  */
3613                 if (!list_empty(&cache->io_list)) {
3614                         list_del_init(&cache->io_list);
3615                         btrfs_wait_cache_io(root, trans, cache,
3616                                             &cache->io_ctl, path,
3617                                             cache->key.objectid);
3618                         btrfs_put_block_group(cache);
3619                 }
3620
3621                 /*
3622                  * don't remove from the dirty list until after we've waited
3623                  * on any pending IO
3624                  */
3625                 list_del_init(&cache->dirty_list);
3626                 should_put = 1;
3627
3628                 cache_save_setup(cache, trans, path);
3629
3630                 if (!ret)
3631                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3632
3633                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3634                         cache->io_ctl.inode = NULL;
3635                         ret = btrfs_write_out_cache(root, trans, cache, path);
3636                         if (ret == 0 && cache->io_ctl.inode) {
3637                                 num_started++;
3638                                 should_put = 0;
3639                                 list_add_tail(&cache->io_list, io);
3640                         } else {
3641                                 /*
3642                                  * if we failed to write the cache, the
3643                                  * generation will be bad and life goes on
3644                                  */
3645                                 ret = 0;
3646                         }
3647                 }
3648                 if (!ret) {
3649                         ret = write_one_cache_group(trans, root, path, cache);
3650                         if (ret)
3651                                 btrfs_abort_transaction(trans, root, ret);
3652                 }
3653
3654                 /* if its not on the io list, we need to put the block group */
3655                 if (should_put)
3656                         btrfs_put_block_group(cache);
3657         }
3658
3659         while (!list_empty(io)) {
3660                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3661                                          io_list);
3662                 list_del_init(&cache->io_list);
3663                 btrfs_wait_cache_io(root, trans, cache,
3664                                     &cache->io_ctl, path, cache->key.objectid);
3665                 btrfs_put_block_group(cache);
3666         }
3667
3668         btrfs_free_path(path);
3669         return ret;
3670 }
3671
3672 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3673 {
3674         struct btrfs_block_group_cache *block_group;
3675         int readonly = 0;
3676
3677         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3678         if (!block_group || block_group->ro)
3679                 readonly = 1;
3680         if (block_group)
3681                 btrfs_put_block_group(block_group);
3682         return readonly;
3683 }
3684
3685 static const char *alloc_name(u64 flags)
3686 {
3687         switch (flags) {
3688         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3689                 return "mixed";
3690         case BTRFS_BLOCK_GROUP_METADATA:
3691                 return "metadata";
3692         case BTRFS_BLOCK_GROUP_DATA:
3693                 return "data";
3694         case BTRFS_BLOCK_GROUP_SYSTEM:
3695                 return "system";
3696         default:
3697                 WARN_ON(1);
3698                 return "invalid-combination";
3699         };
3700 }
3701
3702 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3703                              u64 total_bytes, u64 bytes_used,
3704                              struct btrfs_space_info **space_info)
3705 {
3706         struct btrfs_space_info *found;
3707         int i;
3708         int factor;
3709         int ret;
3710
3711         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3712                      BTRFS_BLOCK_GROUP_RAID10))
3713                 factor = 2;
3714         else
3715                 factor = 1;
3716
3717         found = __find_space_info(info, flags);
3718         if (found) {
3719                 spin_lock(&found->lock);
3720                 found->total_bytes += total_bytes;
3721                 found->disk_total += total_bytes * factor;
3722                 found->bytes_used += bytes_used;
3723                 found->disk_used += bytes_used * factor;
3724                 if (total_bytes > 0)
3725                         found->full = 0;
3726                 spin_unlock(&found->lock);
3727                 *space_info = found;
3728                 return 0;
3729         }
3730         found = kzalloc(sizeof(*found), GFP_NOFS);
3731         if (!found)
3732                 return -ENOMEM;
3733
3734         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3735         if (ret) {
3736                 kfree(found);
3737                 return ret;
3738         }
3739
3740         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3741                 INIT_LIST_HEAD(&found->block_groups[i]);
3742         init_rwsem(&found->groups_sem);
3743         spin_lock_init(&found->lock);
3744         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3745         found->total_bytes = total_bytes;
3746         found->disk_total = total_bytes * factor;
3747         found->bytes_used = bytes_used;
3748         found->disk_used = bytes_used * factor;
3749         found->bytes_pinned = 0;
3750         found->bytes_reserved = 0;
3751         found->bytes_readonly = 0;
3752         found->bytes_may_use = 0;
3753         found->full = 0;
3754         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3755         found->chunk_alloc = 0;
3756         found->flush = 0;
3757         init_waitqueue_head(&found->wait);
3758         INIT_LIST_HEAD(&found->ro_bgs);
3759
3760         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3761                                     info->space_info_kobj, "%s",
3762                                     alloc_name(found->flags));
3763         if (ret) {
3764                 kfree(found);
3765                 return ret;
3766         }
3767
3768         *space_info = found;
3769         list_add_rcu(&found->list, &info->space_info);
3770         if (flags & BTRFS_BLOCK_GROUP_DATA)
3771                 info->data_sinfo = found;
3772
3773         return ret;
3774 }
3775
3776 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3777 {
3778         u64 extra_flags = chunk_to_extended(flags) &
3779                                 BTRFS_EXTENDED_PROFILE_MASK;
3780
3781         write_seqlock(&fs_info->profiles_lock);
3782         if (flags & BTRFS_BLOCK_GROUP_DATA)
3783                 fs_info->avail_data_alloc_bits |= extra_flags;
3784         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3785                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3786         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3787                 fs_info->avail_system_alloc_bits |= extra_flags;
3788         write_sequnlock(&fs_info->profiles_lock);
3789 }
3790
3791 /*
3792  * returns target flags in extended format or 0 if restripe for this
3793  * chunk_type is not in progress
3794  *
3795  * should be called with either volume_mutex or balance_lock held
3796  */
3797 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3798 {
3799         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3800         u64 target = 0;
3801
3802         if (!bctl)
3803                 return 0;
3804
3805         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3806             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3807                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3808         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3809                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3810                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3811         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3812                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3813                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3814         }
3815
3816         return target;
3817 }
3818
3819 /*
3820  * @flags: available profiles in extended format (see ctree.h)
3821  *
3822  * Returns reduced profile in chunk format.  If profile changing is in
3823  * progress (either running or paused) picks the target profile (if it's
3824  * already available), otherwise falls back to plain reducing.
3825  */
3826 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3827 {
3828         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3829         u64 target;
3830         u64 raid_type;
3831         u64 allowed = 0;
3832
3833         /*
3834          * see if restripe for this chunk_type is in progress, if so
3835          * try to reduce to the target profile
3836          */
3837         spin_lock(&root->fs_info->balance_lock);
3838         target = get_restripe_target(root->fs_info, flags);
3839         if (target) {
3840                 /* pick target profile only if it's already available */
3841                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3842                         spin_unlock(&root->fs_info->balance_lock);
3843                         return extended_to_chunk(target);
3844                 }
3845         }
3846         spin_unlock(&root->fs_info->balance_lock);
3847
3848         /* First, mask out the RAID levels which aren't possible */
3849         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
3850                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
3851                         allowed |= btrfs_raid_group[raid_type];
3852         }
3853         allowed &= flags;
3854
3855         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
3856                 allowed = BTRFS_BLOCK_GROUP_RAID6;
3857         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
3858                 allowed = BTRFS_BLOCK_GROUP_RAID5;
3859         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
3860                 allowed = BTRFS_BLOCK_GROUP_RAID10;
3861         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
3862                 allowed = BTRFS_BLOCK_GROUP_RAID1;
3863         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
3864                 allowed = BTRFS_BLOCK_GROUP_RAID0;
3865
3866         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
3867
3868         return extended_to_chunk(flags | allowed);
3869 }
3870
3871 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3872 {
3873         unsigned seq;
3874         u64 flags;
3875
3876         do {
3877                 flags = orig_flags;
3878                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3879
3880                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3881                         flags |= root->fs_info->avail_data_alloc_bits;
3882                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3883                         flags |= root->fs_info->avail_system_alloc_bits;
3884                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3885                         flags |= root->fs_info->avail_metadata_alloc_bits;
3886         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3887
3888         return btrfs_reduce_alloc_profile(root, flags);
3889 }
3890
3891 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3892 {
3893         u64 flags;
3894         u64 ret;
3895
3896         if (data)
3897                 flags = BTRFS_BLOCK_GROUP_DATA;
3898         else if (root == root->fs_info->chunk_root)
3899                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3900         else
3901                 flags = BTRFS_BLOCK_GROUP_METADATA;
3902
3903         ret = get_alloc_profile(root, flags);
3904         return ret;
3905 }
3906
3907 int btrfs_alloc_data_chunk_ondemand(struct inode *inode, u64 bytes)
3908 {
3909         struct btrfs_space_info *data_sinfo;
3910         struct btrfs_root *root = BTRFS_I(inode)->root;
3911         struct btrfs_fs_info *fs_info = root->fs_info;
3912         u64 used;
3913         int ret = 0;
3914         int need_commit = 2;
3915         int have_pinned_space;
3916
3917         /* make sure bytes are sectorsize aligned */
3918         bytes = ALIGN(bytes, root->sectorsize);
3919
3920         if (btrfs_is_free_space_inode(inode)) {
3921                 need_commit = 0;
3922                 ASSERT(current->journal_info);
3923         }
3924
3925         data_sinfo = fs_info->data_sinfo;
3926         if (!data_sinfo)
3927                 goto alloc;
3928
3929 again:
3930         /* make sure we have enough space to handle the data first */
3931         spin_lock(&data_sinfo->lock);
3932         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3933                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3934                 data_sinfo->bytes_may_use;
3935
3936         if (used + bytes > data_sinfo->total_bytes) {
3937                 struct btrfs_trans_handle *trans;
3938
3939                 /*
3940                  * if we don't have enough free bytes in this space then we need
3941                  * to alloc a new chunk.
3942                  */
3943                 if (!data_sinfo->full) {
3944                         u64 alloc_target;
3945
3946                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3947                         spin_unlock(&data_sinfo->lock);
3948 alloc:
3949                         alloc_target = btrfs_get_alloc_profile(root, 1);
3950                         /*
3951                          * It is ugly that we don't call nolock join
3952                          * transaction for the free space inode case here.
3953                          * But it is safe because we only do the data space
3954                          * reservation for the free space cache in the
3955                          * transaction context, the common join transaction
3956                          * just increase the counter of the current transaction
3957                          * handler, doesn't try to acquire the trans_lock of
3958                          * the fs.
3959                          */
3960                         trans = btrfs_join_transaction(root);
3961                         if (IS_ERR(trans))
3962                                 return PTR_ERR(trans);
3963
3964                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3965                                              alloc_target,
3966                                              CHUNK_ALLOC_NO_FORCE);
3967                         btrfs_end_transaction(trans, root);
3968                         if (ret < 0) {
3969                                 if (ret != -ENOSPC)
3970                                         return ret;
3971                                 else {
3972                                         have_pinned_space = 1;
3973                                         goto commit_trans;
3974                                 }
3975                         }
3976
3977                         if (!data_sinfo)
3978                                 data_sinfo = fs_info->data_sinfo;
3979
3980                         goto again;
3981                 }
3982
3983                 /*
3984                  * If we don't have enough pinned space to deal with this
3985                  * allocation, and no removed chunk in current transaction,
3986                  * don't bother committing the transaction.
3987                  */
3988                 have_pinned_space = percpu_counter_compare(
3989                         &data_sinfo->total_bytes_pinned,
3990                         used + bytes - data_sinfo->total_bytes);
3991                 spin_unlock(&data_sinfo->lock);
3992
3993                 /* commit the current transaction and try again */
3994 commit_trans:
3995                 if (need_commit &&
3996                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3997                         need_commit--;
3998
3999                         if (need_commit > 0)
4000                                 btrfs_wait_ordered_roots(fs_info, -1);
4001
4002                         trans = btrfs_join_transaction(root);
4003                         if (IS_ERR(trans))
4004                                 return PTR_ERR(trans);
4005                         if (have_pinned_space >= 0 ||
4006                             trans->transaction->have_free_bgs ||
4007                             need_commit > 0) {
4008                                 ret = btrfs_commit_transaction(trans, root);
4009                                 if (ret)
4010                                         return ret;
4011                                 /*
4012                                  * make sure that all running delayed iput are
4013                                  * done
4014                                  */
4015                                 down_write(&root->fs_info->delayed_iput_sem);
4016                                 up_write(&root->fs_info->delayed_iput_sem);
4017                                 goto again;
4018                         } else {
4019                                 btrfs_end_transaction(trans, root);
4020                         }
4021                 }
4022
4023                 trace_btrfs_space_reservation(root->fs_info,
4024                                               "space_info:enospc",
4025                                               data_sinfo->flags, bytes, 1);
4026                 return -ENOSPC;
4027         }
4028         data_sinfo->bytes_may_use += bytes;
4029         trace_btrfs_space_reservation(root->fs_info, "space_info",
4030                                       data_sinfo->flags, bytes, 1);
4031         spin_unlock(&data_sinfo->lock);
4032
4033         return ret;
4034 }
4035
4036 /*
4037  * This will check the space that the inode allocates from to make sure we have
4038  * enough space for bytes.
4039  */
4040 int btrfs_check_data_free_space(struct inode *inode, u64 bytes, u64 write_bytes)
4041 {
4042         struct btrfs_root *root = BTRFS_I(inode)->root;
4043         int ret;
4044
4045         ret = btrfs_alloc_data_chunk_ondemand(inode, bytes);
4046         if (ret < 0)
4047                 return ret;
4048         ret = btrfs_qgroup_reserve(root, write_bytes);
4049         return ret;
4050 }
4051
4052 /*
4053  * New check_data_free_space() with ability for precious data reservation
4054  * Will replace old btrfs_check_data_free_space(), but for patch split,
4055  * add a new function first and then replace it.
4056  */
4057 int __btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4058 {
4059         struct btrfs_root *root = BTRFS_I(inode)->root;
4060         int ret;
4061
4062         /* align the range */
4063         len = round_up(start + len, root->sectorsize) -
4064               round_down(start, root->sectorsize);
4065         start = round_down(start, root->sectorsize);
4066
4067         ret = btrfs_alloc_data_chunk_ondemand(inode, len);
4068         if (ret < 0)
4069                 return ret;
4070
4071         /* Use new btrfs_qgroup_reserve_data to reserve precious data space */
4072         ret = btrfs_qgroup_reserve_data(inode, start, len);
4073         return ret;
4074 }
4075
4076 /*
4077  * Called if we need to clear a data reservation for this inode.
4078  */
4079 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
4080 {
4081         struct btrfs_root *root = BTRFS_I(inode)->root;
4082         struct btrfs_space_info *data_sinfo;
4083
4084         /* make sure bytes are sectorsize aligned */
4085         bytes = ALIGN(bytes, root->sectorsize);
4086
4087         data_sinfo = root->fs_info->data_sinfo;
4088         spin_lock(&data_sinfo->lock);
4089         WARN_ON(data_sinfo->bytes_may_use < bytes);
4090         data_sinfo->bytes_may_use -= bytes;
4091         trace_btrfs_space_reservation(root->fs_info, "space_info",
4092                                       data_sinfo->flags, bytes, 0);
4093         spin_unlock(&data_sinfo->lock);
4094 }
4095
4096 /*
4097  * Called if we need to clear a data reservation for this inode
4098  * Normally in a error case.
4099  *
4100  * This one will handle the per-indoe data rsv map for accurate reserved
4101  * space framework.
4102  */
4103 void __btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4104 {
4105         struct btrfs_root *root = BTRFS_I(inode)->root;
4106         struct btrfs_space_info *data_sinfo;
4107
4108         /* Make sure the range is aligned to sectorsize */
4109         len = round_up(start + len, root->sectorsize) -
4110               round_down(start, root->sectorsize);
4111         start = round_down(start, root->sectorsize);
4112
4113         /*
4114          * Free any reserved qgroup data space first
4115          * As it will alloc memory, we can't do it with data sinfo
4116          * spinlock hold.
4117          */
4118         btrfs_qgroup_free_data(inode, start, len);
4119
4120         data_sinfo = root->fs_info->data_sinfo;
4121         spin_lock(&data_sinfo->lock);
4122         if (WARN_ON(data_sinfo->bytes_may_use < len))
4123                 data_sinfo->bytes_may_use = 0;
4124         else
4125                 data_sinfo->bytes_may_use -= len;
4126         trace_btrfs_space_reservation(root->fs_info, "space_info",
4127                                       data_sinfo->flags, len, 0);
4128         spin_unlock(&data_sinfo->lock);
4129 }
4130
4131 static void force_metadata_allocation(struct btrfs_fs_info *info)
4132 {
4133         struct list_head *head = &info->space_info;
4134         struct btrfs_space_info *found;
4135
4136         rcu_read_lock();
4137         list_for_each_entry_rcu(found, head, list) {
4138                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4139                         found->force_alloc = CHUNK_ALLOC_FORCE;
4140         }
4141         rcu_read_unlock();
4142 }
4143
4144 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4145 {
4146         return (global->size << 1);
4147 }
4148
4149 static int should_alloc_chunk(struct btrfs_root *root,
4150                               struct btrfs_space_info *sinfo, int force)
4151 {
4152         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4153         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4154         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4155         u64 thresh;
4156
4157         if (force == CHUNK_ALLOC_FORCE)
4158                 return 1;
4159
4160         /*
4161          * We need to take into account the global rsv because for all intents
4162          * and purposes it's used space.  Don't worry about locking the
4163          * global_rsv, it doesn't change except when the transaction commits.
4164          */
4165         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4166                 num_allocated += calc_global_rsv_need_space(global_rsv);
4167
4168         /*
4169          * in limited mode, we want to have some free space up to
4170          * about 1% of the FS size.
4171          */
4172         if (force == CHUNK_ALLOC_LIMITED) {
4173                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4174                 thresh = max_t(u64, 64 * 1024 * 1024,
4175                                div_factor_fine(thresh, 1));
4176
4177                 if (num_bytes - num_allocated < thresh)
4178                         return 1;
4179         }
4180
4181         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
4182                 return 0;
4183         return 1;
4184 }
4185
4186 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4187 {
4188         u64 num_dev;
4189
4190         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4191                     BTRFS_BLOCK_GROUP_RAID0 |
4192                     BTRFS_BLOCK_GROUP_RAID5 |
4193                     BTRFS_BLOCK_GROUP_RAID6))
4194                 num_dev = root->fs_info->fs_devices->rw_devices;
4195         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4196                 num_dev = 2;
4197         else
4198                 num_dev = 1;    /* DUP or single */
4199
4200         return num_dev;
4201 }
4202
4203 /*
4204  * If @is_allocation is true, reserve space in the system space info necessary
4205  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4206  * removing a chunk.
4207  */
4208 void check_system_chunk(struct btrfs_trans_handle *trans,
4209                         struct btrfs_root *root,
4210                         u64 type)
4211 {
4212         struct btrfs_space_info *info;
4213         u64 left;
4214         u64 thresh;
4215         int ret = 0;
4216         u64 num_devs;
4217
4218         /*
4219          * Needed because we can end up allocating a system chunk and for an
4220          * atomic and race free space reservation in the chunk block reserve.
4221          */
4222         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4223
4224         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4225         spin_lock(&info->lock);
4226         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4227                 info->bytes_reserved - info->bytes_readonly -
4228                 info->bytes_may_use;
4229         spin_unlock(&info->lock);
4230
4231         num_devs = get_profile_num_devs(root, type);
4232
4233         /* num_devs device items to update and 1 chunk item to add or remove */
4234         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4235                 btrfs_calc_trans_metadata_size(root, 1);
4236
4237         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4238                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4239                         left, thresh, type);
4240                 dump_space_info(info, 0, 0);
4241         }
4242
4243         if (left < thresh) {
4244                 u64 flags;
4245
4246                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4247                 /*
4248                  * Ignore failure to create system chunk. We might end up not
4249                  * needing it, as we might not need to COW all nodes/leafs from
4250                  * the paths we visit in the chunk tree (they were already COWed
4251                  * or created in the current transaction for example).
4252                  */
4253                 ret = btrfs_alloc_chunk(trans, root, flags);
4254         }
4255
4256         if (!ret) {
4257                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4258                                           &root->fs_info->chunk_block_rsv,
4259                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4260                 if (!ret)
4261                         trans->chunk_bytes_reserved += thresh;
4262         }
4263 }
4264
4265 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4266                           struct btrfs_root *extent_root, u64 flags, int force)
4267 {
4268         struct btrfs_space_info *space_info;
4269         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4270         int wait_for_alloc = 0;
4271         int ret = 0;
4272
4273         /* Don't re-enter if we're already allocating a chunk */
4274         if (trans->allocating_chunk)
4275                 return -ENOSPC;
4276
4277         space_info = __find_space_info(extent_root->fs_info, flags);
4278         if (!space_info) {
4279                 ret = update_space_info(extent_root->fs_info, flags,
4280                                         0, 0, &space_info);
4281                 BUG_ON(ret); /* -ENOMEM */
4282         }
4283         BUG_ON(!space_info); /* Logic error */
4284
4285 again:
4286         spin_lock(&space_info->lock);
4287         if (force < space_info->force_alloc)
4288                 force = space_info->force_alloc;
4289         if (space_info->full) {
4290                 if (should_alloc_chunk(extent_root, space_info, force))
4291                         ret = -ENOSPC;
4292                 else
4293                         ret = 0;
4294                 spin_unlock(&space_info->lock);
4295                 return ret;
4296         }
4297
4298         if (!should_alloc_chunk(extent_root, space_info, force)) {
4299                 spin_unlock(&space_info->lock);
4300                 return 0;
4301         } else if (space_info->chunk_alloc) {
4302                 wait_for_alloc = 1;
4303         } else {
4304                 space_info->chunk_alloc = 1;
4305         }
4306
4307         spin_unlock(&space_info->lock);
4308
4309         mutex_lock(&fs_info->chunk_mutex);
4310
4311         /*
4312          * The chunk_mutex is held throughout the entirety of a chunk
4313          * allocation, so once we've acquired the chunk_mutex we know that the
4314          * other guy is done and we need to recheck and see if we should
4315          * allocate.
4316          */
4317         if (wait_for_alloc) {
4318                 mutex_unlock(&fs_info->chunk_mutex);
4319                 wait_for_alloc = 0;
4320                 goto again;
4321         }
4322
4323         trans->allocating_chunk = true;
4324
4325         /*
4326          * If we have mixed data/metadata chunks we want to make sure we keep
4327          * allocating mixed chunks instead of individual chunks.
4328          */
4329         if (btrfs_mixed_space_info(space_info))
4330                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4331
4332         /*
4333          * if we're doing a data chunk, go ahead and make sure that
4334          * we keep a reasonable number of metadata chunks allocated in the
4335          * FS as well.
4336          */
4337         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4338                 fs_info->data_chunk_allocations++;
4339                 if (!(fs_info->data_chunk_allocations %
4340                       fs_info->metadata_ratio))
4341                         force_metadata_allocation(fs_info);
4342         }
4343
4344         /*
4345          * Check if we have enough space in SYSTEM chunk because we may need
4346          * to update devices.
4347          */
4348         check_system_chunk(trans, extent_root, flags);
4349
4350         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4351         trans->allocating_chunk = false;
4352
4353         spin_lock(&space_info->lock);
4354         if (ret < 0 && ret != -ENOSPC)
4355                 goto out;
4356         if (ret)
4357                 space_info->full = 1;
4358         else
4359                 ret = 1;
4360
4361         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4362 out:
4363         space_info->chunk_alloc = 0;
4364         spin_unlock(&space_info->lock);
4365         mutex_unlock(&fs_info->chunk_mutex);
4366         /*
4367          * When we allocate a new chunk we reserve space in the chunk block
4368          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4369          * add new nodes/leafs to it if we end up needing to do it when
4370          * inserting the chunk item and updating device items as part of the
4371          * second phase of chunk allocation, performed by
4372          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4373          * large number of new block groups to create in our transaction
4374          * handle's new_bgs list to avoid exhausting the chunk block reserve
4375          * in extreme cases - like having a single transaction create many new
4376          * block groups when starting to write out the free space caches of all
4377          * the block groups that were made dirty during the lifetime of the
4378          * transaction.
4379          */
4380         if (trans->can_flush_pending_bgs &&
4381             trans->chunk_bytes_reserved >= (2 * 1024 * 1024ull)) {
4382                 btrfs_create_pending_block_groups(trans, trans->root);
4383                 btrfs_trans_release_chunk_metadata(trans);
4384         }
4385         return ret;
4386 }
4387
4388 static int can_overcommit(struct btrfs_root *root,
4389                           struct btrfs_space_info *space_info, u64 bytes,
4390                           enum btrfs_reserve_flush_enum flush)
4391 {
4392         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4393         u64 profile = btrfs_get_alloc_profile(root, 0);
4394         u64 space_size;
4395         u64 avail;
4396         u64 used;
4397
4398         used = space_info->bytes_used + space_info->bytes_reserved +
4399                 space_info->bytes_pinned + space_info->bytes_readonly;
4400
4401         /*
4402          * We only want to allow over committing if we have lots of actual space
4403          * free, but if we don't have enough space to handle the global reserve
4404          * space then we could end up having a real enospc problem when trying
4405          * to allocate a chunk or some other such important allocation.
4406          */
4407         spin_lock(&global_rsv->lock);
4408         space_size = calc_global_rsv_need_space(global_rsv);
4409         spin_unlock(&global_rsv->lock);
4410         if (used + space_size >= space_info->total_bytes)
4411                 return 0;
4412
4413         used += space_info->bytes_may_use;
4414
4415         spin_lock(&root->fs_info->free_chunk_lock);
4416         avail = root->fs_info->free_chunk_space;
4417         spin_unlock(&root->fs_info->free_chunk_lock);
4418
4419         /*
4420          * If we have dup, raid1 or raid10 then only half of the free
4421          * space is actually useable.  For raid56, the space info used
4422          * doesn't include the parity drive, so we don't have to
4423          * change the math
4424          */
4425         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4426                        BTRFS_BLOCK_GROUP_RAID1 |
4427                        BTRFS_BLOCK_GROUP_RAID10))
4428                 avail >>= 1;
4429
4430         /*
4431          * If we aren't flushing all things, let us overcommit up to
4432          * 1/2th of the space. If we can flush, don't let us overcommit
4433          * too much, let it overcommit up to 1/8 of the space.
4434          */
4435         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4436                 avail >>= 3;
4437         else
4438                 avail >>= 1;
4439
4440         if (used + bytes < space_info->total_bytes + avail)
4441                 return 1;
4442         return 0;
4443 }
4444
4445 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4446                                          unsigned long nr_pages, int nr_items)
4447 {
4448         struct super_block *sb = root->fs_info->sb;
4449
4450         if (down_read_trylock(&sb->s_umount)) {
4451                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4452                 up_read(&sb->s_umount);
4453         } else {
4454                 /*
4455                  * We needn't worry the filesystem going from r/w to r/o though
4456                  * we don't acquire ->s_umount mutex, because the filesystem
4457                  * should guarantee the delalloc inodes list be empty after
4458                  * the filesystem is readonly(all dirty pages are written to
4459                  * the disk).
4460                  */
4461                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4462                 if (!current->journal_info)
4463                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4464         }
4465 }
4466
4467 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4468 {
4469         u64 bytes;
4470         int nr;
4471
4472         bytes = btrfs_calc_trans_metadata_size(root, 1);
4473         nr = (int)div64_u64(to_reclaim, bytes);
4474         if (!nr)
4475                 nr = 1;
4476         return nr;
4477 }
4478
4479 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4480
4481 /*
4482  * shrink metadata reservation for delalloc
4483  */
4484 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4485                             bool wait_ordered)
4486 {
4487         struct btrfs_block_rsv *block_rsv;
4488         struct btrfs_space_info *space_info;
4489         struct btrfs_trans_handle *trans;
4490         u64 delalloc_bytes;
4491         u64 max_reclaim;
4492         long time_left;
4493         unsigned long nr_pages;
4494         int loops;
4495         int items;
4496         enum btrfs_reserve_flush_enum flush;
4497
4498         /* Calc the number of the pages we need flush for space reservation */
4499         items = calc_reclaim_items_nr(root, to_reclaim);
4500         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4501
4502         trans = (struct btrfs_trans_handle *)current->journal_info;
4503         block_rsv = &root->fs_info->delalloc_block_rsv;
4504         space_info = block_rsv->space_info;
4505
4506         delalloc_bytes = percpu_counter_sum_positive(
4507                                                 &root->fs_info->delalloc_bytes);
4508         if (delalloc_bytes == 0) {
4509                 if (trans)
4510                         return;
4511                 if (wait_ordered)
4512                         btrfs_wait_ordered_roots(root->fs_info, items);
4513                 return;
4514         }
4515
4516         loops = 0;
4517         while (delalloc_bytes && loops < 3) {
4518                 max_reclaim = min(delalloc_bytes, to_reclaim);
4519                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4520                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4521                 /*
4522                  * We need to wait for the async pages to actually start before
4523                  * we do anything.
4524                  */
4525                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4526                 if (!max_reclaim)
4527                         goto skip_async;
4528
4529                 if (max_reclaim <= nr_pages)
4530                         max_reclaim = 0;
4531                 else
4532                         max_reclaim -= nr_pages;
4533
4534                 wait_event(root->fs_info->async_submit_wait,
4535                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4536                            (int)max_reclaim);
4537 skip_async:
4538                 if (!trans)
4539                         flush = BTRFS_RESERVE_FLUSH_ALL;
4540                 else
4541                         flush = BTRFS_RESERVE_NO_FLUSH;
4542                 spin_lock(&space_info->lock);
4543                 if (can_overcommit(root, space_info, orig, flush)) {
4544                         spin_unlock(&space_info->lock);
4545                         break;
4546                 }
4547                 spin_unlock(&space_info->lock);
4548
4549                 loops++;
4550                 if (wait_ordered && !trans) {
4551                         btrfs_wait_ordered_roots(root->fs_info, items);
4552                 } else {
4553                         time_left = schedule_timeout_killable(1);
4554                         if (time_left)
4555                                 break;
4556                 }
4557                 delalloc_bytes = percpu_counter_sum_positive(
4558                                                 &root->fs_info->delalloc_bytes);
4559         }
4560 }
4561
4562 /**
4563  * maybe_commit_transaction - possibly commit the transaction if its ok to
4564  * @root - the root we're allocating for
4565  * @bytes - the number of bytes we want to reserve
4566  * @force - force the commit
4567  *
4568  * This will check to make sure that committing the transaction will actually
4569  * get us somewhere and then commit the transaction if it does.  Otherwise it
4570  * will return -ENOSPC.
4571  */
4572 static int may_commit_transaction(struct btrfs_root *root,
4573                                   struct btrfs_space_info *space_info,
4574                                   u64 bytes, int force)
4575 {
4576         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4577         struct btrfs_trans_handle *trans;
4578
4579         trans = (struct btrfs_trans_handle *)current->journal_info;
4580         if (trans)
4581                 return -EAGAIN;
4582
4583         if (force)
4584                 goto commit;
4585
4586         /* See if there is enough pinned space to make this reservation */
4587         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4588                                    bytes) >= 0)
4589                 goto commit;
4590
4591         /*
4592          * See if there is some space in the delayed insertion reservation for
4593          * this reservation.
4594          */
4595         if (space_info != delayed_rsv->space_info)
4596                 return -ENOSPC;
4597
4598         spin_lock(&delayed_rsv->lock);
4599         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4600                                    bytes - delayed_rsv->size) >= 0) {
4601                 spin_unlock(&delayed_rsv->lock);
4602                 return -ENOSPC;
4603         }
4604         spin_unlock(&delayed_rsv->lock);
4605
4606 commit:
4607         trans = btrfs_join_transaction(root);
4608         if (IS_ERR(trans))
4609                 return -ENOSPC;
4610
4611         return btrfs_commit_transaction(trans, root);
4612 }
4613
4614 enum flush_state {
4615         FLUSH_DELAYED_ITEMS_NR  =       1,
4616         FLUSH_DELAYED_ITEMS     =       2,
4617         FLUSH_DELALLOC          =       3,
4618         FLUSH_DELALLOC_WAIT     =       4,
4619         ALLOC_CHUNK             =       5,
4620         COMMIT_TRANS            =       6,
4621 };
4622
4623 static int flush_space(struct btrfs_root *root,
4624                        struct btrfs_space_info *space_info, u64 num_bytes,
4625                        u64 orig_bytes, int state)
4626 {
4627         struct btrfs_trans_handle *trans;
4628         int nr;
4629         int ret = 0;
4630
4631         switch (state) {
4632         case FLUSH_DELAYED_ITEMS_NR:
4633         case FLUSH_DELAYED_ITEMS:
4634                 if (state == FLUSH_DELAYED_ITEMS_NR)
4635                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4636                 else
4637                         nr = -1;
4638
4639                 trans = btrfs_join_transaction(root);
4640                 if (IS_ERR(trans)) {
4641                         ret = PTR_ERR(trans);
4642                         break;
4643                 }
4644                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4645                 btrfs_end_transaction(trans, root);
4646                 break;
4647         case FLUSH_DELALLOC:
4648         case FLUSH_DELALLOC_WAIT:
4649                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4650                                 state == FLUSH_DELALLOC_WAIT);
4651                 break;
4652         case ALLOC_CHUNK:
4653                 trans = btrfs_join_transaction(root);
4654                 if (IS_ERR(trans)) {
4655                         ret = PTR_ERR(trans);
4656                         break;
4657                 }
4658                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4659                                      btrfs_get_alloc_profile(root, 0),
4660                                      CHUNK_ALLOC_NO_FORCE);
4661                 btrfs_end_transaction(trans, root);
4662                 if (ret == -ENOSPC)
4663                         ret = 0;
4664                 break;
4665         case COMMIT_TRANS:
4666                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4667                 break;
4668         default:
4669                 ret = -ENOSPC;
4670                 break;
4671         }
4672
4673         return ret;
4674 }
4675
4676 static inline u64
4677 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4678                                  struct btrfs_space_info *space_info)
4679 {
4680         u64 used;
4681         u64 expected;
4682         u64 to_reclaim;
4683
4684         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4685                                 16 * 1024 * 1024);
4686         spin_lock(&space_info->lock);
4687         if (can_overcommit(root, space_info, to_reclaim,
4688                            BTRFS_RESERVE_FLUSH_ALL)) {
4689                 to_reclaim = 0;
4690                 goto out;
4691         }
4692
4693         used = space_info->bytes_used + space_info->bytes_reserved +
4694                space_info->bytes_pinned + space_info->bytes_readonly +
4695                space_info->bytes_may_use;
4696         if (can_overcommit(root, space_info, 1024 * 1024,
4697                            BTRFS_RESERVE_FLUSH_ALL))
4698                 expected = div_factor_fine(space_info->total_bytes, 95);
4699         else
4700                 expected = div_factor_fine(space_info->total_bytes, 90);
4701
4702         if (used > expected)
4703                 to_reclaim = used - expected;
4704         else
4705                 to_reclaim = 0;
4706         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4707                                      space_info->bytes_reserved);
4708 out:
4709         spin_unlock(&space_info->lock);
4710
4711         return to_reclaim;
4712 }
4713
4714 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4715                                         struct btrfs_fs_info *fs_info, u64 used)
4716 {
4717         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4718
4719         /* If we're just plain full then async reclaim just slows us down. */
4720         if (space_info->bytes_used >= thresh)
4721                 return 0;
4722
4723         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4724                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4725 }
4726
4727 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4728                                        struct btrfs_fs_info *fs_info,
4729                                        int flush_state)
4730 {
4731         u64 used;
4732
4733         spin_lock(&space_info->lock);
4734         /*
4735          * We run out of space and have not got any free space via flush_space,
4736          * so don't bother doing async reclaim.
4737          */
4738         if (flush_state > COMMIT_TRANS && space_info->full) {
4739                 spin_unlock(&space_info->lock);
4740                 return 0;
4741         }
4742
4743         used = space_info->bytes_used + space_info->bytes_reserved +
4744                space_info->bytes_pinned + space_info->bytes_readonly +
4745                space_info->bytes_may_use;
4746         if (need_do_async_reclaim(space_info, fs_info, used)) {
4747                 spin_unlock(&space_info->lock);
4748                 return 1;
4749         }
4750         spin_unlock(&space_info->lock);
4751
4752         return 0;
4753 }
4754
4755 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4756 {
4757         struct btrfs_fs_info *fs_info;
4758         struct btrfs_space_info *space_info;
4759         u64 to_reclaim;
4760         int flush_state;
4761
4762         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4763         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4764
4765         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4766                                                       space_info);
4767         if (!to_reclaim)
4768                 return;
4769
4770         flush_state = FLUSH_DELAYED_ITEMS_NR;
4771         do {
4772                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4773                             to_reclaim, flush_state);
4774                 flush_state++;
4775                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4776                                                  flush_state))
4777                         return;
4778         } while (flush_state < COMMIT_TRANS);
4779 }
4780
4781 void btrfs_init_async_reclaim_work(struct work_struct *work)
4782 {
4783         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4784 }
4785
4786 /**
4787  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4788  * @root - the root we're allocating for
4789  * @block_rsv - the block_rsv we're allocating for
4790  * @orig_bytes - the number of bytes we want
4791  * @flush - whether or not we can flush to make our reservation
4792  *
4793  * This will reserve orgi_bytes number of bytes from the space info associated
4794  * with the block_rsv.  If there is not enough space it will make an attempt to
4795  * flush out space to make room.  It will do this by flushing delalloc if
4796  * possible or committing the transaction.  If flush is 0 then no attempts to
4797  * regain reservations will be made and this will fail if there is not enough
4798  * space already.
4799  */
4800 static int reserve_metadata_bytes(struct btrfs_root *root,
4801                                   struct btrfs_block_rsv *block_rsv,
4802                                   u64 orig_bytes,
4803                                   enum btrfs_reserve_flush_enum flush)
4804 {
4805         struct btrfs_space_info *space_info = block_rsv->space_info;
4806         u64 used;
4807         u64 num_bytes = orig_bytes;
4808         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4809         int ret = 0;
4810         bool flushing = false;
4811
4812 again:
4813         ret = 0;
4814         spin_lock(&space_info->lock);
4815         /*
4816          * We only want to wait if somebody other than us is flushing and we
4817          * are actually allowed to flush all things.
4818          */
4819         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4820                space_info->flush) {
4821                 spin_unlock(&space_info->lock);
4822                 /*
4823                  * If we have a trans handle we can't wait because the flusher
4824                  * may have to commit the transaction, which would mean we would
4825                  * deadlock since we are waiting for the flusher to finish, but
4826                  * hold the current transaction open.
4827                  */
4828                 if (current->journal_info)
4829                         return -EAGAIN;
4830                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4831                 /* Must have been killed, return */
4832                 if (ret)
4833                         return -EINTR;
4834
4835                 spin_lock(&space_info->lock);
4836         }
4837
4838         ret = -ENOSPC;
4839         used = space_info->bytes_used + space_info->bytes_reserved +
4840                 space_info->bytes_pinned + space_info->bytes_readonly +
4841                 space_info->bytes_may_use;
4842
4843         /*
4844          * The idea here is that we've not already over-reserved the block group
4845          * then we can go ahead and save our reservation first and then start
4846          * flushing if we need to.  Otherwise if we've already overcommitted
4847          * lets start flushing stuff first and then come back and try to make
4848          * our reservation.
4849          */
4850         if (used <= space_info->total_bytes) {
4851                 if (used + orig_bytes <= space_info->total_bytes) {
4852                         space_info->bytes_may_use += orig_bytes;
4853                         trace_btrfs_space_reservation(root->fs_info,
4854                                 "space_info", space_info->flags, orig_bytes, 1);
4855                         ret = 0;
4856                 } else {
4857                         /*
4858                          * Ok set num_bytes to orig_bytes since we aren't
4859                          * overocmmitted, this way we only try and reclaim what
4860                          * we need.
4861                          */
4862                         num_bytes = orig_bytes;
4863                 }
4864         } else {
4865                 /*
4866                  * Ok we're over committed, set num_bytes to the overcommitted
4867                  * amount plus the amount of bytes that we need for this
4868                  * reservation.
4869                  */
4870                 num_bytes = used - space_info->total_bytes +
4871                         (orig_bytes * 2);
4872         }
4873
4874         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4875                 space_info->bytes_may_use += orig_bytes;
4876                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4877                                               space_info->flags, orig_bytes,
4878                                               1);
4879                 ret = 0;
4880         }
4881
4882         /*
4883          * Couldn't make our reservation, save our place so while we're trying
4884          * to reclaim space we can actually use it instead of somebody else
4885          * stealing it from us.
4886          *
4887          * We make the other tasks wait for the flush only when we can flush
4888          * all things.
4889          */
4890         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4891                 flushing = true;
4892                 space_info->flush = 1;
4893         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4894                 used += orig_bytes;
4895                 /*
4896                  * We will do the space reservation dance during log replay,
4897                  * which means we won't have fs_info->fs_root set, so don't do
4898                  * the async reclaim as we will panic.
4899                  */
4900                 if (!root->fs_info->log_root_recovering &&
4901                     need_do_async_reclaim(space_info, root->fs_info, used) &&
4902                     !work_busy(&root->fs_info->async_reclaim_work))
4903                         queue_work(system_unbound_wq,
4904                                    &root->fs_info->async_reclaim_work);
4905         }
4906         spin_unlock(&space_info->lock);
4907
4908         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4909                 goto out;
4910
4911         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4912                           flush_state);
4913         flush_state++;
4914
4915         /*
4916          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4917          * would happen. So skip delalloc flush.
4918          */
4919         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4920             (flush_state == FLUSH_DELALLOC ||
4921              flush_state == FLUSH_DELALLOC_WAIT))
4922                 flush_state = ALLOC_CHUNK;
4923
4924         if (!ret)
4925                 goto again;
4926         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4927                  flush_state < COMMIT_TRANS)
4928                 goto again;
4929         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4930                  flush_state <= COMMIT_TRANS)
4931                 goto again;
4932
4933 out:
4934         if (ret == -ENOSPC &&
4935             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4936                 struct btrfs_block_rsv *global_rsv =
4937                         &root->fs_info->global_block_rsv;
4938
4939                 if (block_rsv != global_rsv &&
4940                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4941                         ret = 0;
4942         }
4943         if (ret == -ENOSPC)
4944                 trace_btrfs_space_reservation(root->fs_info,
4945                                               "space_info:enospc",
4946                                               space_info->flags, orig_bytes, 1);
4947         if (flushing) {
4948                 spin_lock(&space_info->lock);
4949                 space_info->flush = 0;
4950                 wake_up_all(&space_info->wait);
4951                 spin_unlock(&space_info->lock);
4952         }
4953         return ret;
4954 }
4955
4956 static struct btrfs_block_rsv *get_block_rsv(
4957                                         const struct btrfs_trans_handle *trans,
4958                                         const struct btrfs_root *root)
4959 {
4960         struct btrfs_block_rsv *block_rsv = NULL;
4961
4962         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4963             (root == root->fs_info->csum_root && trans->adding_csums) ||
4964              (root == root->fs_info->uuid_root))
4965                 block_rsv = trans->block_rsv;
4966
4967         if (!block_rsv)
4968                 block_rsv = root->block_rsv;
4969
4970         if (!block_rsv)
4971                 block_rsv = &root->fs_info->empty_block_rsv;
4972
4973         return block_rsv;
4974 }
4975
4976 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4977                                u64 num_bytes)
4978 {
4979         int ret = -ENOSPC;
4980         spin_lock(&block_rsv->lock);
4981         if (block_rsv->reserved >= num_bytes) {
4982                 block_rsv->reserved -= num_bytes;
4983                 if (block_rsv->reserved < block_rsv->size)
4984                         block_rsv->full = 0;
4985                 ret = 0;
4986         }
4987         spin_unlock(&block_rsv->lock);
4988         return ret;
4989 }
4990
4991 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4992                                 u64 num_bytes, int update_size)
4993 {
4994         spin_lock(&block_rsv->lock);
4995         block_rsv->reserved += num_bytes;
4996         if (update_size)
4997                 block_rsv->size += num_bytes;
4998         else if (block_rsv->reserved >= block_rsv->size)
4999                 block_rsv->full = 1;
5000         spin_unlock(&block_rsv->lock);
5001 }
5002
5003 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5004                              struct btrfs_block_rsv *dest, u64 num_bytes,
5005                              int min_factor)
5006 {
5007         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5008         u64 min_bytes;
5009
5010         if (global_rsv->space_info != dest->space_info)
5011                 return -ENOSPC;
5012
5013         spin_lock(&global_rsv->lock);
5014         min_bytes = div_factor(global_rsv->size, min_factor);
5015         if (global_rsv->reserved < min_bytes + num_bytes) {
5016                 spin_unlock(&global_rsv->lock);
5017                 return -ENOSPC;
5018         }
5019         global_rsv->reserved -= num_bytes;
5020         if (global_rsv->reserved < global_rsv->size)
5021                 global_rsv->full = 0;
5022         spin_unlock(&global_rsv->lock);
5023
5024         block_rsv_add_bytes(dest, num_bytes, 1);
5025         return 0;
5026 }
5027
5028 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5029                                     struct btrfs_block_rsv *block_rsv,
5030                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5031 {
5032         struct btrfs_space_info *space_info = block_rsv->space_info;
5033
5034         spin_lock(&block_rsv->lock);
5035         if (num_bytes == (u64)-1)
5036                 num_bytes = block_rsv->size;
5037         block_rsv->size -= num_bytes;
5038         if (block_rsv->reserved >= block_rsv->size) {
5039                 num_bytes = block_rsv->reserved - block_rsv->size;
5040                 block_rsv->reserved = block_rsv->size;
5041                 block_rsv->full = 1;
5042         } else {
5043                 num_bytes = 0;
5044         }
5045         spin_unlock(&block_rsv->lock);
5046
5047         if (num_bytes > 0) {
5048                 if (dest) {
5049                         spin_lock(&dest->lock);
5050                         if (!dest->full) {
5051                                 u64 bytes_to_add;
5052
5053                                 bytes_to_add = dest->size - dest->reserved;
5054                                 bytes_to_add = min(num_bytes, bytes_to_add);
5055                                 dest->reserved += bytes_to_add;
5056                                 if (dest->reserved >= dest->size)
5057                                         dest->full = 1;
5058                                 num_bytes -= bytes_to_add;
5059                         }
5060                         spin_unlock(&dest->lock);
5061                 }
5062                 if (num_bytes) {
5063                         spin_lock(&space_info->lock);
5064                         space_info->bytes_may_use -= num_bytes;
5065                         trace_btrfs_space_reservation(fs_info, "space_info",
5066                                         space_info->flags, num_bytes, 0);
5067                         spin_unlock(&space_info->lock);
5068                 }
5069         }
5070 }
5071
5072 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
5073                                    struct btrfs_block_rsv *dst, u64 num_bytes)
5074 {
5075         int ret;
5076
5077         ret = block_rsv_use_bytes(src, num_bytes);
5078         if (ret)
5079                 return ret;
5080
5081         block_rsv_add_bytes(dst, num_bytes, 1);
5082         return 0;
5083 }
5084
5085 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5086 {
5087         memset(rsv, 0, sizeof(*rsv));
5088         spin_lock_init(&rsv->lock);
5089         rsv->type = type;
5090 }
5091
5092 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5093                                               unsigned short type)
5094 {
5095         struct btrfs_block_rsv *block_rsv;
5096         struct btrfs_fs_info *fs_info = root->fs_info;
5097
5098         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5099         if (!block_rsv)
5100                 return NULL;
5101
5102         btrfs_init_block_rsv(block_rsv, type);
5103         block_rsv->space_info = __find_space_info(fs_info,
5104                                                   BTRFS_BLOCK_GROUP_METADATA);
5105         return block_rsv;
5106 }
5107
5108 void btrfs_free_block_rsv(struct btrfs_root *root,
5109                           struct btrfs_block_rsv *rsv)
5110 {
5111         if (!rsv)
5112                 return;
5113         btrfs_block_rsv_release(root, rsv, (u64)-1);
5114         kfree(rsv);
5115 }
5116
5117 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5118 {
5119         kfree(rsv);
5120 }
5121
5122 int btrfs_block_rsv_add(struct btrfs_root *root,
5123                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5124                         enum btrfs_reserve_flush_enum flush)
5125 {
5126         int ret;
5127
5128         if (num_bytes == 0)
5129                 return 0;
5130
5131         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5132         if (!ret) {
5133                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5134                 return 0;
5135         }
5136
5137         return ret;
5138 }
5139
5140 int btrfs_block_rsv_check(struct btrfs_root *root,
5141                           struct btrfs_block_rsv *block_rsv, int min_factor)
5142 {
5143         u64 num_bytes = 0;
5144         int ret = -ENOSPC;
5145
5146         if (!block_rsv)
5147                 return 0;
5148
5149         spin_lock(&block_rsv->lock);
5150         num_bytes = div_factor(block_rsv->size, min_factor);
5151         if (block_rsv->reserved >= num_bytes)
5152                 ret = 0;
5153         spin_unlock(&block_rsv->lock);
5154
5155         return ret;
5156 }
5157
5158 int btrfs_block_rsv_refill(struct btrfs_root *root,
5159                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5160                            enum btrfs_reserve_flush_enum flush)
5161 {
5162         u64 num_bytes = 0;
5163         int ret = -ENOSPC;
5164
5165         if (!block_rsv)
5166                 return 0;
5167
5168         spin_lock(&block_rsv->lock);
5169         num_bytes = min_reserved;
5170         if (block_rsv->reserved >= num_bytes)
5171                 ret = 0;
5172         else
5173                 num_bytes -= block_rsv->reserved;
5174         spin_unlock(&block_rsv->lock);
5175
5176         if (!ret)
5177                 return 0;
5178
5179         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5180         if (!ret) {
5181                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5182                 return 0;
5183         }
5184
5185         return ret;
5186 }
5187
5188 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5189                             struct btrfs_block_rsv *dst_rsv,
5190                             u64 num_bytes)
5191 {
5192         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5193 }
5194
5195 void btrfs_block_rsv_release(struct btrfs_root *root,
5196                              struct btrfs_block_rsv *block_rsv,
5197                              u64 num_bytes)
5198 {
5199         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5200         if (global_rsv == block_rsv ||
5201             block_rsv->space_info != global_rsv->space_info)
5202                 global_rsv = NULL;
5203         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5204                                 num_bytes);
5205 }
5206
5207 /*
5208  * helper to calculate size of global block reservation.
5209  * the desired value is sum of space used by extent tree,
5210  * checksum tree and root tree
5211  */
5212 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5213 {
5214         struct btrfs_space_info *sinfo;
5215         u64 num_bytes;
5216         u64 meta_used;
5217         u64 data_used;
5218         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5219
5220         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5221         spin_lock(&sinfo->lock);
5222         data_used = sinfo->bytes_used;
5223         spin_unlock(&sinfo->lock);
5224
5225         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5226         spin_lock(&sinfo->lock);
5227         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5228                 data_used = 0;
5229         meta_used = sinfo->bytes_used;
5230         spin_unlock(&sinfo->lock);
5231
5232         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5233                     csum_size * 2;
5234         num_bytes += div_u64(data_used + meta_used, 50);
5235
5236         if (num_bytes * 3 > meta_used)
5237                 num_bytes = div_u64(meta_used, 3);
5238
5239         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5240 }
5241
5242 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5243 {
5244         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5245         struct btrfs_space_info *sinfo = block_rsv->space_info;
5246         u64 num_bytes;
5247
5248         num_bytes = calc_global_metadata_size(fs_info);
5249
5250         spin_lock(&sinfo->lock);
5251         spin_lock(&block_rsv->lock);
5252
5253         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
5254
5255         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5256                     sinfo->bytes_reserved + sinfo->bytes_readonly +
5257                     sinfo->bytes_may_use;
5258
5259         if (sinfo->total_bytes > num_bytes) {
5260                 num_bytes = sinfo->total_bytes - num_bytes;
5261                 block_rsv->reserved += num_bytes;
5262                 sinfo->bytes_may_use += num_bytes;
5263                 trace_btrfs_space_reservation(fs_info, "space_info",
5264                                       sinfo->flags, num_bytes, 1);
5265         }
5266
5267         if (block_rsv->reserved >= block_rsv->size) {
5268                 num_bytes = block_rsv->reserved - block_rsv->size;
5269                 sinfo->bytes_may_use -= num_bytes;
5270                 trace_btrfs_space_reservation(fs_info, "space_info",
5271                                       sinfo->flags, num_bytes, 0);
5272                 block_rsv->reserved = block_rsv->size;
5273                 block_rsv->full = 1;
5274         }
5275
5276         spin_unlock(&block_rsv->lock);
5277         spin_unlock(&sinfo->lock);
5278 }
5279
5280 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5281 {
5282         struct btrfs_space_info *space_info;
5283
5284         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5285         fs_info->chunk_block_rsv.space_info = space_info;
5286
5287         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5288         fs_info->global_block_rsv.space_info = space_info;
5289         fs_info->delalloc_block_rsv.space_info = space_info;
5290         fs_info->trans_block_rsv.space_info = space_info;
5291         fs_info->empty_block_rsv.space_info = space_info;
5292         fs_info->delayed_block_rsv.space_info = space_info;
5293
5294         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5295         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5296         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5297         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5298         if (fs_info->quota_root)
5299                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5300         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5301
5302         update_global_block_rsv(fs_info);
5303 }
5304
5305 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5306 {
5307         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5308                                 (u64)-1);
5309         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5310         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5311         WARN_ON(fs_info->trans_block_rsv.size > 0);
5312         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5313         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5314         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5315         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5316         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5317 }
5318
5319 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5320                                   struct btrfs_root *root)
5321 {
5322         if (!trans->block_rsv)
5323                 return;
5324
5325         if (!trans->bytes_reserved)
5326                 return;
5327
5328         trace_btrfs_space_reservation(root->fs_info, "transaction",
5329                                       trans->transid, trans->bytes_reserved, 0);
5330         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5331         trans->bytes_reserved = 0;
5332 }
5333
5334 /*
5335  * To be called after all the new block groups attached to the transaction
5336  * handle have been created (btrfs_create_pending_block_groups()).
5337  */
5338 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5339 {
5340         struct btrfs_fs_info *fs_info = trans->root->fs_info;
5341
5342         if (!trans->chunk_bytes_reserved)
5343                 return;
5344
5345         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5346
5347         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5348                                 trans->chunk_bytes_reserved);
5349         trans->chunk_bytes_reserved = 0;
5350 }
5351
5352 /* Can only return 0 or -ENOSPC */
5353 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5354                                   struct inode *inode)
5355 {
5356         struct btrfs_root *root = BTRFS_I(inode)->root;
5357         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5358         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5359
5360         /*
5361          * We need to hold space in order to delete our orphan item once we've
5362          * added it, so this takes the reservation so we can release it later
5363          * when we are truly done with the orphan item.
5364          */
5365         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5366         trace_btrfs_space_reservation(root->fs_info, "orphan",
5367                                       btrfs_ino(inode), num_bytes, 1);
5368         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5369 }
5370
5371 void btrfs_orphan_release_metadata(struct inode *inode)
5372 {
5373         struct btrfs_root *root = BTRFS_I(inode)->root;
5374         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5375         trace_btrfs_space_reservation(root->fs_info, "orphan",
5376                                       btrfs_ino(inode), num_bytes, 0);
5377         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5378 }
5379
5380 /*
5381  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5382  * root: the root of the parent directory
5383  * rsv: block reservation
5384  * items: the number of items that we need do reservation
5385  * qgroup_reserved: used to return the reserved size in qgroup
5386  *
5387  * This function is used to reserve the space for snapshot/subvolume
5388  * creation and deletion. Those operations are different with the
5389  * common file/directory operations, they change two fs/file trees
5390  * and root tree, the number of items that the qgroup reserves is
5391  * different with the free space reservation. So we can not use
5392  * the space reseravtion mechanism in start_transaction().
5393  */
5394 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5395                                      struct btrfs_block_rsv *rsv,
5396                                      int items,
5397                                      u64 *qgroup_reserved,
5398                                      bool use_global_rsv)
5399 {
5400         u64 num_bytes;
5401         int ret;
5402         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5403
5404         if (root->fs_info->quota_enabled) {
5405                 /* One for parent inode, two for dir entries */
5406                 num_bytes = 3 * root->nodesize;
5407                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5408                 if (ret)
5409                         return ret;
5410         } else {
5411                 num_bytes = 0;
5412         }
5413
5414         *qgroup_reserved = num_bytes;
5415
5416         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5417         rsv->space_info = __find_space_info(root->fs_info,
5418                                             BTRFS_BLOCK_GROUP_METADATA);
5419         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5420                                   BTRFS_RESERVE_FLUSH_ALL);
5421
5422         if (ret == -ENOSPC && use_global_rsv)
5423                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5424
5425         if (ret && *qgroup_reserved)
5426                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5427
5428         return ret;
5429 }
5430
5431 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5432                                       struct btrfs_block_rsv *rsv,
5433                                       u64 qgroup_reserved)
5434 {
5435         btrfs_block_rsv_release(root, rsv, (u64)-1);
5436 }
5437
5438 /**
5439  * drop_outstanding_extent - drop an outstanding extent
5440  * @inode: the inode we're dropping the extent for
5441  * @num_bytes: the number of bytes we're relaseing.
5442  *
5443  * This is called when we are freeing up an outstanding extent, either called
5444  * after an error or after an extent is written.  This will return the number of
5445  * reserved extents that need to be freed.  This must be called with
5446  * BTRFS_I(inode)->lock held.
5447  */
5448 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5449 {
5450         unsigned drop_inode_space = 0;
5451         unsigned dropped_extents = 0;
5452         unsigned num_extents = 0;
5453
5454         num_extents = (unsigned)div64_u64(num_bytes +
5455                                           BTRFS_MAX_EXTENT_SIZE - 1,
5456                                           BTRFS_MAX_EXTENT_SIZE);
5457         ASSERT(num_extents);
5458         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5459         BTRFS_I(inode)->outstanding_extents -= num_extents;
5460
5461         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5462             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5463                                &BTRFS_I(inode)->runtime_flags))
5464                 drop_inode_space = 1;
5465
5466         /*
5467          * If we have more or the same amount of outsanding extents than we have
5468          * reserved then we need to leave the reserved extents count alone.
5469          */
5470         if (BTRFS_I(inode)->outstanding_extents >=
5471             BTRFS_I(inode)->reserved_extents)
5472                 return drop_inode_space;
5473
5474         dropped_extents = BTRFS_I(inode)->reserved_extents -
5475                 BTRFS_I(inode)->outstanding_extents;
5476         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5477         return dropped_extents + drop_inode_space;
5478 }
5479
5480 /**
5481  * calc_csum_metadata_size - return the amount of metada space that must be
5482  *      reserved/free'd for the given bytes.
5483  * @inode: the inode we're manipulating
5484  * @num_bytes: the number of bytes in question
5485  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5486  *
5487  * This adjusts the number of csum_bytes in the inode and then returns the
5488  * correct amount of metadata that must either be reserved or freed.  We
5489  * calculate how many checksums we can fit into one leaf and then divide the
5490  * number of bytes that will need to be checksumed by this value to figure out
5491  * how many checksums will be required.  If we are adding bytes then the number
5492  * may go up and we will return the number of additional bytes that must be
5493  * reserved.  If it is going down we will return the number of bytes that must
5494  * be freed.
5495  *
5496  * This must be called with BTRFS_I(inode)->lock held.
5497  */
5498 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5499                                    int reserve)
5500 {
5501         struct btrfs_root *root = BTRFS_I(inode)->root;
5502         u64 old_csums, num_csums;
5503
5504         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5505             BTRFS_I(inode)->csum_bytes == 0)
5506                 return 0;
5507
5508         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5509         if (reserve)
5510                 BTRFS_I(inode)->csum_bytes += num_bytes;
5511         else
5512                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5513         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5514
5515         /* No change, no need to reserve more */
5516         if (old_csums == num_csums)
5517                 return 0;
5518
5519         if (reserve)
5520                 return btrfs_calc_trans_metadata_size(root,
5521                                                       num_csums - old_csums);
5522
5523         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5524 }
5525
5526 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5527 {
5528         struct btrfs_root *root = BTRFS_I(inode)->root;
5529         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5530         u64 to_reserve = 0;
5531         u64 csum_bytes;
5532         unsigned nr_extents = 0;
5533         int extra_reserve = 0;
5534         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5535         int ret = 0;
5536         bool delalloc_lock = true;
5537         u64 to_free = 0;
5538         unsigned dropped;
5539
5540         /* If we are a free space inode we need to not flush since we will be in
5541          * the middle of a transaction commit.  We also don't need the delalloc
5542          * mutex since we won't race with anybody.  We need this mostly to make
5543          * lockdep shut its filthy mouth.
5544          */
5545         if (btrfs_is_free_space_inode(inode)) {
5546                 flush = BTRFS_RESERVE_NO_FLUSH;
5547                 delalloc_lock = false;
5548         }
5549
5550         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5551             btrfs_transaction_in_commit(root->fs_info))
5552                 schedule_timeout(1);
5553
5554         if (delalloc_lock)
5555                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5556
5557         num_bytes = ALIGN(num_bytes, root->sectorsize);
5558
5559         spin_lock(&BTRFS_I(inode)->lock);
5560         nr_extents = (unsigned)div64_u64(num_bytes +
5561                                          BTRFS_MAX_EXTENT_SIZE - 1,
5562                                          BTRFS_MAX_EXTENT_SIZE);
5563         BTRFS_I(inode)->outstanding_extents += nr_extents;
5564         nr_extents = 0;
5565
5566         if (BTRFS_I(inode)->outstanding_extents >
5567             BTRFS_I(inode)->reserved_extents)
5568                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5569                         BTRFS_I(inode)->reserved_extents;
5570
5571         /*
5572          * Add an item to reserve for updating the inode when we complete the
5573          * delalloc io.
5574          */
5575         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5576                       &BTRFS_I(inode)->runtime_flags)) {
5577                 nr_extents++;
5578                 extra_reserve = 1;
5579         }
5580
5581         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5582         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5583         csum_bytes = BTRFS_I(inode)->csum_bytes;
5584         spin_unlock(&BTRFS_I(inode)->lock);
5585
5586         if (root->fs_info->quota_enabled) {
5587                 ret = btrfs_qgroup_reserve_meta(root,
5588                                 nr_extents * root->nodesize);
5589                 if (ret)
5590                         goto out_fail;
5591         }
5592
5593         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5594         if (unlikely(ret)) {
5595                 btrfs_qgroup_free_meta(root, nr_extents * root->nodesize);
5596                 goto out_fail;
5597         }
5598
5599         spin_lock(&BTRFS_I(inode)->lock);
5600         if (extra_reserve) {
5601                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5602                         &BTRFS_I(inode)->runtime_flags);
5603                 nr_extents--;
5604         }
5605         BTRFS_I(inode)->reserved_extents += nr_extents;
5606         spin_unlock(&BTRFS_I(inode)->lock);
5607
5608         if (delalloc_lock)
5609                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5610
5611         if (to_reserve)
5612                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5613                                               btrfs_ino(inode), to_reserve, 1);
5614         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5615
5616         return 0;
5617
5618 out_fail:
5619         spin_lock(&BTRFS_I(inode)->lock);
5620         dropped = drop_outstanding_extent(inode, num_bytes);
5621         /*
5622          * If the inodes csum_bytes is the same as the original
5623          * csum_bytes then we know we haven't raced with any free()ers
5624          * so we can just reduce our inodes csum bytes and carry on.
5625          */
5626         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5627                 calc_csum_metadata_size(inode, num_bytes, 0);
5628         } else {
5629                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5630                 u64 bytes;
5631
5632                 /*
5633                  * This is tricky, but first we need to figure out how much we
5634                  * free'd from any free-ers that occured during this
5635                  * reservation, so we reset ->csum_bytes to the csum_bytes
5636                  * before we dropped our lock, and then call the free for the
5637                  * number of bytes that were freed while we were trying our
5638                  * reservation.
5639                  */
5640                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5641                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5642                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5643
5644
5645                 /*
5646                  * Now we need to see how much we would have freed had we not
5647                  * been making this reservation and our ->csum_bytes were not
5648                  * artificially inflated.
5649                  */
5650                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5651                 bytes = csum_bytes - orig_csum_bytes;
5652                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5653
5654                 /*
5655                  * Now reset ->csum_bytes to what it should be.  If bytes is
5656                  * more than to_free then we would have free'd more space had we
5657                  * not had an artificially high ->csum_bytes, so we need to free
5658                  * the remainder.  If bytes is the same or less then we don't
5659                  * need to do anything, the other free-ers did the correct
5660                  * thing.
5661                  */
5662                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5663                 if (bytes > to_free)
5664                         to_free = bytes - to_free;
5665                 else
5666                         to_free = 0;
5667         }
5668         spin_unlock(&BTRFS_I(inode)->lock);
5669         if (dropped)
5670                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5671
5672         if (to_free) {
5673                 btrfs_block_rsv_release(root, block_rsv, to_free);
5674                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5675                                               btrfs_ino(inode), to_free, 0);
5676         }
5677         if (delalloc_lock)
5678                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5679         return ret;
5680 }
5681
5682 /**
5683  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5684  * @inode: the inode to release the reservation for
5685  * @num_bytes: the number of bytes we're releasing
5686  *
5687  * This will release the metadata reservation for an inode.  This can be called
5688  * once we complete IO for a given set of bytes to release their metadata
5689  * reservations.
5690  */
5691 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5692 {
5693         struct btrfs_root *root = BTRFS_I(inode)->root;
5694         u64 to_free = 0;
5695         unsigned dropped;
5696
5697         num_bytes = ALIGN(num_bytes, root->sectorsize);
5698         spin_lock(&BTRFS_I(inode)->lock);
5699         dropped = drop_outstanding_extent(inode, num_bytes);
5700
5701         if (num_bytes)
5702                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5703         spin_unlock(&BTRFS_I(inode)->lock);
5704         if (dropped > 0)
5705                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5706
5707         if (btrfs_test_is_dummy_root(root))
5708                 return;
5709
5710         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5711                                       btrfs_ino(inode), to_free, 0);
5712
5713         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5714                                 to_free);
5715 }
5716
5717 /**
5718  * __btrfs_delalloc_reserve_space - reserve data and metadata space for
5719  * delalloc
5720  * @inode: inode we're writing to
5721  * @start: start range we are writing to
5722  * @len: how long the range we are writing to
5723  *
5724  * TODO: This function will finally replace old btrfs_delalloc_reserve_space()
5725  *
5726  * This will do the following things
5727  *
5728  * o reserve space in data space info for num bytes
5729  *   and reserve precious corresponding qgroup space
5730  *   (Done in check_data_free_space)
5731  *
5732  * o reserve space for metadata space, based on the number of outstanding
5733  *   extents and how much csums will be needed
5734  *   also reserve metadata space in a per root over-reserve method.
5735  * o add to the inodes->delalloc_bytes
5736  * o add it to the fs_info's delalloc inodes list.
5737  *   (Above 3 all done in delalloc_reserve_metadata)
5738  *
5739  * Return 0 for success
5740  * Return <0 for error(-ENOSPC or -EQUOT)
5741  */
5742 int __btrfs_delalloc_reserve_space(struct inode *inode, u64 start, u64 len)
5743 {
5744         int ret;
5745
5746         ret = __btrfs_check_data_free_space(inode, start, len);
5747         if (ret < 0)
5748                 return ret;
5749         ret = btrfs_delalloc_reserve_metadata(inode, len);
5750         if (ret < 0)
5751                 __btrfs_free_reserved_data_space(inode, start, len);
5752         return ret;
5753 }
5754
5755 /**
5756  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5757  * @inode: inode we're writing to
5758  * @num_bytes: the number of bytes we want to allocate
5759  *
5760  * This will do the following things
5761  *
5762  * o reserve space in the data space info for num_bytes
5763  * o reserve space in the metadata space info based on number of outstanding
5764  *   extents and how much csums will be needed
5765  * o add to the inodes ->delalloc_bytes
5766  * o add it to the fs_info's delalloc inodes list.
5767  *
5768  * This will return 0 for success and -ENOSPC if there is no space left.
5769  */
5770 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5771 {
5772         int ret;
5773
5774         ret = btrfs_check_data_free_space(inode, num_bytes, num_bytes);
5775         if (ret)
5776                 return ret;
5777
5778         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5779         if (ret) {
5780                 btrfs_free_reserved_data_space(inode, num_bytes);
5781                 return ret;
5782         }
5783
5784         return 0;
5785 }
5786
5787 /**
5788  * __btrfs_delalloc_release_space - release data and metadata space for delalloc
5789  * @inode: inode we're releasing space for
5790  * @start: start position of the space already reserved
5791  * @len: the len of the space already reserved
5792  *
5793  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5794  * called in the case that we don't need the metadata AND data reservations
5795  * anymore.  So if there is an error or we insert an inline extent.
5796  *
5797  * This function will release the metadata space that was not used and will
5798  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5799  * list if there are no delalloc bytes left.
5800  * Also it will handle the qgroup reserved space.
5801  */
5802 void __btrfs_delalloc_release_space(struct inode *inode, u64 start, u64 len)
5803 {
5804         btrfs_delalloc_release_metadata(inode, len);
5805         __btrfs_free_reserved_data_space(inode, start, len);
5806 }
5807
5808 /**
5809  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5810  * @inode: inode we're releasing space for
5811  * @num_bytes: the number of bytes we want to free up
5812  *
5813  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5814  * called in the case that we don't need the metadata AND data reservations
5815  * anymore.  So if there is an error or we insert an inline extent.
5816  *
5817  * This function will release the metadata space that was not used and will
5818  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5819  * list if there are no delalloc bytes left.
5820  */
5821 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5822 {
5823         btrfs_delalloc_release_metadata(inode, num_bytes);
5824         btrfs_free_reserved_data_space(inode, num_bytes);
5825 }
5826
5827 static int update_block_group(struct btrfs_trans_handle *trans,
5828                               struct btrfs_root *root, u64 bytenr,
5829                               u64 num_bytes, int alloc)
5830 {
5831         struct btrfs_block_group_cache *cache = NULL;
5832         struct btrfs_fs_info *info = root->fs_info;
5833         u64 total = num_bytes;
5834         u64 old_val;
5835         u64 byte_in_group;
5836         int factor;
5837
5838         /* block accounting for super block */
5839         spin_lock(&info->delalloc_root_lock);
5840         old_val = btrfs_super_bytes_used(info->super_copy);
5841         if (alloc)
5842                 old_val += num_bytes;
5843         else
5844                 old_val -= num_bytes;
5845         btrfs_set_super_bytes_used(info->super_copy, old_val);
5846         spin_unlock(&info->delalloc_root_lock);
5847
5848         while (total) {
5849                 cache = btrfs_lookup_block_group(info, bytenr);
5850                 if (!cache)
5851                         return -ENOENT;
5852                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5853                                     BTRFS_BLOCK_GROUP_RAID1 |
5854                                     BTRFS_BLOCK_GROUP_RAID10))
5855                         factor = 2;
5856                 else
5857                         factor = 1;
5858                 /*
5859                  * If this block group has free space cache written out, we
5860                  * need to make sure to load it if we are removing space.  This
5861                  * is because we need the unpinning stage to actually add the
5862                  * space back to the block group, otherwise we will leak space.
5863                  */
5864                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5865                         cache_block_group(cache, 1);
5866
5867                 byte_in_group = bytenr - cache->key.objectid;
5868                 WARN_ON(byte_in_group > cache->key.offset);
5869
5870                 spin_lock(&cache->space_info->lock);
5871                 spin_lock(&cache->lock);
5872
5873                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5874                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5875                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5876
5877                 old_val = btrfs_block_group_used(&cache->item);
5878                 num_bytes = min(total, cache->key.offset - byte_in_group);
5879                 if (alloc) {
5880                         old_val += num_bytes;
5881                         btrfs_set_block_group_used(&cache->item, old_val);
5882                         cache->reserved -= num_bytes;
5883                         cache->space_info->bytes_reserved -= num_bytes;
5884                         cache->space_info->bytes_used += num_bytes;
5885                         cache->space_info->disk_used += num_bytes * factor;
5886                         spin_unlock(&cache->lock);
5887                         spin_unlock(&cache->space_info->lock);
5888                 } else {
5889                         old_val -= num_bytes;
5890                         btrfs_set_block_group_used(&cache->item, old_val);
5891                         cache->pinned += num_bytes;
5892                         cache->space_info->bytes_pinned += num_bytes;
5893                         cache->space_info->bytes_used -= num_bytes;
5894                         cache->space_info->disk_used -= num_bytes * factor;
5895                         spin_unlock(&cache->lock);
5896                         spin_unlock(&cache->space_info->lock);
5897
5898                         set_extent_dirty(info->pinned_extents,
5899                                          bytenr, bytenr + num_bytes - 1,
5900                                          GFP_NOFS | __GFP_NOFAIL);
5901                         /*
5902                          * No longer have used bytes in this block group, queue
5903                          * it for deletion.
5904                          */
5905                         if (old_val == 0) {
5906                                 spin_lock(&info->unused_bgs_lock);
5907                                 if (list_empty(&cache->bg_list)) {
5908                                         btrfs_get_block_group(cache);
5909                                         list_add_tail(&cache->bg_list,
5910                                                       &info->unused_bgs);
5911                                 }
5912                                 spin_unlock(&info->unused_bgs_lock);
5913                         }
5914                 }
5915
5916                 spin_lock(&trans->transaction->dirty_bgs_lock);
5917                 if (list_empty(&cache->dirty_list)) {
5918                         list_add_tail(&cache->dirty_list,
5919                                       &trans->transaction->dirty_bgs);
5920                                 trans->transaction->num_dirty_bgs++;
5921                         btrfs_get_block_group(cache);
5922                 }
5923                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5924
5925                 btrfs_put_block_group(cache);
5926                 total -= num_bytes;
5927                 bytenr += num_bytes;
5928         }
5929         return 0;
5930 }
5931
5932 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5933 {
5934         struct btrfs_block_group_cache *cache;
5935         u64 bytenr;
5936
5937         spin_lock(&root->fs_info->block_group_cache_lock);
5938         bytenr = root->fs_info->first_logical_byte;
5939         spin_unlock(&root->fs_info->block_group_cache_lock);
5940
5941         if (bytenr < (u64)-1)
5942                 return bytenr;
5943
5944         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5945         if (!cache)
5946                 return 0;
5947
5948         bytenr = cache->key.objectid;
5949         btrfs_put_block_group(cache);
5950
5951         return bytenr;
5952 }
5953
5954 static int pin_down_extent(struct btrfs_root *root,
5955                            struct btrfs_block_group_cache *cache,
5956                            u64 bytenr, u64 num_bytes, int reserved)
5957 {
5958         spin_lock(&cache->space_info->lock);
5959         spin_lock(&cache->lock);
5960         cache->pinned += num_bytes;
5961         cache->space_info->bytes_pinned += num_bytes;
5962         if (reserved) {
5963                 cache->reserved -= num_bytes;
5964                 cache->space_info->bytes_reserved -= num_bytes;
5965         }
5966         spin_unlock(&cache->lock);
5967         spin_unlock(&cache->space_info->lock);
5968
5969         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5970                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5971         if (reserved)
5972                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5973         return 0;
5974 }
5975
5976 /*
5977  * this function must be called within transaction
5978  */
5979 int btrfs_pin_extent(struct btrfs_root *root,
5980                      u64 bytenr, u64 num_bytes, int reserved)
5981 {
5982         struct btrfs_block_group_cache *cache;
5983
5984         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5985         BUG_ON(!cache); /* Logic error */
5986
5987         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5988
5989         btrfs_put_block_group(cache);
5990         return 0;
5991 }
5992
5993 /*
5994  * this function must be called within transaction
5995  */
5996 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5997                                     u64 bytenr, u64 num_bytes)
5998 {
5999         struct btrfs_block_group_cache *cache;
6000         int ret;
6001
6002         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6003         if (!cache)
6004                 return -EINVAL;
6005
6006         /*
6007          * pull in the free space cache (if any) so that our pin
6008          * removes the free space from the cache.  We have load_only set
6009          * to one because the slow code to read in the free extents does check
6010          * the pinned extents.
6011          */
6012         cache_block_group(cache, 1);
6013
6014         pin_down_extent(root, cache, bytenr, num_bytes, 0);
6015
6016         /* remove us from the free space cache (if we're there at all) */
6017         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6018         btrfs_put_block_group(cache);
6019         return ret;
6020 }
6021
6022 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
6023 {
6024         int ret;
6025         struct btrfs_block_group_cache *block_group;
6026         struct btrfs_caching_control *caching_ctl;
6027
6028         block_group = btrfs_lookup_block_group(root->fs_info, start);
6029         if (!block_group)
6030                 return -EINVAL;
6031
6032         cache_block_group(block_group, 0);
6033         caching_ctl = get_caching_control(block_group);
6034
6035         if (!caching_ctl) {
6036                 /* Logic error */
6037                 BUG_ON(!block_group_cache_done(block_group));
6038                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6039         } else {
6040                 mutex_lock(&caching_ctl->mutex);
6041
6042                 if (start >= caching_ctl->progress) {
6043                         ret = add_excluded_extent(root, start, num_bytes);
6044                 } else if (start + num_bytes <= caching_ctl->progress) {
6045                         ret = btrfs_remove_free_space(block_group,
6046                                                       start, num_bytes);
6047                 } else {
6048                         num_bytes = caching_ctl->progress - start;
6049                         ret = btrfs_remove_free_space(block_group,
6050                                                       start, num_bytes);
6051                         if (ret)
6052                                 goto out_lock;
6053
6054                         num_bytes = (start + num_bytes) -
6055                                 caching_ctl->progress;
6056                         start = caching_ctl->progress;
6057                         ret = add_excluded_extent(root, start, num_bytes);
6058                 }
6059 out_lock:
6060                 mutex_unlock(&caching_ctl->mutex);
6061                 put_caching_control(caching_ctl);
6062         }
6063         btrfs_put_block_group(block_group);
6064         return ret;
6065 }
6066
6067 int btrfs_exclude_logged_extents(struct btrfs_root *log,
6068                                  struct extent_buffer *eb)
6069 {
6070         struct btrfs_file_extent_item *item;
6071         struct btrfs_key key;
6072         int found_type;
6073         int i;
6074
6075         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
6076                 return 0;
6077
6078         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6079                 btrfs_item_key_to_cpu(eb, &key, i);
6080                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6081                         continue;
6082                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6083                 found_type = btrfs_file_extent_type(eb, item);
6084                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6085                         continue;
6086                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6087                         continue;
6088                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6089                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6090                 __exclude_logged_extent(log, key.objectid, key.offset);
6091         }
6092
6093         return 0;
6094 }
6095
6096 /**
6097  * btrfs_update_reserved_bytes - update the block_group and space info counters
6098  * @cache:      The cache we are manipulating
6099  * @num_bytes:  The number of bytes in question
6100  * @reserve:    One of the reservation enums
6101  * @delalloc:   The blocks are allocated for the delalloc write
6102  *
6103  * This is called by the allocator when it reserves space, or by somebody who is
6104  * freeing space that was never actually used on disk.  For example if you
6105  * reserve some space for a new leaf in transaction A and before transaction A
6106  * commits you free that leaf, you call this with reserve set to 0 in order to
6107  * clear the reservation.
6108  *
6109  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
6110  * ENOSPC accounting.  For data we handle the reservation through clearing the
6111  * delalloc bits in the io_tree.  We have to do this since we could end up
6112  * allocating less disk space for the amount of data we have reserved in the
6113  * case of compression.
6114  *
6115  * If this is a reservation and the block group has become read only we cannot
6116  * make the reservation and return -EAGAIN, otherwise this function always
6117  * succeeds.
6118  */
6119 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
6120                                        u64 num_bytes, int reserve, int delalloc)
6121 {
6122         struct btrfs_space_info *space_info = cache->space_info;
6123         int ret = 0;
6124
6125         spin_lock(&space_info->lock);
6126         spin_lock(&cache->lock);
6127         if (reserve != RESERVE_FREE) {
6128                 if (cache->ro) {
6129                         ret = -EAGAIN;
6130                 } else {
6131                         cache->reserved += num_bytes;
6132                         space_info->bytes_reserved += num_bytes;
6133                         if (reserve == RESERVE_ALLOC) {
6134                                 trace_btrfs_space_reservation(cache->fs_info,
6135                                                 "space_info", space_info->flags,
6136                                                 num_bytes, 0);
6137                                 space_info->bytes_may_use -= num_bytes;
6138                         }
6139
6140                         if (delalloc)
6141                                 cache->delalloc_bytes += num_bytes;
6142                 }
6143         } else {
6144                 if (cache->ro)
6145                         space_info->bytes_readonly += num_bytes;
6146                 cache->reserved -= num_bytes;
6147                 space_info->bytes_reserved -= num_bytes;
6148
6149                 if (delalloc)
6150                         cache->delalloc_bytes -= num_bytes;
6151         }
6152         spin_unlock(&cache->lock);
6153         spin_unlock(&space_info->lock);
6154         return ret;
6155 }
6156
6157 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6158                                 struct btrfs_root *root)
6159 {
6160         struct btrfs_fs_info *fs_info = root->fs_info;
6161         struct btrfs_caching_control *next;
6162         struct btrfs_caching_control *caching_ctl;
6163         struct btrfs_block_group_cache *cache;
6164
6165         down_write(&fs_info->commit_root_sem);
6166
6167         list_for_each_entry_safe(caching_ctl, next,
6168                                  &fs_info->caching_block_groups, list) {
6169                 cache = caching_ctl->block_group;
6170                 if (block_group_cache_done(cache)) {
6171                         cache->last_byte_to_unpin = (u64)-1;
6172                         list_del_init(&caching_ctl->list);
6173                         put_caching_control(caching_ctl);
6174                 } else {
6175                         cache->last_byte_to_unpin = caching_ctl->progress;
6176                 }
6177         }
6178
6179         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6180                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6181         else
6182                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6183
6184         up_write(&fs_info->commit_root_sem);
6185
6186         update_global_block_rsv(fs_info);
6187 }
6188
6189 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
6190                               const bool return_free_space)
6191 {
6192         struct btrfs_fs_info *fs_info = root->fs_info;
6193         struct btrfs_block_group_cache *cache = NULL;
6194         struct btrfs_space_info *space_info;
6195         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6196         u64 len;
6197         bool readonly;
6198
6199         while (start <= end) {
6200                 readonly = false;
6201                 if (!cache ||
6202                     start >= cache->key.objectid + cache->key.offset) {
6203                         if (cache)
6204                                 btrfs_put_block_group(cache);
6205                         cache = btrfs_lookup_block_group(fs_info, start);
6206                         BUG_ON(!cache); /* Logic error */
6207                 }
6208
6209                 len = cache->key.objectid + cache->key.offset - start;
6210                 len = min(len, end + 1 - start);
6211
6212                 if (start < cache->last_byte_to_unpin) {
6213                         len = min(len, cache->last_byte_to_unpin - start);
6214                         if (return_free_space)
6215                                 btrfs_add_free_space(cache, start, len);
6216                 }
6217
6218                 start += len;
6219                 space_info = cache->space_info;
6220
6221                 spin_lock(&space_info->lock);
6222                 spin_lock(&cache->lock);
6223                 cache->pinned -= len;
6224                 space_info->bytes_pinned -= len;
6225                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6226                 if (cache->ro) {
6227                         space_info->bytes_readonly += len;
6228                         readonly = true;
6229                 }
6230                 spin_unlock(&cache->lock);
6231                 if (!readonly && global_rsv->space_info == space_info) {
6232                         spin_lock(&global_rsv->lock);
6233                         if (!global_rsv->full) {
6234                                 len = min(len, global_rsv->size -
6235                                           global_rsv->reserved);
6236                                 global_rsv->reserved += len;
6237                                 space_info->bytes_may_use += len;
6238                                 if (global_rsv->reserved >= global_rsv->size)
6239                                         global_rsv->full = 1;
6240                         }
6241                         spin_unlock(&global_rsv->lock);
6242                 }
6243                 spin_unlock(&space_info->lock);
6244         }
6245
6246         if (cache)
6247                 btrfs_put_block_group(cache);
6248         return 0;
6249 }
6250
6251 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6252                                struct btrfs_root *root)
6253 {
6254         struct btrfs_fs_info *fs_info = root->fs_info;
6255         struct btrfs_block_group_cache *block_group, *tmp;
6256         struct list_head *deleted_bgs;
6257         struct extent_io_tree *unpin;
6258         u64 start;
6259         u64 end;
6260         int ret;
6261
6262         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6263                 unpin = &fs_info->freed_extents[1];
6264         else
6265                 unpin = &fs_info->freed_extents[0];
6266
6267         while (!trans->aborted) {
6268                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6269                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6270                                             EXTENT_DIRTY, NULL);
6271                 if (ret) {
6272                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6273                         break;
6274                 }
6275
6276                 if (btrfs_test_opt(root, DISCARD))
6277                         ret = btrfs_discard_extent(root, start,
6278                                                    end + 1 - start, NULL);
6279
6280                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
6281                 unpin_extent_range(root, start, end, true);
6282                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6283                 cond_resched();
6284         }
6285
6286         /*
6287          * Transaction is finished.  We don't need the lock anymore.  We
6288          * do need to clean up the block groups in case of a transaction
6289          * abort.
6290          */
6291         deleted_bgs = &trans->transaction->deleted_bgs;
6292         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6293                 u64 trimmed = 0;
6294
6295                 ret = -EROFS;
6296                 if (!trans->aborted)
6297                         ret = btrfs_discard_extent(root,
6298                                                    block_group->key.objectid,
6299                                                    block_group->key.offset,
6300                                                    &trimmed);
6301
6302                 list_del_init(&block_group->bg_list);
6303                 btrfs_put_block_group_trimming(block_group);
6304                 btrfs_put_block_group(block_group);
6305
6306                 if (ret) {
6307                         const char *errstr = btrfs_decode_error(ret);
6308                         btrfs_warn(fs_info,
6309                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6310                                    ret, errstr);
6311                 }
6312         }
6313
6314         return 0;
6315 }
6316
6317 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6318                              u64 owner, u64 root_objectid)
6319 {
6320         struct btrfs_space_info *space_info;
6321         u64 flags;
6322
6323         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6324                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6325                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6326                 else
6327                         flags = BTRFS_BLOCK_GROUP_METADATA;
6328         } else {
6329                 flags = BTRFS_BLOCK_GROUP_DATA;
6330         }
6331
6332         space_info = __find_space_info(fs_info, flags);
6333         BUG_ON(!space_info); /* Logic bug */
6334         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6335 }
6336
6337
6338 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6339                                 struct btrfs_root *root,
6340                                 struct btrfs_delayed_ref_node *node, u64 parent,
6341                                 u64 root_objectid, u64 owner_objectid,
6342                                 u64 owner_offset, int refs_to_drop,
6343                                 struct btrfs_delayed_extent_op *extent_op)
6344 {
6345         struct btrfs_key key;
6346         struct btrfs_path *path;
6347         struct btrfs_fs_info *info = root->fs_info;
6348         struct btrfs_root *extent_root = info->extent_root;
6349         struct extent_buffer *leaf;
6350         struct btrfs_extent_item *ei;
6351         struct btrfs_extent_inline_ref *iref;
6352         int ret;
6353         int is_data;
6354         int extent_slot = 0;
6355         int found_extent = 0;
6356         int num_to_del = 1;
6357         int no_quota = node->no_quota;
6358         u32 item_size;
6359         u64 refs;
6360         u64 bytenr = node->bytenr;
6361         u64 num_bytes = node->num_bytes;
6362         int last_ref = 0;
6363         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6364                                                  SKINNY_METADATA);
6365
6366         if (!info->quota_enabled || !is_fstree(root_objectid))
6367                 no_quota = 1;
6368
6369         path = btrfs_alloc_path();
6370         if (!path)
6371                 return -ENOMEM;
6372
6373         path->reada = 1;
6374         path->leave_spinning = 1;
6375
6376         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6377         BUG_ON(!is_data && refs_to_drop != 1);
6378
6379         if (is_data)
6380                 skinny_metadata = 0;
6381
6382         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6383                                     bytenr, num_bytes, parent,
6384                                     root_objectid, owner_objectid,
6385                                     owner_offset);
6386         if (ret == 0) {
6387                 extent_slot = path->slots[0];
6388                 while (extent_slot >= 0) {
6389                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6390                                               extent_slot);
6391                         if (key.objectid != bytenr)
6392                                 break;
6393                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6394                             key.offset == num_bytes) {
6395                                 found_extent = 1;
6396                                 break;
6397                         }
6398                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6399                             key.offset == owner_objectid) {
6400                                 found_extent = 1;
6401                                 break;
6402                         }
6403                         if (path->slots[0] - extent_slot > 5)
6404                                 break;
6405                         extent_slot--;
6406                 }
6407 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6408                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6409                 if (found_extent && item_size < sizeof(*ei))
6410                         found_extent = 0;
6411 #endif
6412                 if (!found_extent) {
6413                         BUG_ON(iref);
6414                         ret = remove_extent_backref(trans, extent_root, path,
6415                                                     NULL, refs_to_drop,
6416                                                     is_data, &last_ref);
6417                         if (ret) {
6418                                 btrfs_abort_transaction(trans, extent_root, ret);
6419                                 goto out;
6420                         }
6421                         btrfs_release_path(path);
6422                         path->leave_spinning = 1;
6423
6424                         key.objectid = bytenr;
6425                         key.type = BTRFS_EXTENT_ITEM_KEY;
6426                         key.offset = num_bytes;
6427
6428                         if (!is_data && skinny_metadata) {
6429                                 key.type = BTRFS_METADATA_ITEM_KEY;
6430                                 key.offset = owner_objectid;
6431                         }
6432
6433                         ret = btrfs_search_slot(trans, extent_root,
6434                                                 &key, path, -1, 1);
6435                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6436                                 /*
6437                                  * Couldn't find our skinny metadata item,
6438                                  * see if we have ye olde extent item.
6439                                  */
6440                                 path->slots[0]--;
6441                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6442                                                       path->slots[0]);
6443                                 if (key.objectid == bytenr &&
6444                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6445                                     key.offset == num_bytes)
6446                                         ret = 0;
6447                         }
6448
6449                         if (ret > 0 && skinny_metadata) {
6450                                 skinny_metadata = false;
6451                                 key.objectid = bytenr;
6452                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6453                                 key.offset = num_bytes;
6454                                 btrfs_release_path(path);
6455                                 ret = btrfs_search_slot(trans, extent_root,
6456                                                         &key, path, -1, 1);
6457                         }
6458
6459                         if (ret) {
6460                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6461                                         ret, bytenr);
6462                                 if (ret > 0)
6463                                         btrfs_print_leaf(extent_root,
6464                                                          path->nodes[0]);
6465                         }
6466                         if (ret < 0) {
6467                                 btrfs_abort_transaction(trans, extent_root, ret);
6468                                 goto out;
6469                         }
6470                         extent_slot = path->slots[0];
6471                 }
6472         } else if (WARN_ON(ret == -ENOENT)) {
6473                 btrfs_print_leaf(extent_root, path->nodes[0]);
6474                 btrfs_err(info,
6475                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6476                         bytenr, parent, root_objectid, owner_objectid,
6477                         owner_offset);
6478                 btrfs_abort_transaction(trans, extent_root, ret);
6479                 goto out;
6480         } else {
6481                 btrfs_abort_transaction(trans, extent_root, ret);
6482                 goto out;
6483         }
6484
6485         leaf = path->nodes[0];
6486         item_size = btrfs_item_size_nr(leaf, extent_slot);
6487 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6488         if (item_size < sizeof(*ei)) {
6489                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6490                 ret = convert_extent_item_v0(trans, extent_root, path,
6491                                              owner_objectid, 0);
6492                 if (ret < 0) {
6493                         btrfs_abort_transaction(trans, extent_root, ret);
6494                         goto out;
6495                 }
6496
6497                 btrfs_release_path(path);
6498                 path->leave_spinning = 1;
6499
6500                 key.objectid = bytenr;
6501                 key.type = BTRFS_EXTENT_ITEM_KEY;
6502                 key.offset = num_bytes;
6503
6504                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6505                                         -1, 1);
6506                 if (ret) {
6507                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6508                                 ret, bytenr);
6509                         btrfs_print_leaf(extent_root, path->nodes[0]);
6510                 }
6511                 if (ret < 0) {
6512                         btrfs_abort_transaction(trans, extent_root, ret);
6513                         goto out;
6514                 }
6515
6516                 extent_slot = path->slots[0];
6517                 leaf = path->nodes[0];
6518                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6519         }
6520 #endif
6521         BUG_ON(item_size < sizeof(*ei));
6522         ei = btrfs_item_ptr(leaf, extent_slot,
6523                             struct btrfs_extent_item);
6524         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6525             key.type == BTRFS_EXTENT_ITEM_KEY) {
6526                 struct btrfs_tree_block_info *bi;
6527                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6528                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6529                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6530         }
6531
6532         refs = btrfs_extent_refs(leaf, ei);
6533         if (refs < refs_to_drop) {
6534                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6535                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6536                 ret = -EINVAL;
6537                 btrfs_abort_transaction(trans, extent_root, ret);
6538                 goto out;
6539         }
6540         refs -= refs_to_drop;
6541
6542         if (refs > 0) {
6543                 if (extent_op)
6544                         __run_delayed_extent_op(extent_op, leaf, ei);
6545                 /*
6546                  * In the case of inline back ref, reference count will
6547                  * be updated by remove_extent_backref
6548                  */
6549                 if (iref) {
6550                         BUG_ON(!found_extent);
6551                 } else {
6552                         btrfs_set_extent_refs(leaf, ei, refs);
6553                         btrfs_mark_buffer_dirty(leaf);
6554                 }
6555                 if (found_extent) {
6556                         ret = remove_extent_backref(trans, extent_root, path,
6557                                                     iref, refs_to_drop,
6558                                                     is_data, &last_ref);
6559                         if (ret) {
6560                                 btrfs_abort_transaction(trans, extent_root, ret);
6561                                 goto out;
6562                         }
6563                 }
6564                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6565                                  root_objectid);
6566         } else {
6567                 if (found_extent) {
6568                         BUG_ON(is_data && refs_to_drop !=
6569                                extent_data_ref_count(path, iref));
6570                         if (iref) {
6571                                 BUG_ON(path->slots[0] != extent_slot);
6572                         } else {
6573                                 BUG_ON(path->slots[0] != extent_slot + 1);
6574                                 path->slots[0] = extent_slot;
6575                                 num_to_del = 2;
6576                         }
6577                 }
6578
6579                 last_ref = 1;
6580                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6581                                       num_to_del);
6582                 if (ret) {
6583                         btrfs_abort_transaction(trans, extent_root, ret);
6584                         goto out;
6585                 }
6586                 btrfs_release_path(path);
6587
6588                 if (is_data) {
6589                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6590                         if (ret) {
6591                                 btrfs_abort_transaction(trans, extent_root, ret);
6592                                 goto out;
6593                         }
6594                 }
6595
6596                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6597                 if (ret) {
6598                         btrfs_abort_transaction(trans, extent_root, ret);
6599                         goto out;
6600                 }
6601         }
6602         btrfs_release_path(path);
6603
6604 out:
6605         btrfs_free_path(path);
6606         return ret;
6607 }
6608
6609 /*
6610  * when we free an block, it is possible (and likely) that we free the last
6611  * delayed ref for that extent as well.  This searches the delayed ref tree for
6612  * a given extent, and if there are no other delayed refs to be processed, it
6613  * removes it from the tree.
6614  */
6615 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6616                                       struct btrfs_root *root, u64 bytenr)
6617 {
6618         struct btrfs_delayed_ref_head *head;
6619         struct btrfs_delayed_ref_root *delayed_refs;
6620         int ret = 0;
6621
6622         delayed_refs = &trans->transaction->delayed_refs;
6623         spin_lock(&delayed_refs->lock);
6624         head = btrfs_find_delayed_ref_head(trans, bytenr);
6625         if (!head)
6626                 goto out_delayed_unlock;
6627
6628         spin_lock(&head->lock);
6629         if (!list_empty(&head->ref_list))
6630                 goto out;
6631
6632         if (head->extent_op) {
6633                 if (!head->must_insert_reserved)
6634                         goto out;
6635                 btrfs_free_delayed_extent_op(head->extent_op);
6636                 head->extent_op = NULL;
6637         }
6638
6639         /*
6640          * waiting for the lock here would deadlock.  If someone else has it
6641          * locked they are already in the process of dropping it anyway
6642          */
6643         if (!mutex_trylock(&head->mutex))
6644                 goto out;
6645
6646         /*
6647          * at this point we have a head with no other entries.  Go
6648          * ahead and process it.
6649          */
6650         head->node.in_tree = 0;
6651         rb_erase(&head->href_node, &delayed_refs->href_root);
6652
6653         atomic_dec(&delayed_refs->num_entries);
6654
6655         /*
6656          * we don't take a ref on the node because we're removing it from the
6657          * tree, so we just steal the ref the tree was holding.
6658          */
6659         delayed_refs->num_heads--;
6660         if (head->processing == 0)
6661                 delayed_refs->num_heads_ready--;
6662         head->processing = 0;
6663         spin_unlock(&head->lock);
6664         spin_unlock(&delayed_refs->lock);
6665
6666         BUG_ON(head->extent_op);
6667         if (head->must_insert_reserved)
6668                 ret = 1;
6669
6670         mutex_unlock(&head->mutex);
6671         btrfs_put_delayed_ref(&head->node);
6672         return ret;
6673 out:
6674         spin_unlock(&head->lock);
6675
6676 out_delayed_unlock:
6677         spin_unlock(&delayed_refs->lock);
6678         return 0;
6679 }
6680
6681 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6682                            struct btrfs_root *root,
6683                            struct extent_buffer *buf,
6684                            u64 parent, int last_ref)
6685 {
6686         int pin = 1;
6687         int ret;
6688
6689         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6690                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6691                                         buf->start, buf->len,
6692                                         parent, root->root_key.objectid,
6693                                         btrfs_header_level(buf),
6694                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6695                 BUG_ON(ret); /* -ENOMEM */
6696         }
6697
6698         if (!last_ref)
6699                 return;
6700
6701         if (btrfs_header_generation(buf) == trans->transid) {
6702                 struct btrfs_block_group_cache *cache;
6703
6704                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6705                         ret = check_ref_cleanup(trans, root, buf->start);
6706                         if (!ret)
6707                                 goto out;
6708                 }
6709
6710                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6711
6712                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6713                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6714                         btrfs_put_block_group(cache);
6715                         goto out;
6716                 }
6717
6718                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6719
6720                 btrfs_add_free_space(cache, buf->start, buf->len);
6721                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6722                 btrfs_put_block_group(cache);
6723                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6724                 pin = 0;
6725         }
6726 out:
6727         if (pin)
6728                 add_pinned_bytes(root->fs_info, buf->len,
6729                                  btrfs_header_level(buf),
6730                                  root->root_key.objectid);
6731
6732         /*
6733          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6734          * anymore.
6735          */
6736         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6737 }
6738
6739 /* Can return -ENOMEM */
6740 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6741                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6742                       u64 owner, u64 offset, int no_quota)
6743 {
6744         int ret;
6745         struct btrfs_fs_info *fs_info = root->fs_info;
6746
6747         if (btrfs_test_is_dummy_root(root))
6748                 return 0;
6749
6750         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6751
6752         /*
6753          * tree log blocks never actually go into the extent allocation
6754          * tree, just update pinning info and exit early.
6755          */
6756         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6757                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6758                 /* unlocks the pinned mutex */
6759                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6760                 ret = 0;
6761         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6762                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6763                                         num_bytes,
6764                                         parent, root_objectid, (int)owner,
6765                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6766         } else {
6767                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6768                                                 num_bytes,
6769                                                 parent, root_objectid, owner,
6770                                                 offset, BTRFS_DROP_DELAYED_REF,
6771                                                 NULL, no_quota);
6772         }
6773         return ret;
6774 }
6775
6776 /*
6777  * when we wait for progress in the block group caching, its because
6778  * our allocation attempt failed at least once.  So, we must sleep
6779  * and let some progress happen before we try again.
6780  *
6781  * This function will sleep at least once waiting for new free space to
6782  * show up, and then it will check the block group free space numbers
6783  * for our min num_bytes.  Another option is to have it go ahead
6784  * and look in the rbtree for a free extent of a given size, but this
6785  * is a good start.
6786  *
6787  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6788  * any of the information in this block group.
6789  */
6790 static noinline void
6791 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6792                                 u64 num_bytes)
6793 {
6794         struct btrfs_caching_control *caching_ctl;
6795
6796         caching_ctl = get_caching_control(cache);
6797         if (!caching_ctl)
6798                 return;
6799
6800         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6801                    (cache->free_space_ctl->free_space >= num_bytes));
6802
6803         put_caching_control(caching_ctl);
6804 }
6805
6806 static noinline int
6807 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6808 {
6809         struct btrfs_caching_control *caching_ctl;
6810         int ret = 0;
6811
6812         caching_ctl = get_caching_control(cache);
6813         if (!caching_ctl)
6814                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6815
6816         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6817         if (cache->cached == BTRFS_CACHE_ERROR)
6818                 ret = -EIO;
6819         put_caching_control(caching_ctl);
6820         return ret;
6821 }
6822
6823 int __get_raid_index(u64 flags)
6824 {
6825         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6826                 return BTRFS_RAID_RAID10;
6827         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6828                 return BTRFS_RAID_RAID1;
6829         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6830                 return BTRFS_RAID_DUP;
6831         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6832                 return BTRFS_RAID_RAID0;
6833         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6834                 return BTRFS_RAID_RAID5;
6835         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6836                 return BTRFS_RAID_RAID6;
6837
6838         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6839 }
6840
6841 int get_block_group_index(struct btrfs_block_group_cache *cache)
6842 {
6843         return __get_raid_index(cache->flags);
6844 }
6845
6846 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6847         [BTRFS_RAID_RAID10]     = "raid10",
6848         [BTRFS_RAID_RAID1]      = "raid1",
6849         [BTRFS_RAID_DUP]        = "dup",
6850         [BTRFS_RAID_RAID0]      = "raid0",
6851         [BTRFS_RAID_SINGLE]     = "single",
6852         [BTRFS_RAID_RAID5]      = "raid5",
6853         [BTRFS_RAID_RAID6]      = "raid6",
6854 };
6855
6856 static const char *get_raid_name(enum btrfs_raid_types type)
6857 {
6858         if (type >= BTRFS_NR_RAID_TYPES)
6859                 return NULL;
6860
6861         return btrfs_raid_type_names[type];
6862 }
6863
6864 enum btrfs_loop_type {
6865         LOOP_CACHING_NOWAIT = 0,
6866         LOOP_CACHING_WAIT = 1,
6867         LOOP_ALLOC_CHUNK = 2,
6868         LOOP_NO_EMPTY_SIZE = 3,
6869 };
6870
6871 static inline void
6872 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6873                        int delalloc)
6874 {
6875         if (delalloc)
6876                 down_read(&cache->data_rwsem);
6877 }
6878
6879 static inline void
6880 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6881                        int delalloc)
6882 {
6883         btrfs_get_block_group(cache);
6884         if (delalloc)
6885                 down_read(&cache->data_rwsem);
6886 }
6887
6888 static struct btrfs_block_group_cache *
6889 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6890                    struct btrfs_free_cluster *cluster,
6891                    int delalloc)
6892 {
6893         struct btrfs_block_group_cache *used_bg;
6894         bool locked = false;
6895 again:
6896         spin_lock(&cluster->refill_lock);
6897         if (locked) {
6898                 if (used_bg == cluster->block_group)
6899                         return used_bg;
6900
6901                 up_read(&used_bg->data_rwsem);
6902                 btrfs_put_block_group(used_bg);
6903         }
6904
6905         used_bg = cluster->block_group;
6906         if (!used_bg)
6907                 return NULL;
6908
6909         if (used_bg == block_group)
6910                 return used_bg;
6911
6912         btrfs_get_block_group(used_bg);
6913
6914         if (!delalloc)
6915                 return used_bg;
6916
6917         if (down_read_trylock(&used_bg->data_rwsem))
6918                 return used_bg;
6919
6920         spin_unlock(&cluster->refill_lock);
6921         down_read(&used_bg->data_rwsem);
6922         locked = true;
6923         goto again;
6924 }
6925
6926 static inline void
6927 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6928                          int delalloc)
6929 {
6930         if (delalloc)
6931                 up_read(&cache->data_rwsem);
6932         btrfs_put_block_group(cache);
6933 }
6934
6935 /*
6936  * walks the btree of allocated extents and find a hole of a given size.
6937  * The key ins is changed to record the hole:
6938  * ins->objectid == start position
6939  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6940  * ins->offset == the size of the hole.
6941  * Any available blocks before search_start are skipped.
6942  *
6943  * If there is no suitable free space, we will record the max size of
6944  * the free space extent currently.
6945  */
6946 static noinline int find_free_extent(struct btrfs_root *orig_root,
6947                                      u64 num_bytes, u64 empty_size,
6948                                      u64 hint_byte, struct btrfs_key *ins,
6949                                      u64 flags, int delalloc)
6950 {
6951         int ret = 0;
6952         struct btrfs_root *root = orig_root->fs_info->extent_root;
6953         struct btrfs_free_cluster *last_ptr = NULL;
6954         struct btrfs_block_group_cache *block_group = NULL;
6955         u64 search_start = 0;
6956         u64 max_extent_size = 0;
6957         int empty_cluster = 2 * 1024 * 1024;
6958         struct btrfs_space_info *space_info;
6959         int loop = 0;
6960         int index = __get_raid_index(flags);
6961         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6962                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6963         bool failed_cluster_refill = false;
6964         bool failed_alloc = false;
6965         bool use_cluster = true;
6966         bool have_caching_bg = false;
6967
6968         WARN_ON(num_bytes < root->sectorsize);
6969         ins->type = BTRFS_EXTENT_ITEM_KEY;
6970         ins->objectid = 0;
6971         ins->offset = 0;
6972
6973         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6974
6975         space_info = __find_space_info(root->fs_info, flags);
6976         if (!space_info) {
6977                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6978                 return -ENOSPC;
6979         }
6980
6981         /*
6982          * If the space info is for both data and metadata it means we have a
6983          * small filesystem and we can't use the clustering stuff.
6984          */
6985         if (btrfs_mixed_space_info(space_info))
6986                 use_cluster = false;
6987
6988         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6989                 last_ptr = &root->fs_info->meta_alloc_cluster;
6990                 if (!btrfs_test_opt(root, SSD))
6991                         empty_cluster = 64 * 1024;
6992         }
6993
6994         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6995             btrfs_test_opt(root, SSD)) {
6996                 last_ptr = &root->fs_info->data_alloc_cluster;
6997         }
6998
6999         if (last_ptr) {
7000                 spin_lock(&last_ptr->lock);
7001                 if (last_ptr->block_group)
7002                         hint_byte = last_ptr->window_start;
7003                 spin_unlock(&last_ptr->lock);
7004         }
7005
7006         search_start = max(search_start, first_logical_byte(root, 0));
7007         search_start = max(search_start, hint_byte);
7008
7009         if (!last_ptr)
7010                 empty_cluster = 0;
7011
7012         if (search_start == hint_byte) {
7013                 block_group = btrfs_lookup_block_group(root->fs_info,
7014                                                        search_start);
7015                 /*
7016                  * we don't want to use the block group if it doesn't match our
7017                  * allocation bits, or if its not cached.
7018                  *
7019                  * However if we are re-searching with an ideal block group
7020                  * picked out then we don't care that the block group is cached.
7021                  */
7022                 if (block_group && block_group_bits(block_group, flags) &&
7023                     block_group->cached != BTRFS_CACHE_NO) {
7024                         down_read(&space_info->groups_sem);
7025                         if (list_empty(&block_group->list) ||
7026                             block_group->ro) {
7027                                 /*
7028                                  * someone is removing this block group,
7029                                  * we can't jump into the have_block_group
7030                                  * target because our list pointers are not
7031                                  * valid
7032                                  */
7033                                 btrfs_put_block_group(block_group);
7034                                 up_read(&space_info->groups_sem);
7035                         } else {
7036                                 index = get_block_group_index(block_group);
7037                                 btrfs_lock_block_group(block_group, delalloc);
7038                                 goto have_block_group;
7039                         }
7040                 } else if (block_group) {
7041                         btrfs_put_block_group(block_group);
7042                 }
7043         }
7044 search:
7045         have_caching_bg = false;
7046         down_read(&space_info->groups_sem);
7047         list_for_each_entry(block_group, &space_info->block_groups[index],
7048                             list) {
7049                 u64 offset;
7050                 int cached;
7051
7052                 btrfs_grab_block_group(block_group, delalloc);
7053                 search_start = block_group->key.objectid;
7054
7055                 /*
7056                  * this can happen if we end up cycling through all the
7057                  * raid types, but we want to make sure we only allocate
7058                  * for the proper type.
7059                  */
7060                 if (!block_group_bits(block_group, flags)) {
7061                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7062                                 BTRFS_BLOCK_GROUP_RAID1 |
7063                                 BTRFS_BLOCK_GROUP_RAID5 |
7064                                 BTRFS_BLOCK_GROUP_RAID6 |
7065                                 BTRFS_BLOCK_GROUP_RAID10;
7066
7067                         /*
7068                          * if they asked for extra copies and this block group
7069                          * doesn't provide them, bail.  This does allow us to
7070                          * fill raid0 from raid1.
7071                          */
7072                         if ((flags & extra) && !(block_group->flags & extra))
7073                                 goto loop;
7074                 }
7075
7076 have_block_group:
7077                 cached = block_group_cache_done(block_group);
7078                 if (unlikely(!cached)) {
7079                         ret = cache_block_group(block_group, 0);
7080                         BUG_ON(ret < 0);
7081                         ret = 0;
7082                 }
7083
7084                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7085                         goto loop;
7086                 if (unlikely(block_group->ro))
7087                         goto loop;
7088
7089                 /*
7090                  * Ok we want to try and use the cluster allocator, so
7091                  * lets look there
7092                  */
7093                 if (last_ptr) {
7094                         struct btrfs_block_group_cache *used_block_group;
7095                         unsigned long aligned_cluster;
7096                         /*
7097                          * the refill lock keeps out other
7098                          * people trying to start a new cluster
7099                          */
7100                         used_block_group = btrfs_lock_cluster(block_group,
7101                                                               last_ptr,
7102                                                               delalloc);
7103                         if (!used_block_group)
7104                                 goto refill_cluster;
7105
7106                         if (used_block_group != block_group &&
7107                             (used_block_group->ro ||
7108                              !block_group_bits(used_block_group, flags)))
7109                                 goto release_cluster;
7110
7111                         offset = btrfs_alloc_from_cluster(used_block_group,
7112                                                 last_ptr,
7113                                                 num_bytes,
7114                                                 used_block_group->key.objectid,
7115                                                 &max_extent_size);
7116                         if (offset) {
7117                                 /* we have a block, we're done */
7118                                 spin_unlock(&last_ptr->refill_lock);
7119                                 trace_btrfs_reserve_extent_cluster(root,
7120                                                 used_block_group,
7121                                                 search_start, num_bytes);
7122                                 if (used_block_group != block_group) {
7123                                         btrfs_release_block_group(block_group,
7124                                                                   delalloc);
7125                                         block_group = used_block_group;
7126                                 }
7127                                 goto checks;
7128                         }
7129
7130                         WARN_ON(last_ptr->block_group != used_block_group);
7131 release_cluster:
7132                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7133                          * set up a new clusters, so lets just skip it
7134                          * and let the allocator find whatever block
7135                          * it can find.  If we reach this point, we
7136                          * will have tried the cluster allocator
7137                          * plenty of times and not have found
7138                          * anything, so we are likely way too
7139                          * fragmented for the clustering stuff to find
7140                          * anything.
7141                          *
7142                          * However, if the cluster is taken from the
7143                          * current block group, release the cluster
7144                          * first, so that we stand a better chance of
7145                          * succeeding in the unclustered
7146                          * allocation.  */
7147                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7148                             used_block_group != block_group) {
7149                                 spin_unlock(&last_ptr->refill_lock);
7150                                 btrfs_release_block_group(used_block_group,
7151                                                           delalloc);
7152                                 goto unclustered_alloc;
7153                         }
7154
7155                         /*
7156                          * this cluster didn't work out, free it and
7157                          * start over
7158                          */
7159                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7160
7161                         if (used_block_group != block_group)
7162                                 btrfs_release_block_group(used_block_group,
7163                                                           delalloc);
7164 refill_cluster:
7165                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7166                                 spin_unlock(&last_ptr->refill_lock);
7167                                 goto unclustered_alloc;
7168                         }
7169
7170                         aligned_cluster = max_t(unsigned long,
7171                                                 empty_cluster + empty_size,
7172                                               block_group->full_stripe_len);
7173
7174                         /* allocate a cluster in this block group */
7175                         ret = btrfs_find_space_cluster(root, block_group,
7176                                                        last_ptr, search_start,
7177                                                        num_bytes,
7178                                                        aligned_cluster);
7179                         if (ret == 0) {
7180                                 /*
7181                                  * now pull our allocation out of this
7182                                  * cluster
7183                                  */
7184                                 offset = btrfs_alloc_from_cluster(block_group,
7185                                                         last_ptr,
7186                                                         num_bytes,
7187                                                         search_start,
7188                                                         &max_extent_size);
7189                                 if (offset) {
7190                                         /* we found one, proceed */
7191                                         spin_unlock(&last_ptr->refill_lock);
7192                                         trace_btrfs_reserve_extent_cluster(root,
7193                                                 block_group, search_start,
7194                                                 num_bytes);
7195                                         goto checks;
7196                                 }
7197                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7198                                    && !failed_cluster_refill) {
7199                                 spin_unlock(&last_ptr->refill_lock);
7200
7201                                 failed_cluster_refill = true;
7202                                 wait_block_group_cache_progress(block_group,
7203                                        num_bytes + empty_cluster + empty_size);
7204                                 goto have_block_group;
7205                         }
7206
7207                         /*
7208                          * at this point we either didn't find a cluster
7209                          * or we weren't able to allocate a block from our
7210                          * cluster.  Free the cluster we've been trying
7211                          * to use, and go to the next block group
7212                          */
7213                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7214                         spin_unlock(&last_ptr->refill_lock);
7215                         goto loop;
7216                 }
7217
7218 unclustered_alloc:
7219                 spin_lock(&block_group->free_space_ctl->tree_lock);
7220                 if (cached &&
7221                     block_group->free_space_ctl->free_space <
7222                     num_bytes + empty_cluster + empty_size) {
7223                         if (block_group->free_space_ctl->free_space >
7224                             max_extent_size)
7225                                 max_extent_size =
7226                                         block_group->free_space_ctl->free_space;
7227                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7228                         goto loop;
7229                 }
7230                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7231
7232                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7233                                                     num_bytes, empty_size,
7234                                                     &max_extent_size);
7235                 /*
7236                  * If we didn't find a chunk, and we haven't failed on this
7237                  * block group before, and this block group is in the middle of
7238                  * caching and we are ok with waiting, then go ahead and wait
7239                  * for progress to be made, and set failed_alloc to true.
7240                  *
7241                  * If failed_alloc is true then we've already waited on this
7242                  * block group once and should move on to the next block group.
7243                  */
7244                 if (!offset && !failed_alloc && !cached &&
7245                     loop > LOOP_CACHING_NOWAIT) {
7246                         wait_block_group_cache_progress(block_group,
7247                                                 num_bytes + empty_size);
7248                         failed_alloc = true;
7249                         goto have_block_group;
7250                 } else if (!offset) {
7251                         if (!cached)
7252                                 have_caching_bg = true;
7253                         goto loop;
7254                 }
7255 checks:
7256                 search_start = ALIGN(offset, root->stripesize);
7257
7258                 /* move on to the next group */
7259                 if (search_start + num_bytes >
7260                     block_group->key.objectid + block_group->key.offset) {
7261                         btrfs_add_free_space(block_group, offset, num_bytes);
7262                         goto loop;
7263                 }
7264
7265                 if (offset < search_start)
7266                         btrfs_add_free_space(block_group, offset,
7267                                              search_start - offset);
7268                 BUG_ON(offset > search_start);
7269
7270                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7271                                                   alloc_type, delalloc);
7272                 if (ret == -EAGAIN) {
7273                         btrfs_add_free_space(block_group, offset, num_bytes);
7274                         goto loop;
7275                 }
7276
7277                 /* we are all good, lets return */
7278                 ins->objectid = search_start;
7279                 ins->offset = num_bytes;
7280
7281                 trace_btrfs_reserve_extent(orig_root, block_group,
7282                                            search_start, num_bytes);
7283                 btrfs_release_block_group(block_group, delalloc);
7284                 break;
7285 loop:
7286                 failed_cluster_refill = false;
7287                 failed_alloc = false;
7288                 BUG_ON(index != get_block_group_index(block_group));
7289                 btrfs_release_block_group(block_group, delalloc);
7290         }
7291         up_read(&space_info->groups_sem);
7292
7293         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7294                 goto search;
7295
7296         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7297                 goto search;
7298
7299         /*
7300          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7301          *                      caching kthreads as we move along
7302          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7303          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7304          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7305          *                      again
7306          */
7307         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7308                 index = 0;
7309                 loop++;
7310                 if (loop == LOOP_ALLOC_CHUNK) {
7311                         struct btrfs_trans_handle *trans;
7312                         int exist = 0;
7313
7314                         trans = current->journal_info;
7315                         if (trans)
7316                                 exist = 1;
7317                         else
7318                                 trans = btrfs_join_transaction(root);
7319
7320                         if (IS_ERR(trans)) {
7321                                 ret = PTR_ERR(trans);
7322                                 goto out;
7323                         }
7324
7325                         ret = do_chunk_alloc(trans, root, flags,
7326                                              CHUNK_ALLOC_FORCE);
7327                         /*
7328                          * Do not bail out on ENOSPC since we
7329                          * can do more things.
7330                          */
7331                         if (ret < 0 && ret != -ENOSPC)
7332                                 btrfs_abort_transaction(trans,
7333                                                         root, ret);
7334                         else
7335                                 ret = 0;
7336                         if (!exist)
7337                                 btrfs_end_transaction(trans, root);
7338                         if (ret)
7339                                 goto out;
7340                 }
7341
7342                 if (loop == LOOP_NO_EMPTY_SIZE) {
7343                         empty_size = 0;
7344                         empty_cluster = 0;
7345                 }
7346
7347                 goto search;
7348         } else if (!ins->objectid) {
7349                 ret = -ENOSPC;
7350         } else if (ins->objectid) {
7351                 ret = 0;
7352         }
7353 out:
7354         if (ret == -ENOSPC)
7355                 ins->offset = max_extent_size;
7356         return ret;
7357 }
7358
7359 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7360                             int dump_block_groups)
7361 {
7362         struct btrfs_block_group_cache *cache;
7363         int index = 0;
7364
7365         spin_lock(&info->lock);
7366         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7367                info->flags,
7368                info->total_bytes - info->bytes_used - info->bytes_pinned -
7369                info->bytes_reserved - info->bytes_readonly,
7370                (info->full) ? "" : "not ");
7371         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7372                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7373                info->total_bytes, info->bytes_used, info->bytes_pinned,
7374                info->bytes_reserved, info->bytes_may_use,
7375                info->bytes_readonly);
7376         spin_unlock(&info->lock);
7377
7378         if (!dump_block_groups)
7379                 return;
7380
7381         down_read(&info->groups_sem);
7382 again:
7383         list_for_each_entry(cache, &info->block_groups[index], list) {
7384                 spin_lock(&cache->lock);
7385                 printk(KERN_INFO "BTRFS: "
7386                            "block group %llu has %llu bytes, "
7387                            "%llu used %llu pinned %llu reserved %s\n",
7388                        cache->key.objectid, cache->key.offset,
7389                        btrfs_block_group_used(&cache->item), cache->pinned,
7390                        cache->reserved, cache->ro ? "[readonly]" : "");
7391                 btrfs_dump_free_space(cache, bytes);
7392                 spin_unlock(&cache->lock);
7393         }
7394         if (++index < BTRFS_NR_RAID_TYPES)
7395                 goto again;
7396         up_read(&info->groups_sem);
7397 }
7398
7399 int btrfs_reserve_extent(struct btrfs_root *root,
7400                          u64 num_bytes, u64 min_alloc_size,
7401                          u64 empty_size, u64 hint_byte,
7402                          struct btrfs_key *ins, int is_data, int delalloc)
7403 {
7404         bool final_tried = false;
7405         u64 flags;
7406         int ret;
7407
7408         flags = btrfs_get_alloc_profile(root, is_data);
7409 again:
7410         WARN_ON(num_bytes < root->sectorsize);
7411         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7412                                flags, delalloc);
7413
7414         if (ret == -ENOSPC) {
7415                 if (!final_tried && ins->offset) {
7416                         num_bytes = min(num_bytes >> 1, ins->offset);
7417                         num_bytes = round_down(num_bytes, root->sectorsize);
7418                         num_bytes = max(num_bytes, min_alloc_size);
7419                         if (num_bytes == min_alloc_size)
7420                                 final_tried = true;
7421                         goto again;
7422                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7423                         struct btrfs_space_info *sinfo;
7424
7425                         sinfo = __find_space_info(root->fs_info, flags);
7426                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7427                                 flags, num_bytes);
7428                         if (sinfo)
7429                                 dump_space_info(sinfo, num_bytes, 1);
7430                 }
7431         }
7432
7433         return ret;
7434 }
7435
7436 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7437                                         u64 start, u64 len,
7438                                         int pin, int delalloc)
7439 {
7440         struct btrfs_block_group_cache *cache;
7441         int ret = 0;
7442
7443         cache = btrfs_lookup_block_group(root->fs_info, start);
7444         if (!cache) {
7445                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7446                         start);
7447                 return -ENOSPC;
7448         }
7449
7450         if (pin)
7451                 pin_down_extent(root, cache, start, len, 1);
7452         else {
7453                 if (btrfs_test_opt(root, DISCARD))
7454                         ret = btrfs_discard_extent(root, start, len, NULL);
7455                 btrfs_add_free_space(cache, start, len);
7456                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7457         }
7458
7459         btrfs_put_block_group(cache);
7460
7461         trace_btrfs_reserved_extent_free(root, start, len);
7462
7463         return ret;
7464 }
7465
7466 int btrfs_free_reserved_extent(struct btrfs_root *root,
7467                                u64 start, u64 len, int delalloc)
7468 {
7469         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7470 }
7471
7472 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7473                                        u64 start, u64 len)
7474 {
7475         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7476 }
7477
7478 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7479                                       struct btrfs_root *root,
7480                                       u64 parent, u64 root_objectid,
7481                                       u64 flags, u64 owner, u64 offset,
7482                                       struct btrfs_key *ins, int ref_mod)
7483 {
7484         int ret;
7485         struct btrfs_fs_info *fs_info = root->fs_info;
7486         struct btrfs_extent_item *extent_item;
7487         struct btrfs_extent_inline_ref *iref;
7488         struct btrfs_path *path;
7489         struct extent_buffer *leaf;
7490         int type;
7491         u32 size;
7492
7493         if (parent > 0)
7494                 type = BTRFS_SHARED_DATA_REF_KEY;
7495         else
7496                 type = BTRFS_EXTENT_DATA_REF_KEY;
7497
7498         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7499
7500         path = btrfs_alloc_path();
7501         if (!path)
7502                 return -ENOMEM;
7503
7504         path->leave_spinning = 1;
7505         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7506                                       ins, size);
7507         if (ret) {
7508                 btrfs_free_path(path);
7509                 return ret;
7510         }
7511
7512         leaf = path->nodes[0];
7513         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7514                                      struct btrfs_extent_item);
7515         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7516         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7517         btrfs_set_extent_flags(leaf, extent_item,
7518                                flags | BTRFS_EXTENT_FLAG_DATA);
7519
7520         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7521         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7522         if (parent > 0) {
7523                 struct btrfs_shared_data_ref *ref;
7524                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7525                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7526                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7527         } else {
7528                 struct btrfs_extent_data_ref *ref;
7529                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7530                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7531                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7532                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7533                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7534         }
7535
7536         btrfs_mark_buffer_dirty(path->nodes[0]);
7537         btrfs_free_path(path);
7538
7539         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7540         if (ret) { /* -ENOENT, logic error */
7541                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7542                         ins->objectid, ins->offset);
7543                 BUG();
7544         }
7545         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7546         return ret;
7547 }
7548
7549 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7550                                      struct btrfs_root *root,
7551                                      u64 parent, u64 root_objectid,
7552                                      u64 flags, struct btrfs_disk_key *key,
7553                                      int level, struct btrfs_key *ins,
7554                                      int no_quota)
7555 {
7556         int ret;
7557         struct btrfs_fs_info *fs_info = root->fs_info;
7558         struct btrfs_extent_item *extent_item;
7559         struct btrfs_tree_block_info *block_info;
7560         struct btrfs_extent_inline_ref *iref;
7561         struct btrfs_path *path;
7562         struct extent_buffer *leaf;
7563         u32 size = sizeof(*extent_item) + sizeof(*iref);
7564         u64 num_bytes = ins->offset;
7565         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7566                                                  SKINNY_METADATA);
7567
7568         if (!skinny_metadata)
7569                 size += sizeof(*block_info);
7570
7571         path = btrfs_alloc_path();
7572         if (!path) {
7573                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7574                                                    root->nodesize);
7575                 return -ENOMEM;
7576         }
7577
7578         path->leave_spinning = 1;
7579         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7580                                       ins, size);
7581         if (ret) {
7582                 btrfs_free_path(path);
7583                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7584                                                    root->nodesize);
7585                 return ret;
7586         }
7587
7588         leaf = path->nodes[0];
7589         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7590                                      struct btrfs_extent_item);
7591         btrfs_set_extent_refs(leaf, extent_item, 1);
7592         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7593         btrfs_set_extent_flags(leaf, extent_item,
7594                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7595
7596         if (skinny_metadata) {
7597                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7598                 num_bytes = root->nodesize;
7599         } else {
7600                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7601                 btrfs_set_tree_block_key(leaf, block_info, key);
7602                 btrfs_set_tree_block_level(leaf, block_info, level);
7603                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7604         }
7605
7606         if (parent > 0) {
7607                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7608                 btrfs_set_extent_inline_ref_type(leaf, iref,
7609                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7610                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7611         } else {
7612                 btrfs_set_extent_inline_ref_type(leaf, iref,
7613                                                  BTRFS_TREE_BLOCK_REF_KEY);
7614                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7615         }
7616
7617         btrfs_mark_buffer_dirty(leaf);
7618         btrfs_free_path(path);
7619
7620         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7621                                  1);
7622         if (ret) { /* -ENOENT, logic error */
7623                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7624                         ins->objectid, ins->offset);
7625                 BUG();
7626         }
7627
7628         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7629         return ret;
7630 }
7631
7632 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7633                                      struct btrfs_root *root,
7634                                      u64 root_objectid, u64 owner,
7635                                      u64 offset, struct btrfs_key *ins)
7636 {
7637         int ret;
7638
7639         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7640
7641         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7642                                          ins->offset, 0,
7643                                          root_objectid, owner, offset,
7644                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7645         return ret;
7646 }
7647
7648 /*
7649  * this is used by the tree logging recovery code.  It records that
7650  * an extent has been allocated and makes sure to clear the free
7651  * space cache bits as well
7652  */
7653 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7654                                    struct btrfs_root *root,
7655                                    u64 root_objectid, u64 owner, u64 offset,
7656                                    struct btrfs_key *ins)
7657 {
7658         int ret;
7659         struct btrfs_block_group_cache *block_group;
7660
7661         /*
7662          * Mixed block groups will exclude before processing the log so we only
7663          * need to do the exlude dance if this fs isn't mixed.
7664          */
7665         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7666                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7667                 if (ret)
7668                         return ret;
7669         }
7670
7671         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7672         if (!block_group)
7673                 return -EINVAL;
7674
7675         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7676                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7677         BUG_ON(ret); /* logic error */
7678         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7679                                          0, owner, offset, ins, 1);
7680         btrfs_put_block_group(block_group);
7681         return ret;
7682 }
7683
7684 static struct extent_buffer *
7685 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7686                       u64 bytenr, int level)
7687 {
7688         struct extent_buffer *buf;
7689
7690         buf = btrfs_find_create_tree_block(root, bytenr);
7691         if (!buf)
7692                 return ERR_PTR(-ENOMEM);
7693         btrfs_set_header_generation(buf, trans->transid);
7694         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7695         btrfs_tree_lock(buf);
7696         clean_tree_block(trans, root->fs_info, buf);
7697         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7698
7699         btrfs_set_lock_blocking(buf);
7700         btrfs_set_buffer_uptodate(buf);
7701
7702         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7703                 buf->log_index = root->log_transid % 2;
7704                 /*
7705                  * we allow two log transactions at a time, use different
7706                  * EXENT bit to differentiate dirty pages.
7707                  */
7708                 if (buf->log_index == 0)
7709                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7710                                         buf->start + buf->len - 1, GFP_NOFS);
7711                 else
7712                         set_extent_new(&root->dirty_log_pages, buf->start,
7713                                         buf->start + buf->len - 1, GFP_NOFS);
7714         } else {
7715                 buf->log_index = -1;
7716                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7717                          buf->start + buf->len - 1, GFP_NOFS);
7718         }
7719         trans->blocks_used++;
7720         /* this returns a buffer locked for blocking */
7721         return buf;
7722 }
7723
7724 static struct btrfs_block_rsv *
7725 use_block_rsv(struct btrfs_trans_handle *trans,
7726               struct btrfs_root *root, u32 blocksize)
7727 {
7728         struct btrfs_block_rsv *block_rsv;
7729         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7730         int ret;
7731         bool global_updated = false;
7732
7733         block_rsv = get_block_rsv(trans, root);
7734
7735         if (unlikely(block_rsv->size == 0))
7736                 goto try_reserve;
7737 again:
7738         ret = block_rsv_use_bytes(block_rsv, blocksize);
7739         if (!ret)
7740                 return block_rsv;
7741
7742         if (block_rsv->failfast)
7743                 return ERR_PTR(ret);
7744
7745         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7746                 global_updated = true;
7747                 update_global_block_rsv(root->fs_info);
7748                 goto again;
7749         }
7750
7751         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7752                 static DEFINE_RATELIMIT_STATE(_rs,
7753                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7754                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7755                 if (__ratelimit(&_rs))
7756                         WARN(1, KERN_DEBUG
7757                                 "BTRFS: block rsv returned %d\n", ret);
7758         }
7759 try_reserve:
7760         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7761                                      BTRFS_RESERVE_NO_FLUSH);
7762         if (!ret)
7763                 return block_rsv;
7764         /*
7765          * If we couldn't reserve metadata bytes try and use some from
7766          * the global reserve if its space type is the same as the global
7767          * reservation.
7768          */
7769         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7770             block_rsv->space_info == global_rsv->space_info) {
7771                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7772                 if (!ret)
7773                         return global_rsv;
7774         }
7775         return ERR_PTR(ret);
7776 }
7777
7778 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7779                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7780 {
7781         block_rsv_add_bytes(block_rsv, blocksize, 0);
7782         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7783 }
7784
7785 /*
7786  * finds a free extent and does all the dirty work required for allocation
7787  * returns the tree buffer or an ERR_PTR on error.
7788  */
7789 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7790                                         struct btrfs_root *root,
7791                                         u64 parent, u64 root_objectid,
7792                                         struct btrfs_disk_key *key, int level,
7793                                         u64 hint, u64 empty_size)
7794 {
7795         struct btrfs_key ins;
7796         struct btrfs_block_rsv *block_rsv;
7797         struct extent_buffer *buf;
7798         struct btrfs_delayed_extent_op *extent_op;
7799         u64 flags = 0;
7800         int ret;
7801         u32 blocksize = root->nodesize;
7802         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7803                                                  SKINNY_METADATA);
7804
7805         if (btrfs_test_is_dummy_root(root)) {
7806                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7807                                             level);
7808                 if (!IS_ERR(buf))
7809                         root->alloc_bytenr += blocksize;
7810                 return buf;
7811         }
7812
7813         block_rsv = use_block_rsv(trans, root, blocksize);
7814         if (IS_ERR(block_rsv))
7815                 return ERR_CAST(block_rsv);
7816
7817         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7818                                    empty_size, hint, &ins, 0, 0);
7819         if (ret)
7820                 goto out_unuse;
7821
7822         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
7823         if (IS_ERR(buf)) {
7824                 ret = PTR_ERR(buf);
7825                 goto out_free_reserved;
7826         }
7827
7828         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7829                 if (parent == 0)
7830                         parent = ins.objectid;
7831                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7832         } else
7833                 BUG_ON(parent > 0);
7834
7835         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7836                 extent_op = btrfs_alloc_delayed_extent_op();
7837                 if (!extent_op) {
7838                         ret = -ENOMEM;
7839                         goto out_free_buf;
7840                 }
7841                 if (key)
7842                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7843                 else
7844                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7845                 extent_op->flags_to_set = flags;
7846                 if (skinny_metadata)
7847                         extent_op->update_key = 0;
7848                 else
7849                         extent_op->update_key = 1;
7850                 extent_op->update_flags = 1;
7851                 extent_op->is_data = 0;
7852                 extent_op->level = level;
7853
7854                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7855                                                  ins.objectid, ins.offset,
7856                                                  parent, root_objectid, level,
7857                                                  BTRFS_ADD_DELAYED_EXTENT,
7858                                                  extent_op, 0);
7859                 if (ret)
7860                         goto out_free_delayed;
7861         }
7862         return buf;
7863
7864 out_free_delayed:
7865         btrfs_free_delayed_extent_op(extent_op);
7866 out_free_buf:
7867         free_extent_buffer(buf);
7868 out_free_reserved:
7869         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
7870 out_unuse:
7871         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7872         return ERR_PTR(ret);
7873 }
7874
7875 struct walk_control {
7876         u64 refs[BTRFS_MAX_LEVEL];
7877         u64 flags[BTRFS_MAX_LEVEL];
7878         struct btrfs_key update_progress;
7879         int stage;
7880         int level;
7881         int shared_level;
7882         int update_ref;
7883         int keep_locks;
7884         int reada_slot;
7885         int reada_count;
7886         int for_reloc;
7887 };
7888
7889 #define DROP_REFERENCE  1
7890 #define UPDATE_BACKREF  2
7891
7892 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7893                                      struct btrfs_root *root,
7894                                      struct walk_control *wc,
7895                                      struct btrfs_path *path)
7896 {
7897         u64 bytenr;
7898         u64 generation;
7899         u64 refs;
7900         u64 flags;
7901         u32 nritems;
7902         u32 blocksize;
7903         struct btrfs_key key;
7904         struct extent_buffer *eb;
7905         int ret;
7906         int slot;
7907         int nread = 0;
7908
7909         if (path->slots[wc->level] < wc->reada_slot) {
7910                 wc->reada_count = wc->reada_count * 2 / 3;
7911                 wc->reada_count = max(wc->reada_count, 2);
7912         } else {
7913                 wc->reada_count = wc->reada_count * 3 / 2;
7914                 wc->reada_count = min_t(int, wc->reada_count,
7915                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7916         }
7917
7918         eb = path->nodes[wc->level];
7919         nritems = btrfs_header_nritems(eb);
7920         blocksize = root->nodesize;
7921
7922         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7923                 if (nread >= wc->reada_count)
7924                         break;
7925
7926                 cond_resched();
7927                 bytenr = btrfs_node_blockptr(eb, slot);
7928                 generation = btrfs_node_ptr_generation(eb, slot);
7929
7930                 if (slot == path->slots[wc->level])
7931                         goto reada;
7932
7933                 if (wc->stage == UPDATE_BACKREF &&
7934                     generation <= root->root_key.offset)
7935                         continue;
7936
7937                 /* We don't lock the tree block, it's OK to be racy here */
7938                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7939                                                wc->level - 1, 1, &refs,
7940                                                &flags);
7941                 /* We don't care about errors in readahead. */
7942                 if (ret < 0)
7943                         continue;
7944                 BUG_ON(refs == 0);
7945
7946                 if (wc->stage == DROP_REFERENCE) {
7947                         if (refs == 1)
7948                                 goto reada;
7949
7950                         if (wc->level == 1 &&
7951                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7952                                 continue;
7953                         if (!wc->update_ref ||
7954                             generation <= root->root_key.offset)
7955                                 continue;
7956                         btrfs_node_key_to_cpu(eb, &key, slot);
7957                         ret = btrfs_comp_cpu_keys(&key,
7958                                                   &wc->update_progress);
7959                         if (ret < 0)
7960                                 continue;
7961                 } else {
7962                         if (wc->level == 1 &&
7963                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7964                                 continue;
7965                 }
7966 reada:
7967                 readahead_tree_block(root, bytenr);
7968                 nread++;
7969         }
7970         wc->reada_slot = slot;
7971 }
7972
7973 /*
7974  * TODO: Modify related function to add related node/leaf to dirty_extent_root,
7975  * for later qgroup accounting.
7976  *
7977  * Current, this function does nothing.
7978  */
7979 static int account_leaf_items(struct btrfs_trans_handle *trans,
7980                               struct btrfs_root *root,
7981                               struct extent_buffer *eb)
7982 {
7983         int nr = btrfs_header_nritems(eb);
7984         int i, extent_type;
7985         struct btrfs_key key;
7986         struct btrfs_file_extent_item *fi;
7987         u64 bytenr, num_bytes;
7988
7989         for (i = 0; i < nr; i++) {
7990                 btrfs_item_key_to_cpu(eb, &key, i);
7991
7992                 if (key.type != BTRFS_EXTENT_DATA_KEY)
7993                         continue;
7994
7995                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
7996                 /* filter out non qgroup-accountable extents  */
7997                 extent_type = btrfs_file_extent_type(eb, fi);
7998
7999                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
8000                         continue;
8001
8002                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
8003                 if (!bytenr)
8004                         continue;
8005
8006                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
8007         }
8008         return 0;
8009 }
8010
8011 /*
8012  * Walk up the tree from the bottom, freeing leaves and any interior
8013  * nodes which have had all slots visited. If a node (leaf or
8014  * interior) is freed, the node above it will have it's slot
8015  * incremented. The root node will never be freed.
8016  *
8017  * At the end of this function, we should have a path which has all
8018  * slots incremented to the next position for a search. If we need to
8019  * read a new node it will be NULL and the node above it will have the
8020  * correct slot selected for a later read.
8021  *
8022  * If we increment the root nodes slot counter past the number of
8023  * elements, 1 is returned to signal completion of the search.
8024  */
8025 static int adjust_slots_upwards(struct btrfs_root *root,
8026                                 struct btrfs_path *path, int root_level)
8027 {
8028         int level = 0;
8029         int nr, slot;
8030         struct extent_buffer *eb;
8031
8032         if (root_level == 0)
8033                 return 1;
8034
8035         while (level <= root_level) {
8036                 eb = path->nodes[level];
8037                 nr = btrfs_header_nritems(eb);
8038                 path->slots[level]++;
8039                 slot = path->slots[level];
8040                 if (slot >= nr || level == 0) {
8041                         /*
8042                          * Don't free the root -  we will detect this
8043                          * condition after our loop and return a
8044                          * positive value for caller to stop walking the tree.
8045                          */
8046                         if (level != root_level) {
8047                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8048                                 path->locks[level] = 0;
8049
8050                                 free_extent_buffer(eb);
8051                                 path->nodes[level] = NULL;
8052                                 path->slots[level] = 0;
8053                         }
8054                 } else {
8055                         /*
8056                          * We have a valid slot to walk back down
8057                          * from. Stop here so caller can process these
8058                          * new nodes.
8059                          */
8060                         break;
8061                 }
8062
8063                 level++;
8064         }
8065
8066         eb = path->nodes[root_level];
8067         if (path->slots[root_level] >= btrfs_header_nritems(eb))
8068                 return 1;
8069
8070         return 0;
8071 }
8072
8073 /*
8074  * root_eb is the subtree root and is locked before this function is called.
8075  * TODO: Modify this function to mark all (including complete shared node)
8076  * to dirty_extent_root to allow it get accounted in qgroup.
8077  */
8078 static int account_shared_subtree(struct btrfs_trans_handle *trans,
8079                                   struct btrfs_root *root,
8080                                   struct extent_buffer *root_eb,
8081                                   u64 root_gen,
8082                                   int root_level)
8083 {
8084         int ret = 0;
8085         int level;
8086         struct extent_buffer *eb = root_eb;
8087         struct btrfs_path *path = NULL;
8088
8089         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
8090         BUG_ON(root_eb == NULL);
8091
8092         if (!root->fs_info->quota_enabled)
8093                 return 0;
8094
8095         if (!extent_buffer_uptodate(root_eb)) {
8096                 ret = btrfs_read_buffer(root_eb, root_gen);
8097                 if (ret)
8098                         goto out;
8099         }
8100
8101         if (root_level == 0) {
8102                 ret = account_leaf_items(trans, root, root_eb);
8103                 goto out;
8104         }
8105
8106         path = btrfs_alloc_path();
8107         if (!path)
8108                 return -ENOMEM;
8109
8110         /*
8111          * Walk down the tree.  Missing extent blocks are filled in as
8112          * we go. Metadata is accounted every time we read a new
8113          * extent block.
8114          *
8115          * When we reach a leaf, we account for file extent items in it,
8116          * walk back up the tree (adjusting slot pointers as we go)
8117          * and restart the search process.
8118          */
8119         extent_buffer_get(root_eb); /* For path */
8120         path->nodes[root_level] = root_eb;
8121         path->slots[root_level] = 0;
8122         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
8123 walk_down:
8124         level = root_level;
8125         while (level >= 0) {
8126                 if (path->nodes[level] == NULL) {
8127                         int parent_slot;
8128                         u64 child_gen;
8129                         u64 child_bytenr;
8130
8131                         /* We need to get child blockptr/gen from
8132                          * parent before we can read it. */
8133                         eb = path->nodes[level + 1];
8134                         parent_slot = path->slots[level + 1];
8135                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8136                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8137
8138                         eb = read_tree_block(root, child_bytenr, child_gen);
8139                         if (IS_ERR(eb)) {
8140                                 ret = PTR_ERR(eb);
8141                                 goto out;
8142                         } else if (!extent_buffer_uptodate(eb)) {
8143                                 free_extent_buffer(eb);
8144                                 ret = -EIO;
8145                                 goto out;
8146                         }
8147
8148                         path->nodes[level] = eb;
8149                         path->slots[level] = 0;
8150
8151                         btrfs_tree_read_lock(eb);
8152                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8153                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8154                 }
8155
8156                 if (level == 0) {
8157                         ret = account_leaf_items(trans, root, path->nodes[level]);
8158                         if (ret)
8159                                 goto out;
8160
8161                         /* Nonzero return here means we completed our search */
8162                         ret = adjust_slots_upwards(root, path, root_level);
8163                         if (ret)
8164                                 break;
8165
8166                         /* Restart search with new slots */
8167                         goto walk_down;
8168                 }
8169
8170                 level--;
8171         }
8172
8173         ret = 0;
8174 out:
8175         btrfs_free_path(path);
8176
8177         return ret;
8178 }
8179
8180 /*
8181  * helper to process tree block while walking down the tree.
8182  *
8183  * when wc->stage == UPDATE_BACKREF, this function updates
8184  * back refs for pointers in the block.
8185  *
8186  * NOTE: return value 1 means we should stop walking down.
8187  */
8188 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8189                                    struct btrfs_root *root,
8190                                    struct btrfs_path *path,
8191                                    struct walk_control *wc, int lookup_info)
8192 {
8193         int level = wc->level;
8194         struct extent_buffer *eb = path->nodes[level];
8195         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8196         int ret;
8197
8198         if (wc->stage == UPDATE_BACKREF &&
8199             btrfs_header_owner(eb) != root->root_key.objectid)
8200                 return 1;
8201
8202         /*
8203          * when reference count of tree block is 1, it won't increase
8204          * again. once full backref flag is set, we never clear it.
8205          */
8206         if (lookup_info &&
8207             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8208              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8209                 BUG_ON(!path->locks[level]);
8210                 ret = btrfs_lookup_extent_info(trans, root,
8211                                                eb->start, level, 1,
8212                                                &wc->refs[level],
8213                                                &wc->flags[level]);
8214                 BUG_ON(ret == -ENOMEM);
8215                 if (ret)
8216                         return ret;
8217                 BUG_ON(wc->refs[level] == 0);
8218         }
8219
8220         if (wc->stage == DROP_REFERENCE) {
8221                 if (wc->refs[level] > 1)
8222                         return 1;
8223
8224                 if (path->locks[level] && !wc->keep_locks) {
8225                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8226                         path->locks[level] = 0;
8227                 }
8228                 return 0;
8229         }
8230
8231         /* wc->stage == UPDATE_BACKREF */
8232         if (!(wc->flags[level] & flag)) {
8233                 BUG_ON(!path->locks[level]);
8234                 ret = btrfs_inc_ref(trans, root, eb, 1);
8235                 BUG_ON(ret); /* -ENOMEM */
8236                 ret = btrfs_dec_ref(trans, root, eb, 0);
8237                 BUG_ON(ret); /* -ENOMEM */
8238                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8239                                                   eb->len, flag,
8240                                                   btrfs_header_level(eb), 0);
8241                 BUG_ON(ret); /* -ENOMEM */
8242                 wc->flags[level] |= flag;
8243         }
8244
8245         /*
8246          * the block is shared by multiple trees, so it's not good to
8247          * keep the tree lock
8248          */
8249         if (path->locks[level] && level > 0) {
8250                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8251                 path->locks[level] = 0;
8252         }
8253         return 0;
8254 }
8255
8256 /*
8257  * helper to process tree block pointer.
8258  *
8259  * when wc->stage == DROP_REFERENCE, this function checks
8260  * reference count of the block pointed to. if the block
8261  * is shared and we need update back refs for the subtree
8262  * rooted at the block, this function changes wc->stage to
8263  * UPDATE_BACKREF. if the block is shared and there is no
8264  * need to update back, this function drops the reference
8265  * to the block.
8266  *
8267  * NOTE: return value 1 means we should stop walking down.
8268  */
8269 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8270                                  struct btrfs_root *root,
8271                                  struct btrfs_path *path,
8272                                  struct walk_control *wc, int *lookup_info)
8273 {
8274         u64 bytenr;
8275         u64 generation;
8276         u64 parent;
8277         u32 blocksize;
8278         struct btrfs_key key;
8279         struct extent_buffer *next;
8280         int level = wc->level;
8281         int reada = 0;
8282         int ret = 0;
8283         bool need_account = false;
8284
8285         generation = btrfs_node_ptr_generation(path->nodes[level],
8286                                                path->slots[level]);
8287         /*
8288          * if the lower level block was created before the snapshot
8289          * was created, we know there is no need to update back refs
8290          * for the subtree
8291          */
8292         if (wc->stage == UPDATE_BACKREF &&
8293             generation <= root->root_key.offset) {
8294                 *lookup_info = 1;
8295                 return 1;
8296         }
8297
8298         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8299         blocksize = root->nodesize;
8300
8301         next = btrfs_find_tree_block(root->fs_info, bytenr);
8302         if (!next) {
8303                 next = btrfs_find_create_tree_block(root, bytenr);
8304                 if (!next)
8305                         return -ENOMEM;
8306                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8307                                                level - 1);
8308                 reada = 1;
8309         }
8310         btrfs_tree_lock(next);
8311         btrfs_set_lock_blocking(next);
8312
8313         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8314                                        &wc->refs[level - 1],
8315                                        &wc->flags[level - 1]);
8316         if (ret < 0) {
8317                 btrfs_tree_unlock(next);
8318                 return ret;
8319         }
8320
8321         if (unlikely(wc->refs[level - 1] == 0)) {
8322                 btrfs_err(root->fs_info, "Missing references.");
8323                 BUG();
8324         }
8325         *lookup_info = 0;
8326
8327         if (wc->stage == DROP_REFERENCE) {
8328                 if (wc->refs[level - 1] > 1) {
8329                         need_account = true;
8330                         if (level == 1 &&
8331                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8332                                 goto skip;
8333
8334                         if (!wc->update_ref ||
8335                             generation <= root->root_key.offset)
8336                                 goto skip;
8337
8338                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8339                                               path->slots[level]);
8340                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8341                         if (ret < 0)
8342                                 goto skip;
8343
8344                         wc->stage = UPDATE_BACKREF;
8345                         wc->shared_level = level - 1;
8346                 }
8347         } else {
8348                 if (level == 1 &&
8349                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8350                         goto skip;
8351         }
8352
8353         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8354                 btrfs_tree_unlock(next);
8355                 free_extent_buffer(next);
8356                 next = NULL;
8357                 *lookup_info = 1;
8358         }
8359
8360         if (!next) {
8361                 if (reada && level == 1)
8362                         reada_walk_down(trans, root, wc, path);
8363                 next = read_tree_block(root, bytenr, generation);
8364                 if (IS_ERR(next)) {
8365                         return PTR_ERR(next);
8366                 } else if (!extent_buffer_uptodate(next)) {
8367                         free_extent_buffer(next);
8368                         return -EIO;
8369                 }
8370                 btrfs_tree_lock(next);
8371                 btrfs_set_lock_blocking(next);
8372         }
8373
8374         level--;
8375         BUG_ON(level != btrfs_header_level(next));
8376         path->nodes[level] = next;
8377         path->slots[level] = 0;
8378         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8379         wc->level = level;
8380         if (wc->level == 1)
8381                 wc->reada_slot = 0;
8382         return 0;
8383 skip:
8384         wc->refs[level - 1] = 0;
8385         wc->flags[level - 1] = 0;
8386         if (wc->stage == DROP_REFERENCE) {
8387                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8388                         parent = path->nodes[level]->start;
8389                 } else {
8390                         BUG_ON(root->root_key.objectid !=
8391                                btrfs_header_owner(path->nodes[level]));
8392                         parent = 0;
8393                 }
8394
8395                 if (need_account) {
8396                         ret = account_shared_subtree(trans, root, next,
8397                                                      generation, level - 1);
8398                         if (ret) {
8399                                 btrfs_err_rl(root->fs_info,
8400                                         "Error "
8401                                         "%d accounting shared subtree. Quota "
8402                                         "is out of sync, rescan required.",
8403                                         ret);
8404                         }
8405                 }
8406                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8407                                 root->root_key.objectid, level - 1, 0, 0);
8408                 BUG_ON(ret); /* -ENOMEM */
8409         }
8410         btrfs_tree_unlock(next);
8411         free_extent_buffer(next);
8412         *lookup_info = 1;
8413         return 1;
8414 }
8415
8416 /*
8417  * helper to process tree block while walking up the tree.
8418  *
8419  * when wc->stage == DROP_REFERENCE, this function drops
8420  * reference count on the block.
8421  *
8422  * when wc->stage == UPDATE_BACKREF, this function changes
8423  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8424  * to UPDATE_BACKREF previously while processing the block.
8425  *
8426  * NOTE: return value 1 means we should stop walking up.
8427  */
8428 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8429                                  struct btrfs_root *root,
8430                                  struct btrfs_path *path,
8431                                  struct walk_control *wc)
8432 {
8433         int ret;
8434         int level = wc->level;
8435         struct extent_buffer *eb = path->nodes[level];
8436         u64 parent = 0;
8437
8438         if (wc->stage == UPDATE_BACKREF) {
8439                 BUG_ON(wc->shared_level < level);
8440                 if (level < wc->shared_level)
8441                         goto out;
8442
8443                 ret = find_next_key(path, level + 1, &wc->update_progress);
8444                 if (ret > 0)
8445                         wc->update_ref = 0;
8446
8447                 wc->stage = DROP_REFERENCE;
8448                 wc->shared_level = -1;
8449                 path->slots[level] = 0;
8450
8451                 /*
8452                  * check reference count again if the block isn't locked.
8453                  * we should start walking down the tree again if reference
8454                  * count is one.
8455                  */
8456                 if (!path->locks[level]) {
8457                         BUG_ON(level == 0);
8458                         btrfs_tree_lock(eb);
8459                         btrfs_set_lock_blocking(eb);
8460                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8461
8462                         ret = btrfs_lookup_extent_info(trans, root,
8463                                                        eb->start, level, 1,
8464                                                        &wc->refs[level],
8465                                                        &wc->flags[level]);
8466                         if (ret < 0) {
8467                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8468                                 path->locks[level] = 0;
8469                                 return ret;
8470                         }
8471                         BUG_ON(wc->refs[level] == 0);
8472                         if (wc->refs[level] == 1) {
8473                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8474                                 path->locks[level] = 0;
8475                                 return 1;
8476                         }
8477                 }
8478         }
8479
8480         /* wc->stage == DROP_REFERENCE */
8481         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8482
8483         if (wc->refs[level] == 1) {
8484                 if (level == 0) {
8485                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8486                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8487                         else
8488                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8489                         BUG_ON(ret); /* -ENOMEM */
8490                         ret = account_leaf_items(trans, root, eb);
8491                         if (ret) {
8492                                 btrfs_err_rl(root->fs_info,
8493                                         "error "
8494                                         "%d accounting leaf items. Quota "
8495                                         "is out of sync, rescan required.",
8496                                         ret);
8497                         }
8498                 }
8499                 /* make block locked assertion in clean_tree_block happy */
8500                 if (!path->locks[level] &&
8501                     btrfs_header_generation(eb) == trans->transid) {
8502                         btrfs_tree_lock(eb);
8503                         btrfs_set_lock_blocking(eb);
8504                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8505                 }
8506                 clean_tree_block(trans, root->fs_info, eb);
8507         }
8508
8509         if (eb == root->node) {
8510                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8511                         parent = eb->start;
8512                 else
8513                         BUG_ON(root->root_key.objectid !=
8514                                btrfs_header_owner(eb));
8515         } else {
8516                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8517                         parent = path->nodes[level + 1]->start;
8518                 else
8519                         BUG_ON(root->root_key.objectid !=
8520                                btrfs_header_owner(path->nodes[level + 1]));
8521         }
8522
8523         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8524 out:
8525         wc->refs[level] = 0;
8526         wc->flags[level] = 0;
8527         return 0;
8528 }
8529
8530 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8531                                    struct btrfs_root *root,
8532                                    struct btrfs_path *path,
8533                                    struct walk_control *wc)
8534 {
8535         int level = wc->level;
8536         int lookup_info = 1;
8537         int ret;
8538
8539         while (level >= 0) {
8540                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8541                 if (ret > 0)
8542                         break;
8543
8544                 if (level == 0)
8545                         break;
8546
8547                 if (path->slots[level] >=
8548                     btrfs_header_nritems(path->nodes[level]))
8549                         break;
8550
8551                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8552                 if (ret > 0) {
8553                         path->slots[level]++;
8554                         continue;
8555                 } else if (ret < 0)
8556                         return ret;
8557                 level = wc->level;
8558         }
8559         return 0;
8560 }
8561
8562 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8563                                  struct btrfs_root *root,
8564                                  struct btrfs_path *path,
8565                                  struct walk_control *wc, int max_level)
8566 {
8567         int level = wc->level;
8568         int ret;
8569
8570         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8571         while (level < max_level && path->nodes[level]) {
8572                 wc->level = level;
8573                 if (path->slots[level] + 1 <
8574                     btrfs_header_nritems(path->nodes[level])) {
8575                         path->slots[level]++;
8576                         return 0;
8577                 } else {
8578                         ret = walk_up_proc(trans, root, path, wc);
8579                         if (ret > 0)
8580                                 return 0;
8581
8582                         if (path->locks[level]) {
8583                                 btrfs_tree_unlock_rw(path->nodes[level],
8584                                                      path->locks[level]);
8585                                 path->locks[level] = 0;
8586                         }
8587                         free_extent_buffer(path->nodes[level]);
8588                         path->nodes[level] = NULL;
8589                         level++;
8590                 }
8591         }
8592         return 1;
8593 }
8594
8595 /*
8596  * drop a subvolume tree.
8597  *
8598  * this function traverses the tree freeing any blocks that only
8599  * referenced by the tree.
8600  *
8601  * when a shared tree block is found. this function decreases its
8602  * reference count by one. if update_ref is true, this function
8603  * also make sure backrefs for the shared block and all lower level
8604  * blocks are properly updated.
8605  *
8606  * If called with for_reloc == 0, may exit early with -EAGAIN
8607  */
8608 int btrfs_drop_snapshot(struct btrfs_root *root,
8609                          struct btrfs_block_rsv *block_rsv, int update_ref,
8610                          int for_reloc)
8611 {
8612         struct btrfs_path *path;
8613         struct btrfs_trans_handle *trans;
8614         struct btrfs_root *tree_root = root->fs_info->tree_root;
8615         struct btrfs_root_item *root_item = &root->root_item;
8616         struct walk_control *wc;
8617         struct btrfs_key key;
8618         int err = 0;
8619         int ret;
8620         int level;
8621         bool root_dropped = false;
8622
8623         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8624
8625         path = btrfs_alloc_path();
8626         if (!path) {
8627                 err = -ENOMEM;
8628                 goto out;
8629         }
8630
8631         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8632         if (!wc) {
8633                 btrfs_free_path(path);
8634                 err = -ENOMEM;
8635                 goto out;
8636         }
8637
8638         trans = btrfs_start_transaction(tree_root, 0);
8639         if (IS_ERR(trans)) {
8640                 err = PTR_ERR(trans);
8641                 goto out_free;
8642         }
8643
8644         if (block_rsv)
8645                 trans->block_rsv = block_rsv;
8646
8647         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8648                 level = btrfs_header_level(root->node);
8649                 path->nodes[level] = btrfs_lock_root_node(root);
8650                 btrfs_set_lock_blocking(path->nodes[level]);
8651                 path->slots[level] = 0;
8652                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8653                 memset(&wc->update_progress, 0,
8654                        sizeof(wc->update_progress));
8655         } else {
8656                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8657                 memcpy(&wc->update_progress, &key,
8658                        sizeof(wc->update_progress));
8659
8660                 level = root_item->drop_level;
8661                 BUG_ON(level == 0);
8662                 path->lowest_level = level;
8663                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8664                 path->lowest_level = 0;
8665                 if (ret < 0) {
8666                         err = ret;
8667                         goto out_end_trans;
8668                 }
8669                 WARN_ON(ret > 0);
8670
8671                 /*
8672                  * unlock our path, this is safe because only this
8673                  * function is allowed to delete this snapshot
8674                  */
8675                 btrfs_unlock_up_safe(path, 0);
8676
8677                 level = btrfs_header_level(root->node);
8678                 while (1) {
8679                         btrfs_tree_lock(path->nodes[level]);
8680                         btrfs_set_lock_blocking(path->nodes[level]);
8681                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8682
8683                         ret = btrfs_lookup_extent_info(trans, root,
8684                                                 path->nodes[level]->start,
8685                                                 level, 1, &wc->refs[level],
8686                                                 &wc->flags[level]);
8687                         if (ret < 0) {
8688                                 err = ret;
8689                                 goto out_end_trans;
8690                         }
8691                         BUG_ON(wc->refs[level] == 0);
8692
8693                         if (level == root_item->drop_level)
8694                                 break;
8695
8696                         btrfs_tree_unlock(path->nodes[level]);
8697                         path->locks[level] = 0;
8698                         WARN_ON(wc->refs[level] != 1);
8699                         level--;
8700                 }
8701         }
8702
8703         wc->level = level;
8704         wc->shared_level = -1;
8705         wc->stage = DROP_REFERENCE;
8706         wc->update_ref = update_ref;
8707         wc->keep_locks = 0;
8708         wc->for_reloc = for_reloc;
8709         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8710
8711         while (1) {
8712
8713                 ret = walk_down_tree(trans, root, path, wc);
8714                 if (ret < 0) {
8715                         err = ret;
8716                         break;
8717                 }
8718
8719                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8720                 if (ret < 0) {
8721                         err = ret;
8722                         break;
8723                 }
8724
8725                 if (ret > 0) {
8726                         BUG_ON(wc->stage != DROP_REFERENCE);
8727                         break;
8728                 }
8729
8730                 if (wc->stage == DROP_REFERENCE) {
8731                         level = wc->level;
8732                         btrfs_node_key(path->nodes[level],
8733                                        &root_item->drop_progress,
8734                                        path->slots[level]);
8735                         root_item->drop_level = level;
8736                 }
8737
8738                 BUG_ON(wc->level == 0);
8739                 if (btrfs_should_end_transaction(trans, tree_root) ||
8740                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8741                         ret = btrfs_update_root(trans, tree_root,
8742                                                 &root->root_key,
8743                                                 root_item);
8744                         if (ret) {
8745                                 btrfs_abort_transaction(trans, tree_root, ret);
8746                                 err = ret;
8747                                 goto out_end_trans;
8748                         }
8749
8750                         btrfs_end_transaction_throttle(trans, tree_root);
8751                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8752                                 pr_debug("BTRFS: drop snapshot early exit\n");
8753                                 err = -EAGAIN;
8754                                 goto out_free;
8755                         }
8756
8757                         trans = btrfs_start_transaction(tree_root, 0);
8758                         if (IS_ERR(trans)) {
8759                                 err = PTR_ERR(trans);
8760                                 goto out_free;
8761                         }
8762                         if (block_rsv)
8763                                 trans->block_rsv = block_rsv;
8764                 }
8765         }
8766         btrfs_release_path(path);
8767         if (err)
8768                 goto out_end_trans;
8769
8770         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8771         if (ret) {
8772                 btrfs_abort_transaction(trans, tree_root, ret);
8773                 goto out_end_trans;
8774         }
8775
8776         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8777                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8778                                       NULL, NULL);
8779                 if (ret < 0) {
8780                         btrfs_abort_transaction(trans, tree_root, ret);
8781                         err = ret;
8782                         goto out_end_trans;
8783                 } else if (ret > 0) {
8784                         /* if we fail to delete the orphan item this time
8785                          * around, it'll get picked up the next time.
8786                          *
8787                          * The most common failure here is just -ENOENT.
8788                          */
8789                         btrfs_del_orphan_item(trans, tree_root,
8790                                               root->root_key.objectid);
8791                 }
8792         }
8793
8794         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8795                 btrfs_add_dropped_root(trans, root);
8796         } else {
8797                 free_extent_buffer(root->node);
8798                 free_extent_buffer(root->commit_root);
8799                 btrfs_put_fs_root(root);
8800         }
8801         root_dropped = true;
8802 out_end_trans:
8803         btrfs_end_transaction_throttle(trans, tree_root);
8804 out_free:
8805         kfree(wc);
8806         btrfs_free_path(path);
8807 out:
8808         /*
8809          * So if we need to stop dropping the snapshot for whatever reason we
8810          * need to make sure to add it back to the dead root list so that we
8811          * keep trying to do the work later.  This also cleans up roots if we
8812          * don't have it in the radix (like when we recover after a power fail
8813          * or unmount) so we don't leak memory.
8814          */
8815         if (!for_reloc && root_dropped == false)
8816                 btrfs_add_dead_root(root);
8817         if (err && err != -EAGAIN)
8818                 btrfs_std_error(root->fs_info, err, NULL);
8819         return err;
8820 }
8821
8822 /*
8823  * drop subtree rooted at tree block 'node'.
8824  *
8825  * NOTE: this function will unlock and release tree block 'node'
8826  * only used by relocation code
8827  */
8828 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8829                         struct btrfs_root *root,
8830                         struct extent_buffer *node,
8831                         struct extent_buffer *parent)
8832 {
8833         struct btrfs_path *path;
8834         struct walk_control *wc;
8835         int level;
8836         int parent_level;
8837         int ret = 0;
8838         int wret;
8839
8840         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8841
8842         path = btrfs_alloc_path();
8843         if (!path)
8844                 return -ENOMEM;
8845
8846         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8847         if (!wc) {
8848                 btrfs_free_path(path);
8849                 return -ENOMEM;
8850         }
8851
8852         btrfs_assert_tree_locked(parent);
8853         parent_level = btrfs_header_level(parent);
8854         extent_buffer_get(parent);
8855         path->nodes[parent_level] = parent;
8856         path->slots[parent_level] = btrfs_header_nritems(parent);
8857
8858         btrfs_assert_tree_locked(node);
8859         level = btrfs_header_level(node);
8860         path->nodes[level] = node;
8861         path->slots[level] = 0;
8862         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8863
8864         wc->refs[parent_level] = 1;
8865         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8866         wc->level = level;
8867         wc->shared_level = -1;
8868         wc->stage = DROP_REFERENCE;
8869         wc->update_ref = 0;
8870         wc->keep_locks = 1;
8871         wc->for_reloc = 1;
8872         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8873
8874         while (1) {
8875                 wret = walk_down_tree(trans, root, path, wc);
8876                 if (wret < 0) {
8877                         ret = wret;
8878                         break;
8879                 }
8880
8881                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8882                 if (wret < 0)
8883                         ret = wret;
8884                 if (wret != 0)
8885                         break;
8886         }
8887
8888         kfree(wc);
8889         btrfs_free_path(path);
8890         return ret;
8891 }
8892
8893 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8894 {
8895         u64 num_devices;
8896         u64 stripped;
8897
8898         /*
8899          * if restripe for this chunk_type is on pick target profile and
8900          * return, otherwise do the usual balance
8901          */
8902         stripped = get_restripe_target(root->fs_info, flags);
8903         if (stripped)
8904                 return extended_to_chunk(stripped);
8905
8906         num_devices = root->fs_info->fs_devices->rw_devices;
8907
8908         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8909                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8910                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8911
8912         if (num_devices == 1) {
8913                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8914                 stripped = flags & ~stripped;
8915
8916                 /* turn raid0 into single device chunks */
8917                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8918                         return stripped;
8919
8920                 /* turn mirroring into duplication */
8921                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8922                              BTRFS_BLOCK_GROUP_RAID10))
8923                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8924         } else {
8925                 /* they already had raid on here, just return */
8926                 if (flags & stripped)
8927                         return flags;
8928
8929                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8930                 stripped = flags & ~stripped;
8931
8932                 /* switch duplicated blocks with raid1 */
8933                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8934                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8935
8936                 /* this is drive concat, leave it alone */
8937         }
8938
8939         return flags;
8940 }
8941
8942 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8943 {
8944         struct btrfs_space_info *sinfo = cache->space_info;
8945         u64 num_bytes;
8946         u64 min_allocable_bytes;
8947         int ret = -ENOSPC;
8948
8949         /*
8950          * We need some metadata space and system metadata space for
8951          * allocating chunks in some corner cases until we force to set
8952          * it to be readonly.
8953          */
8954         if ((sinfo->flags &
8955              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8956             !force)
8957                 min_allocable_bytes = 1 * 1024 * 1024;
8958         else
8959                 min_allocable_bytes = 0;
8960
8961         spin_lock(&sinfo->lock);
8962         spin_lock(&cache->lock);
8963
8964         if (cache->ro) {
8965                 cache->ro++;
8966                 ret = 0;
8967                 goto out;
8968         }
8969
8970         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8971                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8972
8973         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8974             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8975             min_allocable_bytes <= sinfo->total_bytes) {
8976                 sinfo->bytes_readonly += num_bytes;
8977                 cache->ro++;
8978                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
8979                 ret = 0;
8980         }
8981 out:
8982         spin_unlock(&cache->lock);
8983         spin_unlock(&sinfo->lock);
8984         return ret;
8985 }
8986
8987 int btrfs_inc_block_group_ro(struct btrfs_root *root,
8988                              struct btrfs_block_group_cache *cache)
8989
8990 {
8991         struct btrfs_trans_handle *trans;
8992         u64 alloc_flags;
8993         int ret;
8994
8995 again:
8996         trans = btrfs_join_transaction(root);
8997         if (IS_ERR(trans))
8998                 return PTR_ERR(trans);
8999
9000         /*
9001          * we're not allowed to set block groups readonly after the dirty
9002          * block groups cache has started writing.  If it already started,
9003          * back off and let this transaction commit
9004          */
9005         mutex_lock(&root->fs_info->ro_block_group_mutex);
9006         if (trans->transaction->dirty_bg_run) {
9007                 u64 transid = trans->transid;
9008
9009                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
9010                 btrfs_end_transaction(trans, root);
9011
9012                 ret = btrfs_wait_for_commit(root, transid);
9013                 if (ret)
9014                         return ret;
9015                 goto again;
9016         }
9017
9018         /*
9019          * if we are changing raid levels, try to allocate a corresponding
9020          * block group with the new raid level.
9021          */
9022         alloc_flags = update_block_group_flags(root, cache->flags);
9023         if (alloc_flags != cache->flags) {
9024                 ret = do_chunk_alloc(trans, root, alloc_flags,
9025                                      CHUNK_ALLOC_FORCE);
9026                 /*
9027                  * ENOSPC is allowed here, we may have enough space
9028                  * already allocated at the new raid level to
9029                  * carry on
9030                  */
9031                 if (ret == -ENOSPC)
9032                         ret = 0;
9033                 if (ret < 0)
9034                         goto out;
9035         }
9036
9037         ret = inc_block_group_ro(cache, 0);
9038         if (!ret)
9039                 goto out;
9040         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
9041         ret = do_chunk_alloc(trans, root, alloc_flags,
9042                              CHUNK_ALLOC_FORCE);
9043         if (ret < 0)
9044                 goto out;
9045         ret = inc_block_group_ro(cache, 0);
9046 out:
9047         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9048                 alloc_flags = update_block_group_flags(root, cache->flags);
9049                 lock_chunks(root->fs_info->chunk_root);
9050                 check_system_chunk(trans, root, alloc_flags);
9051                 unlock_chunks(root->fs_info->chunk_root);
9052         }
9053         mutex_unlock(&root->fs_info->ro_block_group_mutex);
9054
9055         btrfs_end_transaction(trans, root);
9056         return ret;
9057 }
9058
9059 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9060                             struct btrfs_root *root, u64 type)
9061 {
9062         u64 alloc_flags = get_alloc_profile(root, type);
9063         return do_chunk_alloc(trans, root, alloc_flags,
9064                               CHUNK_ALLOC_FORCE);
9065 }
9066
9067 /*
9068  * helper to account the unused space of all the readonly block group in the
9069  * space_info. takes mirrors into account.
9070  */
9071 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9072 {
9073         struct btrfs_block_group_cache *block_group;
9074         u64 free_bytes = 0;
9075         int factor;
9076
9077         /* It's df, we don't care if it's racey */
9078         if (list_empty(&sinfo->ro_bgs))
9079                 return 0;
9080
9081         spin_lock(&sinfo->lock);
9082         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9083                 spin_lock(&block_group->lock);
9084
9085                 if (!block_group->ro) {
9086                         spin_unlock(&block_group->lock);
9087                         continue;
9088                 }
9089
9090                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9091                                           BTRFS_BLOCK_GROUP_RAID10 |
9092                                           BTRFS_BLOCK_GROUP_DUP))
9093                         factor = 2;
9094                 else
9095                         factor = 1;
9096
9097                 free_bytes += (block_group->key.offset -
9098                                btrfs_block_group_used(&block_group->item)) *
9099                                factor;
9100
9101                 spin_unlock(&block_group->lock);
9102         }
9103         spin_unlock(&sinfo->lock);
9104
9105         return free_bytes;
9106 }
9107
9108 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9109                               struct btrfs_block_group_cache *cache)
9110 {
9111         struct btrfs_space_info *sinfo = cache->space_info;
9112         u64 num_bytes;
9113
9114         BUG_ON(!cache->ro);
9115
9116         spin_lock(&sinfo->lock);
9117         spin_lock(&cache->lock);
9118         if (!--cache->ro) {
9119                 num_bytes = cache->key.offset - cache->reserved -
9120                             cache->pinned - cache->bytes_super -
9121                             btrfs_block_group_used(&cache->item);
9122                 sinfo->bytes_readonly -= num_bytes;
9123                 list_del_init(&cache->ro_list);
9124         }
9125         spin_unlock(&cache->lock);
9126         spin_unlock(&sinfo->lock);
9127 }
9128
9129 /*
9130  * checks to see if its even possible to relocate this block group.
9131  *
9132  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9133  * ok to go ahead and try.
9134  */
9135 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9136 {
9137         struct btrfs_block_group_cache *block_group;
9138         struct btrfs_space_info *space_info;
9139         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9140         struct btrfs_device *device;
9141         struct btrfs_trans_handle *trans;
9142         u64 min_free;
9143         u64 dev_min = 1;
9144         u64 dev_nr = 0;
9145         u64 target;
9146         int index;
9147         int full = 0;
9148         int ret = 0;
9149
9150         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9151
9152         /* odd, couldn't find the block group, leave it alone */
9153         if (!block_group)
9154                 return -1;
9155
9156         min_free = btrfs_block_group_used(&block_group->item);
9157
9158         /* no bytes used, we're good */
9159         if (!min_free)
9160                 goto out;
9161
9162         space_info = block_group->space_info;
9163         spin_lock(&space_info->lock);
9164
9165         full = space_info->full;
9166
9167         /*
9168          * if this is the last block group we have in this space, we can't
9169          * relocate it unless we're able to allocate a new chunk below.
9170          *
9171          * Otherwise, we need to make sure we have room in the space to handle
9172          * all of the extents from this block group.  If we can, we're good
9173          */
9174         if ((space_info->total_bytes != block_group->key.offset) &&
9175             (space_info->bytes_used + space_info->bytes_reserved +
9176              space_info->bytes_pinned + space_info->bytes_readonly +
9177              min_free < space_info->total_bytes)) {
9178                 spin_unlock(&space_info->lock);
9179                 goto out;
9180         }
9181         spin_unlock(&space_info->lock);
9182
9183         /*
9184          * ok we don't have enough space, but maybe we have free space on our
9185          * devices to allocate new chunks for relocation, so loop through our
9186          * alloc devices and guess if we have enough space.  if this block
9187          * group is going to be restriped, run checks against the target
9188          * profile instead of the current one.
9189          */
9190         ret = -1;
9191
9192         /*
9193          * index:
9194          *      0: raid10
9195          *      1: raid1
9196          *      2: dup
9197          *      3: raid0
9198          *      4: single
9199          */
9200         target = get_restripe_target(root->fs_info, block_group->flags);
9201         if (target) {
9202                 index = __get_raid_index(extended_to_chunk(target));
9203         } else {
9204                 /*
9205                  * this is just a balance, so if we were marked as full
9206                  * we know there is no space for a new chunk
9207                  */
9208                 if (full)
9209                         goto out;
9210
9211                 index = get_block_group_index(block_group);
9212         }
9213
9214         if (index == BTRFS_RAID_RAID10) {
9215                 dev_min = 4;
9216                 /* Divide by 2 */
9217                 min_free >>= 1;
9218         } else if (index == BTRFS_RAID_RAID1) {
9219                 dev_min = 2;
9220         } else if (index == BTRFS_RAID_DUP) {
9221                 /* Multiply by 2 */
9222                 min_free <<= 1;
9223         } else if (index == BTRFS_RAID_RAID0) {
9224                 dev_min = fs_devices->rw_devices;
9225                 min_free = div64_u64(min_free, dev_min);
9226         }
9227
9228         /* We need to do this so that we can look at pending chunks */
9229         trans = btrfs_join_transaction(root);
9230         if (IS_ERR(trans)) {
9231                 ret = PTR_ERR(trans);
9232                 goto out;
9233         }
9234
9235         mutex_lock(&root->fs_info->chunk_mutex);
9236         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9237                 u64 dev_offset;
9238
9239                 /*
9240                  * check to make sure we can actually find a chunk with enough
9241                  * space to fit our block group in.
9242                  */
9243                 if (device->total_bytes > device->bytes_used + min_free &&
9244                     !device->is_tgtdev_for_dev_replace) {
9245                         ret = find_free_dev_extent(trans, device, min_free,
9246                                                    &dev_offset, NULL);
9247                         if (!ret)
9248                                 dev_nr++;
9249
9250                         if (dev_nr >= dev_min)
9251                                 break;
9252
9253                         ret = -1;
9254                 }
9255         }
9256         mutex_unlock(&root->fs_info->chunk_mutex);
9257         btrfs_end_transaction(trans, root);
9258 out:
9259         btrfs_put_block_group(block_group);
9260         return ret;
9261 }
9262
9263 static int find_first_block_group(struct btrfs_root *root,
9264                 struct btrfs_path *path, struct btrfs_key *key)
9265 {
9266         int ret = 0;
9267         struct btrfs_key found_key;
9268         struct extent_buffer *leaf;
9269         int slot;
9270
9271         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9272         if (ret < 0)
9273                 goto out;
9274
9275         while (1) {
9276                 slot = path->slots[0];
9277                 leaf = path->nodes[0];
9278                 if (slot >= btrfs_header_nritems(leaf)) {
9279                         ret = btrfs_next_leaf(root, path);
9280                         if (ret == 0)
9281                                 continue;
9282                         if (ret < 0)
9283                                 goto out;
9284                         break;
9285                 }
9286                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9287
9288                 if (found_key.objectid >= key->objectid &&
9289                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9290                         ret = 0;
9291                         goto out;
9292                 }
9293                 path->slots[0]++;
9294         }
9295 out:
9296         return ret;
9297 }
9298
9299 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9300 {
9301         struct btrfs_block_group_cache *block_group;
9302         u64 last = 0;
9303
9304         while (1) {
9305                 struct inode *inode;
9306
9307                 block_group = btrfs_lookup_first_block_group(info, last);
9308                 while (block_group) {
9309                         spin_lock(&block_group->lock);
9310                         if (block_group->iref)
9311                                 break;
9312                         spin_unlock(&block_group->lock);
9313                         block_group = next_block_group(info->tree_root,
9314                                                        block_group);
9315                 }
9316                 if (!block_group) {
9317                         if (last == 0)
9318                                 break;
9319                         last = 0;
9320                         continue;
9321                 }
9322
9323                 inode = block_group->inode;
9324                 block_group->iref = 0;
9325                 block_group->inode = NULL;
9326                 spin_unlock(&block_group->lock);
9327                 iput(inode);
9328                 last = block_group->key.objectid + block_group->key.offset;
9329                 btrfs_put_block_group(block_group);
9330         }
9331 }
9332
9333 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9334 {
9335         struct btrfs_block_group_cache *block_group;
9336         struct btrfs_space_info *space_info;
9337         struct btrfs_caching_control *caching_ctl;
9338         struct rb_node *n;
9339
9340         down_write(&info->commit_root_sem);
9341         while (!list_empty(&info->caching_block_groups)) {
9342                 caching_ctl = list_entry(info->caching_block_groups.next,
9343                                          struct btrfs_caching_control, list);
9344                 list_del(&caching_ctl->list);
9345                 put_caching_control(caching_ctl);
9346         }
9347         up_write(&info->commit_root_sem);
9348
9349         spin_lock(&info->unused_bgs_lock);
9350         while (!list_empty(&info->unused_bgs)) {
9351                 block_group = list_first_entry(&info->unused_bgs,
9352                                                struct btrfs_block_group_cache,
9353                                                bg_list);
9354                 list_del_init(&block_group->bg_list);
9355                 btrfs_put_block_group(block_group);
9356         }
9357         spin_unlock(&info->unused_bgs_lock);
9358
9359         spin_lock(&info->block_group_cache_lock);
9360         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9361                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9362                                        cache_node);
9363                 rb_erase(&block_group->cache_node,
9364                          &info->block_group_cache_tree);
9365                 RB_CLEAR_NODE(&block_group->cache_node);
9366                 spin_unlock(&info->block_group_cache_lock);
9367
9368                 down_write(&block_group->space_info->groups_sem);
9369                 list_del(&block_group->list);
9370                 up_write(&block_group->space_info->groups_sem);
9371
9372                 if (block_group->cached == BTRFS_CACHE_STARTED)
9373                         wait_block_group_cache_done(block_group);
9374
9375                 /*
9376                  * We haven't cached this block group, which means we could
9377                  * possibly have excluded extents on this block group.
9378                  */
9379                 if (block_group->cached == BTRFS_CACHE_NO ||
9380                     block_group->cached == BTRFS_CACHE_ERROR)
9381                         free_excluded_extents(info->extent_root, block_group);
9382
9383                 btrfs_remove_free_space_cache(block_group);
9384                 btrfs_put_block_group(block_group);
9385
9386                 spin_lock(&info->block_group_cache_lock);
9387         }
9388         spin_unlock(&info->block_group_cache_lock);
9389
9390         /* now that all the block groups are freed, go through and
9391          * free all the space_info structs.  This is only called during
9392          * the final stages of unmount, and so we know nobody is
9393          * using them.  We call synchronize_rcu() once before we start,
9394          * just to be on the safe side.
9395          */
9396         synchronize_rcu();
9397
9398         release_global_block_rsv(info);
9399
9400         while (!list_empty(&info->space_info)) {
9401                 int i;
9402
9403                 space_info = list_entry(info->space_info.next,
9404                                         struct btrfs_space_info,
9405                                         list);
9406                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9407                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9408                             space_info->bytes_reserved > 0 ||
9409                             space_info->bytes_may_use > 0)) {
9410                                 dump_space_info(space_info, 0, 0);
9411                         }
9412                 }
9413                 list_del(&space_info->list);
9414                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9415                         struct kobject *kobj;
9416                         kobj = space_info->block_group_kobjs[i];
9417                         space_info->block_group_kobjs[i] = NULL;
9418                         if (kobj) {
9419                                 kobject_del(kobj);
9420                                 kobject_put(kobj);
9421                         }
9422                 }
9423                 kobject_del(&space_info->kobj);
9424                 kobject_put(&space_info->kobj);
9425         }
9426         return 0;
9427 }
9428
9429 static void __link_block_group(struct btrfs_space_info *space_info,
9430                                struct btrfs_block_group_cache *cache)
9431 {
9432         int index = get_block_group_index(cache);
9433         bool first = false;
9434
9435         down_write(&space_info->groups_sem);
9436         if (list_empty(&space_info->block_groups[index]))
9437                 first = true;
9438         list_add_tail(&cache->list, &space_info->block_groups[index]);
9439         up_write(&space_info->groups_sem);
9440
9441         if (first) {
9442                 struct raid_kobject *rkobj;
9443                 int ret;
9444
9445                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9446                 if (!rkobj)
9447                         goto out_err;
9448                 rkobj->raid_type = index;
9449                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9450                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9451                                   "%s", get_raid_name(index));
9452                 if (ret) {
9453                         kobject_put(&rkobj->kobj);
9454                         goto out_err;
9455                 }
9456                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9457         }
9458
9459         return;
9460 out_err:
9461         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9462 }
9463
9464 static struct btrfs_block_group_cache *
9465 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9466 {
9467         struct btrfs_block_group_cache *cache;
9468
9469         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9470         if (!cache)
9471                 return NULL;
9472
9473         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9474                                         GFP_NOFS);
9475         if (!cache->free_space_ctl) {
9476                 kfree(cache);
9477                 return NULL;
9478         }
9479
9480         cache->key.objectid = start;
9481         cache->key.offset = size;
9482         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9483
9484         cache->sectorsize = root->sectorsize;
9485         cache->fs_info = root->fs_info;
9486         cache->full_stripe_len = btrfs_full_stripe_len(root,
9487                                                &root->fs_info->mapping_tree,
9488                                                start);
9489         atomic_set(&cache->count, 1);
9490         spin_lock_init(&cache->lock);
9491         init_rwsem(&cache->data_rwsem);
9492         INIT_LIST_HEAD(&cache->list);
9493         INIT_LIST_HEAD(&cache->cluster_list);
9494         INIT_LIST_HEAD(&cache->bg_list);
9495         INIT_LIST_HEAD(&cache->ro_list);
9496         INIT_LIST_HEAD(&cache->dirty_list);
9497         INIT_LIST_HEAD(&cache->io_list);
9498         btrfs_init_free_space_ctl(cache);
9499         atomic_set(&cache->trimming, 0);
9500
9501         return cache;
9502 }
9503
9504 int btrfs_read_block_groups(struct btrfs_root *root)
9505 {
9506         struct btrfs_path *path;
9507         int ret;
9508         struct btrfs_block_group_cache *cache;
9509         struct btrfs_fs_info *info = root->fs_info;
9510         struct btrfs_space_info *space_info;
9511         struct btrfs_key key;
9512         struct btrfs_key found_key;
9513         struct extent_buffer *leaf;
9514         int need_clear = 0;
9515         u64 cache_gen;
9516
9517         root = info->extent_root;
9518         key.objectid = 0;
9519         key.offset = 0;
9520         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9521         path = btrfs_alloc_path();
9522         if (!path)
9523                 return -ENOMEM;
9524         path->reada = 1;
9525
9526         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9527         if (btrfs_test_opt(root, SPACE_CACHE) &&
9528             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9529                 need_clear = 1;
9530         if (btrfs_test_opt(root, CLEAR_CACHE))
9531                 need_clear = 1;
9532
9533         while (1) {
9534                 ret = find_first_block_group(root, path, &key);
9535                 if (ret > 0)
9536                         break;
9537                 if (ret != 0)
9538                         goto error;
9539
9540                 leaf = path->nodes[0];
9541                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9542
9543                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9544                                                        found_key.offset);
9545                 if (!cache) {
9546                         ret = -ENOMEM;
9547                         goto error;
9548                 }
9549
9550                 if (need_clear) {
9551                         /*
9552                          * When we mount with old space cache, we need to
9553                          * set BTRFS_DC_CLEAR and set dirty flag.
9554                          *
9555                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9556                          *    truncate the old free space cache inode and
9557                          *    setup a new one.
9558                          * b) Setting 'dirty flag' makes sure that we flush
9559                          *    the new space cache info onto disk.
9560                          */
9561                         if (btrfs_test_opt(root, SPACE_CACHE))
9562                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9563                 }
9564
9565                 read_extent_buffer(leaf, &cache->item,
9566                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9567                                    sizeof(cache->item));
9568                 cache->flags = btrfs_block_group_flags(&cache->item);
9569
9570                 key.objectid = found_key.objectid + found_key.offset;
9571                 btrfs_release_path(path);
9572
9573                 /*
9574                  * We need to exclude the super stripes now so that the space
9575                  * info has super bytes accounted for, otherwise we'll think
9576                  * we have more space than we actually do.
9577                  */
9578                 ret = exclude_super_stripes(root, cache);
9579                 if (ret) {
9580                         /*
9581                          * We may have excluded something, so call this just in
9582                          * case.
9583                          */
9584                         free_excluded_extents(root, cache);
9585                         btrfs_put_block_group(cache);
9586                         goto error;
9587                 }
9588
9589                 /*
9590                  * check for two cases, either we are full, and therefore
9591                  * don't need to bother with the caching work since we won't
9592                  * find any space, or we are empty, and we can just add all
9593                  * the space in and be done with it.  This saves us _alot_ of
9594                  * time, particularly in the full case.
9595                  */
9596                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9597                         cache->last_byte_to_unpin = (u64)-1;
9598                         cache->cached = BTRFS_CACHE_FINISHED;
9599                         free_excluded_extents(root, cache);
9600                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9601                         cache->last_byte_to_unpin = (u64)-1;
9602                         cache->cached = BTRFS_CACHE_FINISHED;
9603                         add_new_free_space(cache, root->fs_info,
9604                                            found_key.objectid,
9605                                            found_key.objectid +
9606                                            found_key.offset);
9607                         free_excluded_extents(root, cache);
9608                 }
9609
9610                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9611                 if (ret) {
9612                         btrfs_remove_free_space_cache(cache);
9613                         btrfs_put_block_group(cache);
9614                         goto error;
9615                 }
9616
9617                 ret = update_space_info(info, cache->flags, found_key.offset,
9618                                         btrfs_block_group_used(&cache->item),
9619                                         &space_info);
9620                 if (ret) {
9621                         btrfs_remove_free_space_cache(cache);
9622                         spin_lock(&info->block_group_cache_lock);
9623                         rb_erase(&cache->cache_node,
9624                                  &info->block_group_cache_tree);
9625                         RB_CLEAR_NODE(&cache->cache_node);
9626                         spin_unlock(&info->block_group_cache_lock);
9627                         btrfs_put_block_group(cache);
9628                         goto error;
9629                 }
9630
9631                 cache->space_info = space_info;
9632                 spin_lock(&cache->space_info->lock);
9633                 cache->space_info->bytes_readonly += cache->bytes_super;
9634                 spin_unlock(&cache->space_info->lock);
9635
9636                 __link_block_group(space_info, cache);
9637
9638                 set_avail_alloc_bits(root->fs_info, cache->flags);
9639                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9640                         inc_block_group_ro(cache, 1);
9641                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9642                         spin_lock(&info->unused_bgs_lock);
9643                         /* Should always be true but just in case. */
9644                         if (list_empty(&cache->bg_list)) {
9645                                 btrfs_get_block_group(cache);
9646                                 list_add_tail(&cache->bg_list,
9647                                               &info->unused_bgs);
9648                         }
9649                         spin_unlock(&info->unused_bgs_lock);
9650                 }
9651         }
9652
9653         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9654                 if (!(get_alloc_profile(root, space_info->flags) &
9655                       (BTRFS_BLOCK_GROUP_RAID10 |
9656                        BTRFS_BLOCK_GROUP_RAID1 |
9657                        BTRFS_BLOCK_GROUP_RAID5 |
9658                        BTRFS_BLOCK_GROUP_RAID6 |
9659                        BTRFS_BLOCK_GROUP_DUP)))
9660                         continue;
9661                 /*
9662                  * avoid allocating from un-mirrored block group if there are
9663                  * mirrored block groups.
9664                  */
9665                 list_for_each_entry(cache,
9666                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9667                                 list)
9668                         inc_block_group_ro(cache, 1);
9669                 list_for_each_entry(cache,
9670                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9671                                 list)
9672                         inc_block_group_ro(cache, 1);
9673         }
9674
9675         init_global_block_rsv(info);
9676         ret = 0;
9677 error:
9678         btrfs_free_path(path);
9679         return ret;
9680 }
9681
9682 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9683                                        struct btrfs_root *root)
9684 {
9685         struct btrfs_block_group_cache *block_group, *tmp;
9686         struct btrfs_root *extent_root = root->fs_info->extent_root;
9687         struct btrfs_block_group_item item;
9688         struct btrfs_key key;
9689         int ret = 0;
9690         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
9691
9692         trans->can_flush_pending_bgs = false;
9693         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9694                 if (ret)
9695                         goto next;
9696
9697                 spin_lock(&block_group->lock);
9698                 memcpy(&item, &block_group->item, sizeof(item));
9699                 memcpy(&key, &block_group->key, sizeof(key));
9700                 spin_unlock(&block_group->lock);
9701
9702                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9703                                         sizeof(item));
9704                 if (ret)
9705                         btrfs_abort_transaction(trans, extent_root, ret);
9706                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9707                                                key.objectid, key.offset);
9708                 if (ret)
9709                         btrfs_abort_transaction(trans, extent_root, ret);
9710 next:
9711                 list_del_init(&block_group->bg_list);
9712         }
9713         trans->can_flush_pending_bgs = can_flush_pending_bgs;
9714 }
9715
9716 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9717                            struct btrfs_root *root, u64 bytes_used,
9718                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9719                            u64 size)
9720 {
9721         int ret;
9722         struct btrfs_root *extent_root;
9723         struct btrfs_block_group_cache *cache;
9724
9725         extent_root = root->fs_info->extent_root;
9726
9727         btrfs_set_log_full_commit(root->fs_info, trans);
9728
9729         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9730         if (!cache)
9731                 return -ENOMEM;
9732
9733         btrfs_set_block_group_used(&cache->item, bytes_used);
9734         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9735         btrfs_set_block_group_flags(&cache->item, type);
9736
9737         cache->flags = type;
9738         cache->last_byte_to_unpin = (u64)-1;
9739         cache->cached = BTRFS_CACHE_FINISHED;
9740         ret = exclude_super_stripes(root, cache);
9741         if (ret) {
9742                 /*
9743                  * We may have excluded something, so call this just in
9744                  * case.
9745                  */
9746                 free_excluded_extents(root, cache);
9747                 btrfs_put_block_group(cache);
9748                 return ret;
9749         }
9750
9751         add_new_free_space(cache, root->fs_info, chunk_offset,
9752                            chunk_offset + size);
9753
9754         free_excluded_extents(root, cache);
9755
9756         /*
9757          * Call to ensure the corresponding space_info object is created and
9758          * assigned to our block group, but don't update its counters just yet.
9759          * We want our bg to be added to the rbtree with its ->space_info set.
9760          */
9761         ret = update_space_info(root->fs_info, cache->flags, 0, 0,
9762                                 &cache->space_info);
9763         if (ret) {
9764                 btrfs_remove_free_space_cache(cache);
9765                 btrfs_put_block_group(cache);
9766                 return ret;
9767         }
9768
9769         ret = btrfs_add_block_group_cache(root->fs_info, cache);
9770         if (ret) {
9771                 btrfs_remove_free_space_cache(cache);
9772                 btrfs_put_block_group(cache);
9773                 return ret;
9774         }
9775
9776         /*
9777          * Now that our block group has its ->space_info set and is inserted in
9778          * the rbtree, update the space info's counters.
9779          */
9780         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
9781                                 &cache->space_info);
9782         if (ret) {
9783                 btrfs_remove_free_space_cache(cache);
9784                 spin_lock(&root->fs_info->block_group_cache_lock);
9785                 rb_erase(&cache->cache_node,
9786                          &root->fs_info->block_group_cache_tree);
9787                 RB_CLEAR_NODE(&cache->cache_node);
9788                 spin_unlock(&root->fs_info->block_group_cache_lock);
9789                 btrfs_put_block_group(cache);
9790                 return ret;
9791         }
9792         update_global_block_rsv(root->fs_info);
9793
9794         spin_lock(&cache->space_info->lock);
9795         cache->space_info->bytes_readonly += cache->bytes_super;
9796         spin_unlock(&cache->space_info->lock);
9797
9798         __link_block_group(cache->space_info, cache);
9799
9800         list_add_tail(&cache->bg_list, &trans->new_bgs);
9801
9802         set_avail_alloc_bits(extent_root->fs_info, type);
9803
9804         return 0;
9805 }
9806
9807 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9808 {
9809         u64 extra_flags = chunk_to_extended(flags) &
9810                                 BTRFS_EXTENDED_PROFILE_MASK;
9811
9812         write_seqlock(&fs_info->profiles_lock);
9813         if (flags & BTRFS_BLOCK_GROUP_DATA)
9814                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9815         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9816                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9817         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9818                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9819         write_sequnlock(&fs_info->profiles_lock);
9820 }
9821
9822 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9823                              struct btrfs_root *root, u64 group_start,
9824                              struct extent_map *em)
9825 {
9826         struct btrfs_path *path;
9827         struct btrfs_block_group_cache *block_group;
9828         struct btrfs_free_cluster *cluster;
9829         struct btrfs_root *tree_root = root->fs_info->tree_root;
9830         struct btrfs_key key;
9831         struct inode *inode;
9832         struct kobject *kobj = NULL;
9833         int ret;
9834         int index;
9835         int factor;
9836         struct btrfs_caching_control *caching_ctl = NULL;
9837         bool remove_em;
9838
9839         root = root->fs_info->extent_root;
9840
9841         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9842         BUG_ON(!block_group);
9843         BUG_ON(!block_group->ro);
9844
9845         /*
9846          * Free the reserved super bytes from this block group before
9847          * remove it.
9848          */
9849         free_excluded_extents(root, block_group);
9850
9851         memcpy(&key, &block_group->key, sizeof(key));
9852         index = get_block_group_index(block_group);
9853         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9854                                   BTRFS_BLOCK_GROUP_RAID1 |
9855                                   BTRFS_BLOCK_GROUP_RAID10))
9856                 factor = 2;
9857         else
9858                 factor = 1;
9859
9860         /* make sure this block group isn't part of an allocation cluster */
9861         cluster = &root->fs_info->data_alloc_cluster;
9862         spin_lock(&cluster->refill_lock);
9863         btrfs_return_cluster_to_free_space(block_group, cluster);
9864         spin_unlock(&cluster->refill_lock);
9865
9866         /*
9867          * make sure this block group isn't part of a metadata
9868          * allocation cluster
9869          */
9870         cluster = &root->fs_info->meta_alloc_cluster;
9871         spin_lock(&cluster->refill_lock);
9872         btrfs_return_cluster_to_free_space(block_group, cluster);
9873         spin_unlock(&cluster->refill_lock);
9874
9875         path = btrfs_alloc_path();
9876         if (!path) {
9877                 ret = -ENOMEM;
9878                 goto out;
9879         }
9880
9881         /*
9882          * get the inode first so any iput calls done for the io_list
9883          * aren't the final iput (no unlinks allowed now)
9884          */
9885         inode = lookup_free_space_inode(tree_root, block_group, path);
9886
9887         mutex_lock(&trans->transaction->cache_write_mutex);
9888         /*
9889          * make sure our free spache cache IO is done before remove the
9890          * free space inode
9891          */
9892         spin_lock(&trans->transaction->dirty_bgs_lock);
9893         if (!list_empty(&block_group->io_list)) {
9894                 list_del_init(&block_group->io_list);
9895
9896                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
9897
9898                 spin_unlock(&trans->transaction->dirty_bgs_lock);
9899                 btrfs_wait_cache_io(root, trans, block_group,
9900                                     &block_group->io_ctl, path,
9901                                     block_group->key.objectid);
9902                 btrfs_put_block_group(block_group);
9903                 spin_lock(&trans->transaction->dirty_bgs_lock);
9904         }
9905
9906         if (!list_empty(&block_group->dirty_list)) {
9907                 list_del_init(&block_group->dirty_list);
9908                 btrfs_put_block_group(block_group);
9909         }
9910         spin_unlock(&trans->transaction->dirty_bgs_lock);
9911         mutex_unlock(&trans->transaction->cache_write_mutex);
9912
9913         if (!IS_ERR(inode)) {
9914                 ret = btrfs_orphan_add(trans, inode);
9915                 if (ret) {
9916                         btrfs_add_delayed_iput(inode);
9917                         goto out;
9918                 }
9919                 clear_nlink(inode);
9920                 /* One for the block groups ref */
9921                 spin_lock(&block_group->lock);
9922                 if (block_group->iref) {
9923                         block_group->iref = 0;
9924                         block_group->inode = NULL;
9925                         spin_unlock(&block_group->lock);
9926                         iput(inode);
9927                 } else {
9928                         spin_unlock(&block_group->lock);
9929                 }
9930                 /* One for our lookup ref */
9931                 btrfs_add_delayed_iput(inode);
9932         }
9933
9934         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9935         key.offset = block_group->key.objectid;
9936         key.type = 0;
9937
9938         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9939         if (ret < 0)
9940                 goto out;
9941         if (ret > 0)
9942                 btrfs_release_path(path);
9943         if (ret == 0) {
9944                 ret = btrfs_del_item(trans, tree_root, path);
9945                 if (ret)
9946                         goto out;
9947                 btrfs_release_path(path);
9948         }
9949
9950         spin_lock(&root->fs_info->block_group_cache_lock);
9951         rb_erase(&block_group->cache_node,
9952                  &root->fs_info->block_group_cache_tree);
9953         RB_CLEAR_NODE(&block_group->cache_node);
9954
9955         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9956                 root->fs_info->first_logical_byte = (u64)-1;
9957         spin_unlock(&root->fs_info->block_group_cache_lock);
9958
9959         down_write(&block_group->space_info->groups_sem);
9960         /*
9961          * we must use list_del_init so people can check to see if they
9962          * are still on the list after taking the semaphore
9963          */
9964         list_del_init(&block_group->list);
9965         if (list_empty(&block_group->space_info->block_groups[index])) {
9966                 kobj = block_group->space_info->block_group_kobjs[index];
9967                 block_group->space_info->block_group_kobjs[index] = NULL;
9968                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9969         }
9970         up_write(&block_group->space_info->groups_sem);
9971         if (kobj) {
9972                 kobject_del(kobj);
9973                 kobject_put(kobj);
9974         }
9975
9976         if (block_group->has_caching_ctl)
9977                 caching_ctl = get_caching_control(block_group);
9978         if (block_group->cached == BTRFS_CACHE_STARTED)
9979                 wait_block_group_cache_done(block_group);
9980         if (block_group->has_caching_ctl) {
9981                 down_write(&root->fs_info->commit_root_sem);
9982                 if (!caching_ctl) {
9983                         struct btrfs_caching_control *ctl;
9984
9985                         list_for_each_entry(ctl,
9986                                     &root->fs_info->caching_block_groups, list)
9987                                 if (ctl->block_group == block_group) {
9988                                         caching_ctl = ctl;
9989                                         atomic_inc(&caching_ctl->count);
9990                                         break;
9991                                 }
9992                 }
9993                 if (caching_ctl)
9994                         list_del_init(&caching_ctl->list);
9995                 up_write(&root->fs_info->commit_root_sem);
9996                 if (caching_ctl) {
9997                         /* Once for the caching bgs list and once for us. */
9998                         put_caching_control(caching_ctl);
9999                         put_caching_control(caching_ctl);
10000                 }
10001         }
10002
10003         spin_lock(&trans->transaction->dirty_bgs_lock);
10004         if (!list_empty(&block_group->dirty_list)) {
10005                 WARN_ON(1);
10006         }
10007         if (!list_empty(&block_group->io_list)) {
10008                 WARN_ON(1);
10009         }
10010         spin_unlock(&trans->transaction->dirty_bgs_lock);
10011         btrfs_remove_free_space_cache(block_group);
10012
10013         spin_lock(&block_group->space_info->lock);
10014         list_del_init(&block_group->ro_list);
10015
10016         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
10017                 WARN_ON(block_group->space_info->total_bytes
10018                         < block_group->key.offset);
10019                 WARN_ON(block_group->space_info->bytes_readonly
10020                         < block_group->key.offset);
10021                 WARN_ON(block_group->space_info->disk_total
10022                         < block_group->key.offset * factor);
10023         }
10024         block_group->space_info->total_bytes -= block_group->key.offset;
10025         block_group->space_info->bytes_readonly -= block_group->key.offset;
10026         block_group->space_info->disk_total -= block_group->key.offset * factor;
10027
10028         spin_unlock(&block_group->space_info->lock);
10029
10030         memcpy(&key, &block_group->key, sizeof(key));
10031
10032         lock_chunks(root);
10033         if (!list_empty(&em->list)) {
10034                 /* We're in the transaction->pending_chunks list. */
10035                 free_extent_map(em);
10036         }
10037         spin_lock(&block_group->lock);
10038         block_group->removed = 1;
10039         /*
10040          * At this point trimming can't start on this block group, because we
10041          * removed the block group from the tree fs_info->block_group_cache_tree
10042          * so no one can't find it anymore and even if someone already got this
10043          * block group before we removed it from the rbtree, they have already
10044          * incremented block_group->trimming - if they didn't, they won't find
10045          * any free space entries because we already removed them all when we
10046          * called btrfs_remove_free_space_cache().
10047          *
10048          * And we must not remove the extent map from the fs_info->mapping_tree
10049          * to prevent the same logical address range and physical device space
10050          * ranges from being reused for a new block group. This is because our
10051          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10052          * completely transactionless, so while it is trimming a range the
10053          * currently running transaction might finish and a new one start,
10054          * allowing for new block groups to be created that can reuse the same
10055          * physical device locations unless we take this special care.
10056          *
10057          * There may also be an implicit trim operation if the file system
10058          * is mounted with -odiscard. The same protections must remain
10059          * in place until the extents have been discarded completely when
10060          * the transaction commit has completed.
10061          */
10062         remove_em = (atomic_read(&block_group->trimming) == 0);
10063         /*
10064          * Make sure a trimmer task always sees the em in the pinned_chunks list
10065          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10066          * before checking block_group->removed).
10067          */
10068         if (!remove_em) {
10069                 /*
10070                  * Our em might be in trans->transaction->pending_chunks which
10071                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10072                  * and so is the fs_info->pinned_chunks list.
10073                  *
10074                  * So at this point we must be holding the chunk_mutex to avoid
10075                  * any races with chunk allocation (more specifically at
10076                  * volumes.c:contains_pending_extent()), to ensure it always
10077                  * sees the em, either in the pending_chunks list or in the
10078                  * pinned_chunks list.
10079                  */
10080                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10081         }
10082         spin_unlock(&block_group->lock);
10083
10084         if (remove_em) {
10085                 struct extent_map_tree *em_tree;
10086
10087                 em_tree = &root->fs_info->mapping_tree.map_tree;
10088                 write_lock(&em_tree->lock);
10089                 /*
10090                  * The em might be in the pending_chunks list, so make sure the
10091                  * chunk mutex is locked, since remove_extent_mapping() will
10092                  * delete us from that list.
10093                  */
10094                 remove_extent_mapping(em_tree, em);
10095                 write_unlock(&em_tree->lock);
10096                 /* once for the tree */
10097                 free_extent_map(em);
10098         }
10099
10100         unlock_chunks(root);
10101
10102         btrfs_put_block_group(block_group);
10103         btrfs_put_block_group(block_group);
10104
10105         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10106         if (ret > 0)
10107                 ret = -EIO;
10108         if (ret < 0)
10109                 goto out;
10110
10111         ret = btrfs_del_item(trans, root, path);
10112 out:
10113         btrfs_free_path(path);
10114         return ret;
10115 }
10116
10117 /*
10118  * Process the unused_bgs list and remove any that don't have any allocated
10119  * space inside of them.
10120  */
10121 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10122 {
10123         struct btrfs_block_group_cache *block_group;
10124         struct btrfs_space_info *space_info;
10125         struct btrfs_root *root = fs_info->extent_root;
10126         struct btrfs_trans_handle *trans;
10127         int ret = 0;
10128
10129         if (!fs_info->open)
10130                 return;
10131
10132         spin_lock(&fs_info->unused_bgs_lock);
10133         while (!list_empty(&fs_info->unused_bgs)) {
10134                 u64 start, end;
10135                 int trimming;
10136
10137                 block_group = list_first_entry(&fs_info->unused_bgs,
10138                                                struct btrfs_block_group_cache,
10139                                                bg_list);
10140                 space_info = block_group->space_info;
10141                 list_del_init(&block_group->bg_list);
10142                 if (ret || btrfs_mixed_space_info(space_info)) {
10143                         btrfs_put_block_group(block_group);
10144                         continue;
10145                 }
10146                 spin_unlock(&fs_info->unused_bgs_lock);
10147
10148                 mutex_lock(&root->fs_info->delete_unused_bgs_mutex);
10149
10150                 /* Don't want to race with allocators so take the groups_sem */
10151                 down_write(&space_info->groups_sem);
10152                 spin_lock(&block_group->lock);
10153                 if (block_group->reserved ||
10154                     btrfs_block_group_used(&block_group->item) ||
10155                     block_group->ro) {
10156                         /*
10157                          * We want to bail if we made new allocations or have
10158                          * outstanding allocations in this block group.  We do
10159                          * the ro check in case balance is currently acting on
10160                          * this block group.
10161                          */
10162                         spin_unlock(&block_group->lock);
10163                         up_write(&space_info->groups_sem);
10164                         goto next;
10165                 }
10166                 spin_unlock(&block_group->lock);
10167
10168                 /* We don't want to force the issue, only flip if it's ok. */
10169                 ret = inc_block_group_ro(block_group, 0);
10170                 up_write(&space_info->groups_sem);
10171                 if (ret < 0) {
10172                         ret = 0;
10173                         goto next;
10174                 }
10175
10176                 /*
10177                  * Want to do this before we do anything else so we can recover
10178                  * properly if we fail to join the transaction.
10179                  */
10180                 /* 1 for btrfs_orphan_reserve_metadata() */
10181                 trans = btrfs_start_transaction(root, 1);
10182                 if (IS_ERR(trans)) {
10183                         btrfs_dec_block_group_ro(root, block_group);
10184                         ret = PTR_ERR(trans);
10185                         goto next;
10186                 }
10187
10188                 /*
10189                  * We could have pending pinned extents for this block group,
10190                  * just delete them, we don't care about them anymore.
10191                  */
10192                 start = block_group->key.objectid;
10193                 end = start + block_group->key.offset - 1;
10194                 /*
10195                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10196                  * btrfs_finish_extent_commit(). If we are at transaction N,
10197                  * another task might be running finish_extent_commit() for the
10198                  * previous transaction N - 1, and have seen a range belonging
10199                  * to the block group in freed_extents[] before we were able to
10200                  * clear the whole block group range from freed_extents[]. This
10201                  * means that task can lookup for the block group after we
10202                  * unpinned it from freed_extents[] and removed it, leading to
10203                  * a BUG_ON() at btrfs_unpin_extent_range().
10204                  */
10205                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10206                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10207                                   EXTENT_DIRTY, GFP_NOFS);
10208                 if (ret) {
10209                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10210                         btrfs_dec_block_group_ro(root, block_group);
10211                         goto end_trans;
10212                 }
10213                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10214                                   EXTENT_DIRTY, GFP_NOFS);
10215                 if (ret) {
10216                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10217                         btrfs_dec_block_group_ro(root, block_group);
10218                         goto end_trans;
10219                 }
10220                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10221
10222                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10223                 spin_lock(&space_info->lock);
10224                 spin_lock(&block_group->lock);
10225
10226                 space_info->bytes_pinned -= block_group->pinned;
10227                 space_info->bytes_readonly += block_group->pinned;
10228                 percpu_counter_add(&space_info->total_bytes_pinned,
10229                                    -block_group->pinned);
10230                 block_group->pinned = 0;
10231
10232                 spin_unlock(&block_group->lock);
10233                 spin_unlock(&space_info->lock);
10234
10235                 /* DISCARD can flip during remount */
10236                 trimming = btrfs_test_opt(root, DISCARD);
10237
10238                 /* Implicit trim during transaction commit. */
10239                 if (trimming)
10240                         btrfs_get_block_group_trimming(block_group);
10241
10242                 /*
10243                  * Btrfs_remove_chunk will abort the transaction if things go
10244                  * horribly wrong.
10245                  */
10246                 ret = btrfs_remove_chunk(trans, root,
10247                                          block_group->key.objectid);
10248
10249                 if (ret) {
10250                         if (trimming)
10251                                 btrfs_put_block_group_trimming(block_group);
10252                         goto end_trans;
10253                 }
10254
10255                 /*
10256                  * If we're not mounted with -odiscard, we can just forget
10257                  * about this block group. Otherwise we'll need to wait
10258                  * until transaction commit to do the actual discard.
10259                  */
10260                 if (trimming) {
10261                         WARN_ON(!list_empty(&block_group->bg_list));
10262                         spin_lock(&trans->transaction->deleted_bgs_lock);
10263                         list_move(&block_group->bg_list,
10264                                   &trans->transaction->deleted_bgs);
10265                         spin_unlock(&trans->transaction->deleted_bgs_lock);
10266                         btrfs_get_block_group(block_group);
10267                 }
10268 end_trans:
10269                 btrfs_end_transaction(trans, root);
10270 next:
10271                 mutex_unlock(&root->fs_info->delete_unused_bgs_mutex);
10272                 btrfs_put_block_group(block_group);
10273                 spin_lock(&fs_info->unused_bgs_lock);
10274         }
10275         spin_unlock(&fs_info->unused_bgs_lock);
10276 }
10277
10278 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10279 {
10280         struct btrfs_space_info *space_info;
10281         struct btrfs_super_block *disk_super;
10282         u64 features;
10283         u64 flags;
10284         int mixed = 0;
10285         int ret;
10286
10287         disk_super = fs_info->super_copy;
10288         if (!btrfs_super_root(disk_super))
10289                 return 1;
10290
10291         features = btrfs_super_incompat_flags(disk_super);
10292         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10293                 mixed = 1;
10294
10295         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10296         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10297         if (ret)
10298                 goto out;
10299
10300         if (mixed) {
10301                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10302                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10303         } else {
10304                 flags = BTRFS_BLOCK_GROUP_METADATA;
10305                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10306                 if (ret)
10307                         goto out;
10308
10309                 flags = BTRFS_BLOCK_GROUP_DATA;
10310                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10311         }
10312 out:
10313         return ret;
10314 }
10315
10316 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10317 {
10318         return unpin_extent_range(root, start, end, false);
10319 }
10320
10321 /*
10322  * It used to be that old block groups would be left around forever.
10323  * Iterating over them would be enough to trim unused space.  Since we
10324  * now automatically remove them, we also need to iterate over unallocated
10325  * space.
10326  *
10327  * We don't want a transaction for this since the discard may take a
10328  * substantial amount of time.  We don't require that a transaction be
10329  * running, but we do need to take a running transaction into account
10330  * to ensure that we're not discarding chunks that were released in
10331  * the current transaction.
10332  *
10333  * Holding the chunks lock will prevent other threads from allocating
10334  * or releasing chunks, but it won't prevent a running transaction
10335  * from committing and releasing the memory that the pending chunks
10336  * list head uses.  For that, we need to take a reference to the
10337  * transaction.
10338  */
10339 static int btrfs_trim_free_extents(struct btrfs_device *device,
10340                                    u64 minlen, u64 *trimmed)
10341 {
10342         u64 start = 0, len = 0;
10343         int ret;
10344
10345         *trimmed = 0;
10346
10347         /* Not writeable = nothing to do. */
10348         if (!device->writeable)
10349                 return 0;
10350
10351         /* No free space = nothing to do. */
10352         if (device->total_bytes <= device->bytes_used)
10353                 return 0;
10354
10355         ret = 0;
10356
10357         while (1) {
10358                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
10359                 struct btrfs_transaction *trans;
10360                 u64 bytes;
10361
10362                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10363                 if (ret)
10364                         return ret;
10365
10366                 down_read(&fs_info->commit_root_sem);
10367
10368                 spin_lock(&fs_info->trans_lock);
10369                 trans = fs_info->running_transaction;
10370                 if (trans)
10371                         atomic_inc(&trans->use_count);
10372                 spin_unlock(&fs_info->trans_lock);
10373
10374                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10375                                                  &start, &len);
10376                 if (trans)
10377                         btrfs_put_transaction(trans);
10378
10379                 if (ret) {
10380                         up_read(&fs_info->commit_root_sem);
10381                         mutex_unlock(&fs_info->chunk_mutex);
10382                         if (ret == -ENOSPC)
10383                                 ret = 0;
10384                         break;
10385                 }
10386
10387                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10388                 up_read(&fs_info->commit_root_sem);
10389                 mutex_unlock(&fs_info->chunk_mutex);
10390
10391                 if (ret)
10392                         break;
10393
10394                 start += len;
10395                 *trimmed += bytes;
10396
10397                 if (fatal_signal_pending(current)) {
10398                         ret = -ERESTARTSYS;
10399                         break;
10400                 }
10401
10402                 cond_resched();
10403         }
10404
10405         return ret;
10406 }
10407
10408 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10409 {
10410         struct btrfs_fs_info *fs_info = root->fs_info;
10411         struct btrfs_block_group_cache *cache = NULL;
10412         struct btrfs_device *device;
10413         struct list_head *devices;
10414         u64 group_trimmed;
10415         u64 start;
10416         u64 end;
10417         u64 trimmed = 0;
10418         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10419         int ret = 0;
10420
10421         /*
10422          * try to trim all FS space, our block group may start from non-zero.
10423          */
10424         if (range->len == total_bytes)
10425                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10426         else
10427                 cache = btrfs_lookup_block_group(fs_info, range->start);
10428
10429         while (cache) {
10430                 if (cache->key.objectid >= (range->start + range->len)) {
10431                         btrfs_put_block_group(cache);
10432                         break;
10433                 }
10434
10435                 start = max(range->start, cache->key.objectid);
10436                 end = min(range->start + range->len,
10437                                 cache->key.objectid + cache->key.offset);
10438
10439                 if (end - start >= range->minlen) {
10440                         if (!block_group_cache_done(cache)) {
10441                                 ret = cache_block_group(cache, 0);
10442                                 if (ret) {
10443                                         btrfs_put_block_group(cache);
10444                                         break;
10445                                 }
10446                                 ret = wait_block_group_cache_done(cache);
10447                                 if (ret) {
10448                                         btrfs_put_block_group(cache);
10449                                         break;
10450                                 }
10451                         }
10452                         ret = btrfs_trim_block_group(cache,
10453                                                      &group_trimmed,
10454                                                      start,
10455                                                      end,
10456                                                      range->minlen);
10457
10458                         trimmed += group_trimmed;
10459                         if (ret) {
10460                                 btrfs_put_block_group(cache);
10461                                 break;
10462                         }
10463                 }
10464
10465                 cache = next_block_group(fs_info->tree_root, cache);
10466         }
10467
10468         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
10469         devices = &root->fs_info->fs_devices->alloc_list;
10470         list_for_each_entry(device, devices, dev_alloc_list) {
10471                 ret = btrfs_trim_free_extents(device, range->minlen,
10472                                               &group_trimmed);
10473                 if (ret)
10474                         break;
10475
10476                 trimmed += group_trimmed;
10477         }
10478         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
10479
10480         range->len = trimmed;
10481         return ret;
10482 }
10483
10484 /*
10485  * btrfs_{start,end}_write_no_snapshoting() are similar to
10486  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10487  * data into the page cache through nocow before the subvolume is snapshoted,
10488  * but flush the data into disk after the snapshot creation, or to prevent
10489  * operations while snapshoting is ongoing and that cause the snapshot to be
10490  * inconsistent (writes followed by expanding truncates for example).
10491  */
10492 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10493 {
10494         percpu_counter_dec(&root->subv_writers->counter);
10495         /*
10496          * Make sure counter is updated before we wake up waiters.
10497          */
10498         smp_mb();
10499         if (waitqueue_active(&root->subv_writers->wait))
10500                 wake_up(&root->subv_writers->wait);
10501 }
10502
10503 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10504 {
10505         if (atomic_read(&root->will_be_snapshoted))
10506                 return 0;
10507
10508         percpu_counter_inc(&root->subv_writers->counter);
10509         /*
10510          * Make sure counter is updated before we check for snapshot creation.
10511          */
10512         smp_mb();
10513         if (atomic_read(&root->will_be_snapshoted)) {
10514                 btrfs_end_write_no_snapshoting(root);
10515                 return 0;
10516         }
10517         return 1;
10518 }