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, num_pages, 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, 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 /*
3908  * This will check the space that the inode allocates from to make sure we have
3909  * enough space for bytes.
3910  */
3911 int btrfs_check_data_free_space(struct inode *inode, u64 bytes, u64 write_bytes)
3912 {
3913         struct btrfs_space_info *data_sinfo;
3914         struct btrfs_root *root = BTRFS_I(inode)->root;
3915         struct btrfs_fs_info *fs_info = root->fs_info;
3916         u64 used;
3917         int ret = 0;
3918         int need_commit = 2;
3919         int have_pinned_space;
3920
3921         /* make sure bytes are sectorsize aligned */
3922         bytes = ALIGN(bytes, root->sectorsize);
3923
3924         if (btrfs_is_free_space_inode(inode)) {
3925                 need_commit = 0;
3926                 ASSERT(current->journal_info);
3927         }
3928
3929         data_sinfo = fs_info->data_sinfo;
3930         if (!data_sinfo)
3931                 goto alloc;
3932
3933 again:
3934         /* make sure we have enough space to handle the data first */
3935         spin_lock(&data_sinfo->lock);
3936         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3937                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3938                 data_sinfo->bytes_may_use;
3939
3940         if (used + bytes > data_sinfo->total_bytes) {
3941                 struct btrfs_trans_handle *trans;
3942
3943                 /*
3944                  * if we don't have enough free bytes in this space then we need
3945                  * to alloc a new chunk.
3946                  */
3947                 if (!data_sinfo->full) {
3948                         u64 alloc_target;
3949
3950                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3951                         spin_unlock(&data_sinfo->lock);
3952 alloc:
3953                         alloc_target = btrfs_get_alloc_profile(root, 1);
3954                         /*
3955                          * It is ugly that we don't call nolock join
3956                          * transaction for the free space inode case here.
3957                          * But it is safe because we only do the data space
3958                          * reservation for the free space cache in the
3959                          * transaction context, the common join transaction
3960                          * just increase the counter of the current transaction
3961                          * handler, doesn't try to acquire the trans_lock of
3962                          * the fs.
3963                          */
3964                         trans = btrfs_join_transaction(root);
3965                         if (IS_ERR(trans))
3966                                 return PTR_ERR(trans);
3967
3968                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3969                                              alloc_target,
3970                                              CHUNK_ALLOC_NO_FORCE);
3971                         btrfs_end_transaction(trans, root);
3972                         if (ret < 0) {
3973                                 if (ret != -ENOSPC)
3974                                         return ret;
3975                                 else {
3976                                         have_pinned_space = 1;
3977                                         goto commit_trans;
3978                                 }
3979                         }
3980
3981                         if (!data_sinfo)
3982                                 data_sinfo = fs_info->data_sinfo;
3983
3984                         goto again;
3985                 }
3986
3987                 /*
3988                  * If we don't have enough pinned space to deal with this
3989                  * allocation, and no removed chunk in current transaction,
3990                  * don't bother committing the transaction.
3991                  */
3992                 have_pinned_space = percpu_counter_compare(
3993                         &data_sinfo->total_bytes_pinned,
3994                         used + bytes - data_sinfo->total_bytes);
3995                 spin_unlock(&data_sinfo->lock);
3996
3997                 /* commit the current transaction and try again */
3998 commit_trans:
3999                 if (need_commit &&
4000                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
4001                         need_commit--;
4002
4003                         if (need_commit > 0)
4004                                 btrfs_wait_ordered_roots(fs_info, -1);
4005
4006                         trans = btrfs_join_transaction(root);
4007                         if (IS_ERR(trans))
4008                                 return PTR_ERR(trans);
4009                         if (have_pinned_space >= 0 ||
4010                             trans->transaction->have_free_bgs ||
4011                             need_commit > 0) {
4012                                 ret = btrfs_commit_transaction(trans, root);
4013                                 if (ret)
4014                                         return ret;
4015                                 /*
4016                                  * make sure that all running delayed iput are
4017                                  * done
4018                                  */
4019                                 down_write(&root->fs_info->delayed_iput_sem);
4020                                 up_write(&root->fs_info->delayed_iput_sem);
4021                                 goto again;
4022                         } else {
4023                                 btrfs_end_transaction(trans, root);
4024                         }
4025                 }
4026
4027                 trace_btrfs_space_reservation(root->fs_info,
4028                                               "space_info:enospc",
4029                                               data_sinfo->flags, bytes, 1);
4030                 return -ENOSPC;
4031         }
4032         ret = btrfs_qgroup_reserve(root, write_bytes);
4033         if (ret)
4034                 goto out;
4035         data_sinfo->bytes_may_use += bytes;
4036         trace_btrfs_space_reservation(root->fs_info, "space_info",
4037                                       data_sinfo->flags, bytes, 1);
4038 out:
4039         spin_unlock(&data_sinfo->lock);
4040
4041         return ret;
4042 }
4043
4044 /*
4045  * Called if we need to clear a data reservation for this inode.
4046  */
4047 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
4048 {
4049         struct btrfs_root *root = BTRFS_I(inode)->root;
4050         struct btrfs_space_info *data_sinfo;
4051
4052         /* make sure bytes are sectorsize aligned */
4053         bytes = ALIGN(bytes, root->sectorsize);
4054
4055         data_sinfo = root->fs_info->data_sinfo;
4056         spin_lock(&data_sinfo->lock);
4057         WARN_ON(data_sinfo->bytes_may_use < bytes);
4058         data_sinfo->bytes_may_use -= bytes;
4059         trace_btrfs_space_reservation(root->fs_info, "space_info",
4060                                       data_sinfo->flags, bytes, 0);
4061         spin_unlock(&data_sinfo->lock);
4062 }
4063
4064 static void force_metadata_allocation(struct btrfs_fs_info *info)
4065 {
4066         struct list_head *head = &info->space_info;
4067         struct btrfs_space_info *found;
4068
4069         rcu_read_lock();
4070         list_for_each_entry_rcu(found, head, list) {
4071                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4072                         found->force_alloc = CHUNK_ALLOC_FORCE;
4073         }
4074         rcu_read_unlock();
4075 }
4076
4077 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4078 {
4079         return (global->size << 1);
4080 }
4081
4082 static int should_alloc_chunk(struct btrfs_root *root,
4083                               struct btrfs_space_info *sinfo, int force)
4084 {
4085         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4086         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4087         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4088         u64 thresh;
4089
4090         if (force == CHUNK_ALLOC_FORCE)
4091                 return 1;
4092
4093         /*
4094          * We need to take into account the global rsv because for all intents
4095          * and purposes it's used space.  Don't worry about locking the
4096          * global_rsv, it doesn't change except when the transaction commits.
4097          */
4098         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4099                 num_allocated += calc_global_rsv_need_space(global_rsv);
4100
4101         /*
4102          * in limited mode, we want to have some free space up to
4103          * about 1% of the FS size.
4104          */
4105         if (force == CHUNK_ALLOC_LIMITED) {
4106                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4107                 thresh = max_t(u64, 64 * 1024 * 1024,
4108                                div_factor_fine(thresh, 1));
4109
4110                 if (num_bytes - num_allocated < thresh)
4111                         return 1;
4112         }
4113
4114         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
4115                 return 0;
4116         return 1;
4117 }
4118
4119 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4120 {
4121         u64 num_dev;
4122
4123         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4124                     BTRFS_BLOCK_GROUP_RAID0 |
4125                     BTRFS_BLOCK_GROUP_RAID5 |
4126                     BTRFS_BLOCK_GROUP_RAID6))
4127                 num_dev = root->fs_info->fs_devices->rw_devices;
4128         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4129                 num_dev = 2;
4130         else
4131                 num_dev = 1;    /* DUP or single */
4132
4133         return num_dev;
4134 }
4135
4136 /*
4137  * If @is_allocation is true, reserve space in the system space info necessary
4138  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4139  * removing a chunk.
4140  */
4141 void check_system_chunk(struct btrfs_trans_handle *trans,
4142                         struct btrfs_root *root,
4143                         u64 type)
4144 {
4145         struct btrfs_space_info *info;
4146         u64 left;
4147         u64 thresh;
4148         int ret = 0;
4149         u64 num_devs;
4150
4151         /*
4152          * Needed because we can end up allocating a system chunk and for an
4153          * atomic and race free space reservation in the chunk block reserve.
4154          */
4155         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4156
4157         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4158         spin_lock(&info->lock);
4159         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4160                 info->bytes_reserved - info->bytes_readonly -
4161                 info->bytes_may_use;
4162         spin_unlock(&info->lock);
4163
4164         num_devs = get_profile_num_devs(root, type);
4165
4166         /* num_devs device items to update and 1 chunk item to add or remove */
4167         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4168                 btrfs_calc_trans_metadata_size(root, 1);
4169
4170         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4171                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4172                         left, thresh, type);
4173                 dump_space_info(info, 0, 0);
4174         }
4175
4176         if (left < thresh) {
4177                 u64 flags;
4178
4179                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4180                 /*
4181                  * Ignore failure to create system chunk. We might end up not
4182                  * needing it, as we might not need to COW all nodes/leafs from
4183                  * the paths we visit in the chunk tree (they were already COWed
4184                  * or created in the current transaction for example).
4185                  */
4186                 ret = btrfs_alloc_chunk(trans, root, flags);
4187         }
4188
4189         if (!ret) {
4190                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4191                                           &root->fs_info->chunk_block_rsv,
4192                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4193                 if (!ret)
4194                         trans->chunk_bytes_reserved += thresh;
4195         }
4196 }
4197
4198 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4199                           struct btrfs_root *extent_root, u64 flags, int force)
4200 {
4201         struct btrfs_space_info *space_info;
4202         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4203         int wait_for_alloc = 0;
4204         int ret = 0;
4205
4206         /* Don't re-enter if we're already allocating a chunk */
4207         if (trans->allocating_chunk)
4208                 return -ENOSPC;
4209
4210         space_info = __find_space_info(extent_root->fs_info, flags);
4211         if (!space_info) {
4212                 ret = update_space_info(extent_root->fs_info, flags,
4213                                         0, 0, &space_info);
4214                 BUG_ON(ret); /* -ENOMEM */
4215         }
4216         BUG_ON(!space_info); /* Logic error */
4217
4218 again:
4219         spin_lock(&space_info->lock);
4220         if (force < space_info->force_alloc)
4221                 force = space_info->force_alloc;
4222         if (space_info->full) {
4223                 if (should_alloc_chunk(extent_root, space_info, force))
4224                         ret = -ENOSPC;
4225                 else
4226                         ret = 0;
4227                 spin_unlock(&space_info->lock);
4228                 return ret;
4229         }
4230
4231         if (!should_alloc_chunk(extent_root, space_info, force)) {
4232                 spin_unlock(&space_info->lock);
4233                 return 0;
4234         } else if (space_info->chunk_alloc) {
4235                 wait_for_alloc = 1;
4236         } else {
4237                 space_info->chunk_alloc = 1;
4238         }
4239
4240         spin_unlock(&space_info->lock);
4241
4242         mutex_lock(&fs_info->chunk_mutex);
4243
4244         /*
4245          * The chunk_mutex is held throughout the entirety of a chunk
4246          * allocation, so once we've acquired the chunk_mutex we know that the
4247          * other guy is done and we need to recheck and see if we should
4248          * allocate.
4249          */
4250         if (wait_for_alloc) {
4251                 mutex_unlock(&fs_info->chunk_mutex);
4252                 wait_for_alloc = 0;
4253                 goto again;
4254         }
4255
4256         trans->allocating_chunk = true;
4257
4258         /*
4259          * If we have mixed data/metadata chunks we want to make sure we keep
4260          * allocating mixed chunks instead of individual chunks.
4261          */
4262         if (btrfs_mixed_space_info(space_info))
4263                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4264
4265         /*
4266          * if we're doing a data chunk, go ahead and make sure that
4267          * we keep a reasonable number of metadata chunks allocated in the
4268          * FS as well.
4269          */
4270         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4271                 fs_info->data_chunk_allocations++;
4272                 if (!(fs_info->data_chunk_allocations %
4273                       fs_info->metadata_ratio))
4274                         force_metadata_allocation(fs_info);
4275         }
4276
4277         /*
4278          * Check if we have enough space in SYSTEM chunk because we may need
4279          * to update devices.
4280          */
4281         check_system_chunk(trans, extent_root, flags);
4282
4283         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4284         trans->allocating_chunk = false;
4285
4286         spin_lock(&space_info->lock);
4287         if (ret < 0 && ret != -ENOSPC)
4288                 goto out;
4289         if (ret)
4290                 space_info->full = 1;
4291         else
4292                 ret = 1;
4293
4294         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4295 out:
4296         space_info->chunk_alloc = 0;
4297         spin_unlock(&space_info->lock);
4298         mutex_unlock(&fs_info->chunk_mutex);
4299         /*
4300          * When we allocate a new chunk we reserve space in the chunk block
4301          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4302          * add new nodes/leafs to it if we end up needing to do it when
4303          * inserting the chunk item and updating device items as part of the
4304          * second phase of chunk allocation, performed by
4305          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4306          * large number of new block groups to create in our transaction
4307          * handle's new_bgs list to avoid exhausting the chunk block reserve
4308          * in extreme cases - like having a single transaction create many new
4309          * block groups when starting to write out the free space caches of all
4310          * the block groups that were made dirty during the lifetime of the
4311          * transaction.
4312          */
4313         if (trans->can_flush_pending_bgs &&
4314             trans->chunk_bytes_reserved >= (2 * 1024 * 1024ull)) {
4315                 btrfs_create_pending_block_groups(trans, trans->root);
4316                 btrfs_trans_release_chunk_metadata(trans);
4317         }
4318         return ret;
4319 }
4320
4321 static int can_overcommit(struct btrfs_root *root,
4322                           struct btrfs_space_info *space_info, u64 bytes,
4323                           enum btrfs_reserve_flush_enum flush)
4324 {
4325         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4326         u64 profile = btrfs_get_alloc_profile(root, 0);
4327         u64 space_size;
4328         u64 avail;
4329         u64 used;
4330
4331         used = space_info->bytes_used + space_info->bytes_reserved +
4332                 space_info->bytes_pinned + space_info->bytes_readonly;
4333
4334         /*
4335          * We only want to allow over committing if we have lots of actual space
4336          * free, but if we don't have enough space to handle the global reserve
4337          * space then we could end up having a real enospc problem when trying
4338          * to allocate a chunk or some other such important allocation.
4339          */
4340         spin_lock(&global_rsv->lock);
4341         space_size = calc_global_rsv_need_space(global_rsv);
4342         spin_unlock(&global_rsv->lock);
4343         if (used + space_size >= space_info->total_bytes)
4344                 return 0;
4345
4346         used += space_info->bytes_may_use;
4347
4348         spin_lock(&root->fs_info->free_chunk_lock);
4349         avail = root->fs_info->free_chunk_space;
4350         spin_unlock(&root->fs_info->free_chunk_lock);
4351
4352         /*
4353          * If we have dup, raid1 or raid10 then only half of the free
4354          * space is actually useable.  For raid56, the space info used
4355          * doesn't include the parity drive, so we don't have to
4356          * change the math
4357          */
4358         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4359                        BTRFS_BLOCK_GROUP_RAID1 |
4360                        BTRFS_BLOCK_GROUP_RAID10))
4361                 avail >>= 1;
4362
4363         /*
4364          * If we aren't flushing all things, let us overcommit up to
4365          * 1/2th of the space. If we can flush, don't let us overcommit
4366          * too much, let it overcommit up to 1/8 of the space.
4367          */
4368         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4369                 avail >>= 3;
4370         else
4371                 avail >>= 1;
4372
4373         if (used + bytes < space_info->total_bytes + avail)
4374                 return 1;
4375         return 0;
4376 }
4377
4378 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4379                                          unsigned long nr_pages, int nr_items)
4380 {
4381         struct super_block *sb = root->fs_info->sb;
4382
4383         if (down_read_trylock(&sb->s_umount)) {
4384                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4385                 up_read(&sb->s_umount);
4386         } else {
4387                 /*
4388                  * We needn't worry the filesystem going from r/w to r/o though
4389                  * we don't acquire ->s_umount mutex, because the filesystem
4390                  * should guarantee the delalloc inodes list be empty after
4391                  * the filesystem is readonly(all dirty pages are written to
4392                  * the disk).
4393                  */
4394                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4395                 if (!current->journal_info)
4396                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4397         }
4398 }
4399
4400 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4401 {
4402         u64 bytes;
4403         int nr;
4404
4405         bytes = btrfs_calc_trans_metadata_size(root, 1);
4406         nr = (int)div64_u64(to_reclaim, bytes);
4407         if (!nr)
4408                 nr = 1;
4409         return nr;
4410 }
4411
4412 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4413
4414 /*
4415  * shrink metadata reservation for delalloc
4416  */
4417 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4418                             bool wait_ordered)
4419 {
4420         struct btrfs_block_rsv *block_rsv;
4421         struct btrfs_space_info *space_info;
4422         struct btrfs_trans_handle *trans;
4423         u64 delalloc_bytes;
4424         u64 max_reclaim;
4425         long time_left;
4426         unsigned long nr_pages;
4427         int loops;
4428         int items;
4429         enum btrfs_reserve_flush_enum flush;
4430
4431         /* Calc the number of the pages we need flush for space reservation */
4432         items = calc_reclaim_items_nr(root, to_reclaim);
4433         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4434
4435         trans = (struct btrfs_trans_handle *)current->journal_info;
4436         block_rsv = &root->fs_info->delalloc_block_rsv;
4437         space_info = block_rsv->space_info;
4438
4439         delalloc_bytes = percpu_counter_sum_positive(
4440                                                 &root->fs_info->delalloc_bytes);
4441         if (delalloc_bytes == 0) {
4442                 if (trans)
4443                         return;
4444                 if (wait_ordered)
4445                         btrfs_wait_ordered_roots(root->fs_info, items);
4446                 return;
4447         }
4448
4449         loops = 0;
4450         while (delalloc_bytes && loops < 3) {
4451                 max_reclaim = min(delalloc_bytes, to_reclaim);
4452                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4453                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4454                 /*
4455                  * We need to wait for the async pages to actually start before
4456                  * we do anything.
4457                  */
4458                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4459                 if (!max_reclaim)
4460                         goto skip_async;
4461
4462                 if (max_reclaim <= nr_pages)
4463                         max_reclaim = 0;
4464                 else
4465                         max_reclaim -= nr_pages;
4466
4467                 wait_event(root->fs_info->async_submit_wait,
4468                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4469                            (int)max_reclaim);
4470 skip_async:
4471                 if (!trans)
4472                         flush = BTRFS_RESERVE_FLUSH_ALL;
4473                 else
4474                         flush = BTRFS_RESERVE_NO_FLUSH;
4475                 spin_lock(&space_info->lock);
4476                 if (can_overcommit(root, space_info, orig, flush)) {
4477                         spin_unlock(&space_info->lock);
4478                         break;
4479                 }
4480                 spin_unlock(&space_info->lock);
4481
4482                 loops++;
4483                 if (wait_ordered && !trans) {
4484                         btrfs_wait_ordered_roots(root->fs_info, items);
4485                 } else {
4486                         time_left = schedule_timeout_killable(1);
4487                         if (time_left)
4488                                 break;
4489                 }
4490                 delalloc_bytes = percpu_counter_sum_positive(
4491                                                 &root->fs_info->delalloc_bytes);
4492         }
4493 }
4494
4495 /**
4496  * maybe_commit_transaction - possibly commit the transaction if its ok to
4497  * @root - the root we're allocating for
4498  * @bytes - the number of bytes we want to reserve
4499  * @force - force the commit
4500  *
4501  * This will check to make sure that committing the transaction will actually
4502  * get us somewhere and then commit the transaction if it does.  Otherwise it
4503  * will return -ENOSPC.
4504  */
4505 static int may_commit_transaction(struct btrfs_root *root,
4506                                   struct btrfs_space_info *space_info,
4507                                   u64 bytes, int force)
4508 {
4509         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4510         struct btrfs_trans_handle *trans;
4511
4512         trans = (struct btrfs_trans_handle *)current->journal_info;
4513         if (trans)
4514                 return -EAGAIN;
4515
4516         if (force)
4517                 goto commit;
4518
4519         /* See if there is enough pinned space to make this reservation */
4520         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4521                                    bytes) >= 0)
4522                 goto commit;
4523
4524         /*
4525          * See if there is some space in the delayed insertion reservation for
4526          * this reservation.
4527          */
4528         if (space_info != delayed_rsv->space_info)
4529                 return -ENOSPC;
4530
4531         spin_lock(&delayed_rsv->lock);
4532         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4533                                    bytes - delayed_rsv->size) >= 0) {
4534                 spin_unlock(&delayed_rsv->lock);
4535                 return -ENOSPC;
4536         }
4537         spin_unlock(&delayed_rsv->lock);
4538
4539 commit:
4540         trans = btrfs_join_transaction(root);
4541         if (IS_ERR(trans))
4542                 return -ENOSPC;
4543
4544         return btrfs_commit_transaction(trans, root);
4545 }
4546
4547 enum flush_state {
4548         FLUSH_DELAYED_ITEMS_NR  =       1,
4549         FLUSH_DELAYED_ITEMS     =       2,
4550         FLUSH_DELALLOC          =       3,
4551         FLUSH_DELALLOC_WAIT     =       4,
4552         ALLOC_CHUNK             =       5,
4553         COMMIT_TRANS            =       6,
4554 };
4555
4556 static int flush_space(struct btrfs_root *root,
4557                        struct btrfs_space_info *space_info, u64 num_bytes,
4558                        u64 orig_bytes, int state)
4559 {
4560         struct btrfs_trans_handle *trans;
4561         int nr;
4562         int ret = 0;
4563
4564         switch (state) {
4565         case FLUSH_DELAYED_ITEMS_NR:
4566         case FLUSH_DELAYED_ITEMS:
4567                 if (state == FLUSH_DELAYED_ITEMS_NR)
4568                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4569                 else
4570                         nr = -1;
4571
4572                 trans = btrfs_join_transaction(root);
4573                 if (IS_ERR(trans)) {
4574                         ret = PTR_ERR(trans);
4575                         break;
4576                 }
4577                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4578                 btrfs_end_transaction(trans, root);
4579                 break;
4580         case FLUSH_DELALLOC:
4581         case FLUSH_DELALLOC_WAIT:
4582                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4583                                 state == FLUSH_DELALLOC_WAIT);
4584                 break;
4585         case ALLOC_CHUNK:
4586                 trans = btrfs_join_transaction(root);
4587                 if (IS_ERR(trans)) {
4588                         ret = PTR_ERR(trans);
4589                         break;
4590                 }
4591                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4592                                      btrfs_get_alloc_profile(root, 0),
4593                                      CHUNK_ALLOC_NO_FORCE);
4594                 btrfs_end_transaction(trans, root);
4595                 if (ret == -ENOSPC)
4596                         ret = 0;
4597                 break;
4598         case COMMIT_TRANS:
4599                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4600                 break;
4601         default:
4602                 ret = -ENOSPC;
4603                 break;
4604         }
4605
4606         return ret;
4607 }
4608
4609 static inline u64
4610 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4611                                  struct btrfs_space_info *space_info)
4612 {
4613         u64 used;
4614         u64 expected;
4615         u64 to_reclaim;
4616
4617         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4618                                 16 * 1024 * 1024);
4619         spin_lock(&space_info->lock);
4620         if (can_overcommit(root, space_info, to_reclaim,
4621                            BTRFS_RESERVE_FLUSH_ALL)) {
4622                 to_reclaim = 0;
4623                 goto out;
4624         }
4625
4626         used = space_info->bytes_used + space_info->bytes_reserved +
4627                space_info->bytes_pinned + space_info->bytes_readonly +
4628                space_info->bytes_may_use;
4629         if (can_overcommit(root, space_info, 1024 * 1024,
4630                            BTRFS_RESERVE_FLUSH_ALL))
4631                 expected = div_factor_fine(space_info->total_bytes, 95);
4632         else
4633                 expected = div_factor_fine(space_info->total_bytes, 90);
4634
4635         if (used > expected)
4636                 to_reclaim = used - expected;
4637         else
4638                 to_reclaim = 0;
4639         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4640                                      space_info->bytes_reserved);
4641 out:
4642         spin_unlock(&space_info->lock);
4643
4644         return to_reclaim;
4645 }
4646
4647 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4648                                         struct btrfs_fs_info *fs_info, u64 used)
4649 {
4650         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4651
4652         /* If we're just plain full then async reclaim just slows us down. */
4653         if (space_info->bytes_used >= thresh)
4654                 return 0;
4655
4656         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4657                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4658 }
4659
4660 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4661                                        struct btrfs_fs_info *fs_info,
4662                                        int flush_state)
4663 {
4664         u64 used;
4665
4666         spin_lock(&space_info->lock);
4667         /*
4668          * We run out of space and have not got any free space via flush_space,
4669          * so don't bother doing async reclaim.
4670          */
4671         if (flush_state > COMMIT_TRANS && space_info->full) {
4672                 spin_unlock(&space_info->lock);
4673                 return 0;
4674         }
4675
4676         used = space_info->bytes_used + space_info->bytes_reserved +
4677                space_info->bytes_pinned + space_info->bytes_readonly +
4678                space_info->bytes_may_use;
4679         if (need_do_async_reclaim(space_info, fs_info, used)) {
4680                 spin_unlock(&space_info->lock);
4681                 return 1;
4682         }
4683         spin_unlock(&space_info->lock);
4684
4685         return 0;
4686 }
4687
4688 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4689 {
4690         struct btrfs_fs_info *fs_info;
4691         struct btrfs_space_info *space_info;
4692         u64 to_reclaim;
4693         int flush_state;
4694
4695         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4696         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4697
4698         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4699                                                       space_info);
4700         if (!to_reclaim)
4701                 return;
4702
4703         flush_state = FLUSH_DELAYED_ITEMS_NR;
4704         do {
4705                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4706                             to_reclaim, flush_state);
4707                 flush_state++;
4708                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4709                                                  flush_state))
4710                         return;
4711         } while (flush_state < COMMIT_TRANS);
4712 }
4713
4714 void btrfs_init_async_reclaim_work(struct work_struct *work)
4715 {
4716         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4717 }
4718
4719 /**
4720  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4721  * @root - the root we're allocating for
4722  * @block_rsv - the block_rsv we're allocating for
4723  * @orig_bytes - the number of bytes we want
4724  * @flush - whether or not we can flush to make our reservation
4725  *
4726  * This will reserve orgi_bytes number of bytes from the space info associated
4727  * with the block_rsv.  If there is not enough space it will make an attempt to
4728  * flush out space to make room.  It will do this by flushing delalloc if
4729  * possible or committing the transaction.  If flush is 0 then no attempts to
4730  * regain reservations will be made and this will fail if there is not enough
4731  * space already.
4732  */
4733 static int reserve_metadata_bytes(struct btrfs_root *root,
4734                                   struct btrfs_block_rsv *block_rsv,
4735                                   u64 orig_bytes,
4736                                   enum btrfs_reserve_flush_enum flush)
4737 {
4738         struct btrfs_space_info *space_info = block_rsv->space_info;
4739         u64 used;
4740         u64 num_bytes = orig_bytes;
4741         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4742         int ret = 0;
4743         bool flushing = false;
4744
4745 again:
4746         ret = 0;
4747         spin_lock(&space_info->lock);
4748         /*
4749          * We only want to wait if somebody other than us is flushing and we
4750          * are actually allowed to flush all things.
4751          */
4752         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4753                space_info->flush) {
4754                 spin_unlock(&space_info->lock);
4755                 /*
4756                  * If we have a trans handle we can't wait because the flusher
4757                  * may have to commit the transaction, which would mean we would
4758                  * deadlock since we are waiting for the flusher to finish, but
4759                  * hold the current transaction open.
4760                  */
4761                 if (current->journal_info)
4762                         return -EAGAIN;
4763                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4764                 /* Must have been killed, return */
4765                 if (ret)
4766                         return -EINTR;
4767
4768                 spin_lock(&space_info->lock);
4769         }
4770
4771         ret = -ENOSPC;
4772         used = space_info->bytes_used + space_info->bytes_reserved +
4773                 space_info->bytes_pinned + space_info->bytes_readonly +
4774                 space_info->bytes_may_use;
4775
4776         /*
4777          * The idea here is that we've not already over-reserved the block group
4778          * then we can go ahead and save our reservation first and then start
4779          * flushing if we need to.  Otherwise if we've already overcommitted
4780          * lets start flushing stuff first and then come back and try to make
4781          * our reservation.
4782          */
4783         if (used <= space_info->total_bytes) {
4784                 if (used + orig_bytes <= space_info->total_bytes) {
4785                         space_info->bytes_may_use += orig_bytes;
4786                         trace_btrfs_space_reservation(root->fs_info,
4787                                 "space_info", space_info->flags, orig_bytes, 1);
4788                         ret = 0;
4789                 } else {
4790                         /*
4791                          * Ok set num_bytes to orig_bytes since we aren't
4792                          * overocmmitted, this way we only try and reclaim what
4793                          * we need.
4794                          */
4795                         num_bytes = orig_bytes;
4796                 }
4797         } else {
4798                 /*
4799                  * Ok we're over committed, set num_bytes to the overcommitted
4800                  * amount plus the amount of bytes that we need for this
4801                  * reservation.
4802                  */
4803                 num_bytes = used - space_info->total_bytes +
4804                         (orig_bytes * 2);
4805         }
4806
4807         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4808                 space_info->bytes_may_use += orig_bytes;
4809                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4810                                               space_info->flags, orig_bytes,
4811                                               1);
4812                 ret = 0;
4813         }
4814
4815         /*
4816          * Couldn't make our reservation, save our place so while we're trying
4817          * to reclaim space we can actually use it instead of somebody else
4818          * stealing it from us.
4819          *
4820          * We make the other tasks wait for the flush only when we can flush
4821          * all things.
4822          */
4823         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4824                 flushing = true;
4825                 space_info->flush = 1;
4826         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4827                 used += orig_bytes;
4828                 /*
4829                  * We will do the space reservation dance during log replay,
4830                  * which means we won't have fs_info->fs_root set, so don't do
4831                  * the async reclaim as we will panic.
4832                  */
4833                 if (!root->fs_info->log_root_recovering &&
4834                     need_do_async_reclaim(space_info, root->fs_info, used) &&
4835                     !work_busy(&root->fs_info->async_reclaim_work))
4836                         queue_work(system_unbound_wq,
4837                                    &root->fs_info->async_reclaim_work);
4838         }
4839         spin_unlock(&space_info->lock);
4840
4841         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4842                 goto out;
4843
4844         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4845                           flush_state);
4846         flush_state++;
4847
4848         /*
4849          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4850          * would happen. So skip delalloc flush.
4851          */
4852         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4853             (flush_state == FLUSH_DELALLOC ||
4854              flush_state == FLUSH_DELALLOC_WAIT))
4855                 flush_state = ALLOC_CHUNK;
4856
4857         if (!ret)
4858                 goto again;
4859         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4860                  flush_state < COMMIT_TRANS)
4861                 goto again;
4862         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4863                  flush_state <= COMMIT_TRANS)
4864                 goto again;
4865
4866 out:
4867         if (ret == -ENOSPC &&
4868             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4869                 struct btrfs_block_rsv *global_rsv =
4870                         &root->fs_info->global_block_rsv;
4871
4872                 if (block_rsv != global_rsv &&
4873                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4874                         ret = 0;
4875         }
4876         if (ret == -ENOSPC)
4877                 trace_btrfs_space_reservation(root->fs_info,
4878                                               "space_info:enospc",
4879                                               space_info->flags, orig_bytes, 1);
4880         if (flushing) {
4881                 spin_lock(&space_info->lock);
4882                 space_info->flush = 0;
4883                 wake_up_all(&space_info->wait);
4884                 spin_unlock(&space_info->lock);
4885         }
4886         return ret;
4887 }
4888
4889 static struct btrfs_block_rsv *get_block_rsv(
4890                                         const struct btrfs_trans_handle *trans,
4891                                         const struct btrfs_root *root)
4892 {
4893         struct btrfs_block_rsv *block_rsv = NULL;
4894
4895         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4896             (root == root->fs_info->csum_root && trans->adding_csums) ||
4897              (root == root->fs_info->uuid_root))
4898                 block_rsv = trans->block_rsv;
4899
4900         if (!block_rsv)
4901                 block_rsv = root->block_rsv;
4902
4903         if (!block_rsv)
4904                 block_rsv = &root->fs_info->empty_block_rsv;
4905
4906         return block_rsv;
4907 }
4908
4909 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4910                                u64 num_bytes)
4911 {
4912         int ret = -ENOSPC;
4913         spin_lock(&block_rsv->lock);
4914         if (block_rsv->reserved >= num_bytes) {
4915                 block_rsv->reserved -= num_bytes;
4916                 if (block_rsv->reserved < block_rsv->size)
4917                         block_rsv->full = 0;
4918                 ret = 0;
4919         }
4920         spin_unlock(&block_rsv->lock);
4921         return ret;
4922 }
4923
4924 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4925                                 u64 num_bytes, int update_size)
4926 {
4927         spin_lock(&block_rsv->lock);
4928         block_rsv->reserved += num_bytes;
4929         if (update_size)
4930                 block_rsv->size += num_bytes;
4931         else if (block_rsv->reserved >= block_rsv->size)
4932                 block_rsv->full = 1;
4933         spin_unlock(&block_rsv->lock);
4934 }
4935
4936 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4937                              struct btrfs_block_rsv *dest, u64 num_bytes,
4938                              int min_factor)
4939 {
4940         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4941         u64 min_bytes;
4942
4943         if (global_rsv->space_info != dest->space_info)
4944                 return -ENOSPC;
4945
4946         spin_lock(&global_rsv->lock);
4947         min_bytes = div_factor(global_rsv->size, min_factor);
4948         if (global_rsv->reserved < min_bytes + num_bytes) {
4949                 spin_unlock(&global_rsv->lock);
4950                 return -ENOSPC;
4951         }
4952         global_rsv->reserved -= num_bytes;
4953         if (global_rsv->reserved < global_rsv->size)
4954                 global_rsv->full = 0;
4955         spin_unlock(&global_rsv->lock);
4956
4957         block_rsv_add_bytes(dest, num_bytes, 1);
4958         return 0;
4959 }
4960
4961 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4962                                     struct btrfs_block_rsv *block_rsv,
4963                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4964 {
4965         struct btrfs_space_info *space_info = block_rsv->space_info;
4966
4967         spin_lock(&block_rsv->lock);
4968         if (num_bytes == (u64)-1)
4969                 num_bytes = block_rsv->size;
4970         block_rsv->size -= num_bytes;
4971         if (block_rsv->reserved >= block_rsv->size) {
4972                 num_bytes = block_rsv->reserved - block_rsv->size;
4973                 block_rsv->reserved = block_rsv->size;
4974                 block_rsv->full = 1;
4975         } else {
4976                 num_bytes = 0;
4977         }
4978         spin_unlock(&block_rsv->lock);
4979
4980         if (num_bytes > 0) {
4981                 if (dest) {
4982                         spin_lock(&dest->lock);
4983                         if (!dest->full) {
4984                                 u64 bytes_to_add;
4985
4986                                 bytes_to_add = dest->size - dest->reserved;
4987                                 bytes_to_add = min(num_bytes, bytes_to_add);
4988                                 dest->reserved += bytes_to_add;
4989                                 if (dest->reserved >= dest->size)
4990                                         dest->full = 1;
4991                                 num_bytes -= bytes_to_add;
4992                         }
4993                         spin_unlock(&dest->lock);
4994                 }
4995                 if (num_bytes) {
4996                         spin_lock(&space_info->lock);
4997                         space_info->bytes_may_use -= num_bytes;
4998                         trace_btrfs_space_reservation(fs_info, "space_info",
4999                                         space_info->flags, num_bytes, 0);
5000                         spin_unlock(&space_info->lock);
5001                 }
5002         }
5003 }
5004
5005 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
5006                                    struct btrfs_block_rsv *dst, u64 num_bytes)
5007 {
5008         int ret;
5009
5010         ret = block_rsv_use_bytes(src, num_bytes);
5011         if (ret)
5012                 return ret;
5013
5014         block_rsv_add_bytes(dst, num_bytes, 1);
5015         return 0;
5016 }
5017
5018 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5019 {
5020         memset(rsv, 0, sizeof(*rsv));
5021         spin_lock_init(&rsv->lock);
5022         rsv->type = type;
5023 }
5024
5025 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5026                                               unsigned short type)
5027 {
5028         struct btrfs_block_rsv *block_rsv;
5029         struct btrfs_fs_info *fs_info = root->fs_info;
5030
5031         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5032         if (!block_rsv)
5033                 return NULL;
5034
5035         btrfs_init_block_rsv(block_rsv, type);
5036         block_rsv->space_info = __find_space_info(fs_info,
5037                                                   BTRFS_BLOCK_GROUP_METADATA);
5038         return block_rsv;
5039 }
5040
5041 void btrfs_free_block_rsv(struct btrfs_root *root,
5042                           struct btrfs_block_rsv *rsv)
5043 {
5044         if (!rsv)
5045                 return;
5046         btrfs_block_rsv_release(root, rsv, (u64)-1);
5047         kfree(rsv);
5048 }
5049
5050 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5051 {
5052         kfree(rsv);
5053 }
5054
5055 int btrfs_block_rsv_add(struct btrfs_root *root,
5056                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5057                         enum btrfs_reserve_flush_enum flush)
5058 {
5059         int ret;
5060
5061         if (num_bytes == 0)
5062                 return 0;
5063
5064         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5065         if (!ret) {
5066                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5067                 return 0;
5068         }
5069
5070         return ret;
5071 }
5072
5073 int btrfs_block_rsv_check(struct btrfs_root *root,
5074                           struct btrfs_block_rsv *block_rsv, int min_factor)
5075 {
5076         u64 num_bytes = 0;
5077         int ret = -ENOSPC;
5078
5079         if (!block_rsv)
5080                 return 0;
5081
5082         spin_lock(&block_rsv->lock);
5083         num_bytes = div_factor(block_rsv->size, min_factor);
5084         if (block_rsv->reserved >= num_bytes)
5085                 ret = 0;
5086         spin_unlock(&block_rsv->lock);
5087
5088         return ret;
5089 }
5090
5091 int btrfs_block_rsv_refill(struct btrfs_root *root,
5092                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5093                            enum btrfs_reserve_flush_enum flush)
5094 {
5095         u64 num_bytes = 0;
5096         int ret = -ENOSPC;
5097
5098         if (!block_rsv)
5099                 return 0;
5100
5101         spin_lock(&block_rsv->lock);
5102         num_bytes = min_reserved;
5103         if (block_rsv->reserved >= num_bytes)
5104                 ret = 0;
5105         else
5106                 num_bytes -= block_rsv->reserved;
5107         spin_unlock(&block_rsv->lock);
5108
5109         if (!ret)
5110                 return 0;
5111
5112         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5113         if (!ret) {
5114                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5115                 return 0;
5116         }
5117
5118         return ret;
5119 }
5120
5121 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5122                             struct btrfs_block_rsv *dst_rsv,
5123                             u64 num_bytes)
5124 {
5125         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5126 }
5127
5128 void btrfs_block_rsv_release(struct btrfs_root *root,
5129                              struct btrfs_block_rsv *block_rsv,
5130                              u64 num_bytes)
5131 {
5132         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5133         if (global_rsv == block_rsv ||
5134             block_rsv->space_info != global_rsv->space_info)
5135                 global_rsv = NULL;
5136         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5137                                 num_bytes);
5138 }
5139
5140 /*
5141  * helper to calculate size of global block reservation.
5142  * the desired value is sum of space used by extent tree,
5143  * checksum tree and root tree
5144  */
5145 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5146 {
5147         struct btrfs_space_info *sinfo;
5148         u64 num_bytes;
5149         u64 meta_used;
5150         u64 data_used;
5151         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5152
5153         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5154         spin_lock(&sinfo->lock);
5155         data_used = sinfo->bytes_used;
5156         spin_unlock(&sinfo->lock);
5157
5158         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5159         spin_lock(&sinfo->lock);
5160         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5161                 data_used = 0;
5162         meta_used = sinfo->bytes_used;
5163         spin_unlock(&sinfo->lock);
5164
5165         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5166                     csum_size * 2;
5167         num_bytes += div_u64(data_used + meta_used, 50);
5168
5169         if (num_bytes * 3 > meta_used)
5170                 num_bytes = div_u64(meta_used, 3);
5171
5172         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5173 }
5174
5175 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5176 {
5177         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5178         struct btrfs_space_info *sinfo = block_rsv->space_info;
5179         u64 num_bytes;
5180
5181         num_bytes = calc_global_metadata_size(fs_info);
5182
5183         spin_lock(&sinfo->lock);
5184         spin_lock(&block_rsv->lock);
5185
5186         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
5187
5188         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5189                     sinfo->bytes_reserved + sinfo->bytes_readonly +
5190                     sinfo->bytes_may_use;
5191
5192         if (sinfo->total_bytes > num_bytes) {
5193                 num_bytes = sinfo->total_bytes - num_bytes;
5194                 block_rsv->reserved += num_bytes;
5195                 sinfo->bytes_may_use += num_bytes;
5196                 trace_btrfs_space_reservation(fs_info, "space_info",
5197                                       sinfo->flags, num_bytes, 1);
5198         }
5199
5200         if (block_rsv->reserved >= block_rsv->size) {
5201                 num_bytes = block_rsv->reserved - block_rsv->size;
5202                 sinfo->bytes_may_use -= num_bytes;
5203                 trace_btrfs_space_reservation(fs_info, "space_info",
5204                                       sinfo->flags, num_bytes, 0);
5205                 block_rsv->reserved = block_rsv->size;
5206                 block_rsv->full = 1;
5207         }
5208
5209         spin_unlock(&block_rsv->lock);
5210         spin_unlock(&sinfo->lock);
5211 }
5212
5213 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5214 {
5215         struct btrfs_space_info *space_info;
5216
5217         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5218         fs_info->chunk_block_rsv.space_info = space_info;
5219
5220         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5221         fs_info->global_block_rsv.space_info = space_info;
5222         fs_info->delalloc_block_rsv.space_info = space_info;
5223         fs_info->trans_block_rsv.space_info = space_info;
5224         fs_info->empty_block_rsv.space_info = space_info;
5225         fs_info->delayed_block_rsv.space_info = space_info;
5226
5227         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5228         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5229         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5230         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5231         if (fs_info->quota_root)
5232                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5233         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5234
5235         update_global_block_rsv(fs_info);
5236 }
5237
5238 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5239 {
5240         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5241                                 (u64)-1);
5242         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5243         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5244         WARN_ON(fs_info->trans_block_rsv.size > 0);
5245         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5246         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5247         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5248         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5249         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5250 }
5251
5252 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5253                                   struct btrfs_root *root)
5254 {
5255         if (!trans->block_rsv)
5256                 return;
5257
5258         if (!trans->bytes_reserved)
5259                 return;
5260
5261         trace_btrfs_space_reservation(root->fs_info, "transaction",
5262                                       trans->transid, trans->bytes_reserved, 0);
5263         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5264         trans->bytes_reserved = 0;
5265 }
5266
5267 /*
5268  * To be called after all the new block groups attached to the transaction
5269  * handle have been created (btrfs_create_pending_block_groups()).
5270  */
5271 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5272 {
5273         struct btrfs_fs_info *fs_info = trans->root->fs_info;
5274
5275         if (!trans->chunk_bytes_reserved)
5276                 return;
5277
5278         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5279
5280         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5281                                 trans->chunk_bytes_reserved);
5282         trans->chunk_bytes_reserved = 0;
5283 }
5284
5285 /* Can only return 0 or -ENOSPC */
5286 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5287                                   struct inode *inode)
5288 {
5289         struct btrfs_root *root = BTRFS_I(inode)->root;
5290         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5291         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5292
5293         /*
5294          * We need to hold space in order to delete our orphan item once we've
5295          * added it, so this takes the reservation so we can release it later
5296          * when we are truly done with the orphan item.
5297          */
5298         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5299         trace_btrfs_space_reservation(root->fs_info, "orphan",
5300                                       btrfs_ino(inode), num_bytes, 1);
5301         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5302 }
5303
5304 void btrfs_orphan_release_metadata(struct inode *inode)
5305 {
5306         struct btrfs_root *root = BTRFS_I(inode)->root;
5307         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5308         trace_btrfs_space_reservation(root->fs_info, "orphan",
5309                                       btrfs_ino(inode), num_bytes, 0);
5310         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5311 }
5312
5313 /*
5314  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5315  * root: the root of the parent directory
5316  * rsv: block reservation
5317  * items: the number of items that we need do reservation
5318  * qgroup_reserved: used to return the reserved size in qgroup
5319  *
5320  * This function is used to reserve the space for snapshot/subvolume
5321  * creation and deletion. Those operations are different with the
5322  * common file/directory operations, they change two fs/file trees
5323  * and root tree, the number of items that the qgroup reserves is
5324  * different with the free space reservation. So we can not use
5325  * the space reseravtion mechanism in start_transaction().
5326  */
5327 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5328                                      struct btrfs_block_rsv *rsv,
5329                                      int items,
5330                                      u64 *qgroup_reserved,
5331                                      bool use_global_rsv)
5332 {
5333         u64 num_bytes;
5334         int ret;
5335         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5336
5337         if (root->fs_info->quota_enabled) {
5338                 /* One for parent inode, two for dir entries */
5339                 num_bytes = 3 * root->nodesize;
5340                 ret = btrfs_qgroup_reserve(root, num_bytes);
5341                 if (ret)
5342                         return ret;
5343         } else {
5344                 num_bytes = 0;
5345         }
5346
5347         *qgroup_reserved = num_bytes;
5348
5349         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5350         rsv->space_info = __find_space_info(root->fs_info,
5351                                             BTRFS_BLOCK_GROUP_METADATA);
5352         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5353                                   BTRFS_RESERVE_FLUSH_ALL);
5354
5355         if (ret == -ENOSPC && use_global_rsv)
5356                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5357
5358         if (ret) {
5359                 if (*qgroup_reserved)
5360                         btrfs_qgroup_free(root, *qgroup_reserved);
5361         }
5362
5363         return ret;
5364 }
5365
5366 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5367                                       struct btrfs_block_rsv *rsv,
5368                                       u64 qgroup_reserved)
5369 {
5370         btrfs_block_rsv_release(root, rsv, (u64)-1);
5371 }
5372
5373 /**
5374  * drop_outstanding_extent - drop an outstanding extent
5375  * @inode: the inode we're dropping the extent for
5376  * @num_bytes: the number of bytes we're relaseing.
5377  *
5378  * This is called when we are freeing up an outstanding extent, either called
5379  * after an error or after an extent is written.  This will return the number of
5380  * reserved extents that need to be freed.  This must be called with
5381  * BTRFS_I(inode)->lock held.
5382  */
5383 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5384 {
5385         unsigned drop_inode_space = 0;
5386         unsigned dropped_extents = 0;
5387         unsigned num_extents = 0;
5388
5389         num_extents = (unsigned)div64_u64(num_bytes +
5390                                           BTRFS_MAX_EXTENT_SIZE - 1,
5391                                           BTRFS_MAX_EXTENT_SIZE);
5392         ASSERT(num_extents);
5393         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5394         BTRFS_I(inode)->outstanding_extents -= num_extents;
5395
5396         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5397             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5398                                &BTRFS_I(inode)->runtime_flags))
5399                 drop_inode_space = 1;
5400
5401         /*
5402          * If we have more or the same amount of outsanding extents than we have
5403          * reserved then we need to leave the reserved extents count alone.
5404          */
5405         if (BTRFS_I(inode)->outstanding_extents >=
5406             BTRFS_I(inode)->reserved_extents)
5407                 return drop_inode_space;
5408
5409         dropped_extents = BTRFS_I(inode)->reserved_extents -
5410                 BTRFS_I(inode)->outstanding_extents;
5411         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5412         return dropped_extents + drop_inode_space;
5413 }
5414
5415 /**
5416  * calc_csum_metadata_size - return the amount of metada space that must be
5417  *      reserved/free'd for the given bytes.
5418  * @inode: the inode we're manipulating
5419  * @num_bytes: the number of bytes in question
5420  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5421  *
5422  * This adjusts the number of csum_bytes in the inode and then returns the
5423  * correct amount of metadata that must either be reserved or freed.  We
5424  * calculate how many checksums we can fit into one leaf and then divide the
5425  * number of bytes that will need to be checksumed by this value to figure out
5426  * how many checksums will be required.  If we are adding bytes then the number
5427  * may go up and we will return the number of additional bytes that must be
5428  * reserved.  If it is going down we will return the number of bytes that must
5429  * be freed.
5430  *
5431  * This must be called with BTRFS_I(inode)->lock held.
5432  */
5433 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5434                                    int reserve)
5435 {
5436         struct btrfs_root *root = BTRFS_I(inode)->root;
5437         u64 old_csums, num_csums;
5438
5439         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5440             BTRFS_I(inode)->csum_bytes == 0)
5441                 return 0;
5442
5443         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5444         if (reserve)
5445                 BTRFS_I(inode)->csum_bytes += num_bytes;
5446         else
5447                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5448         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5449
5450         /* No change, no need to reserve more */
5451         if (old_csums == num_csums)
5452                 return 0;
5453
5454         if (reserve)
5455                 return btrfs_calc_trans_metadata_size(root,
5456                                                       num_csums - old_csums);
5457
5458         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5459 }
5460
5461 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5462 {
5463         struct btrfs_root *root = BTRFS_I(inode)->root;
5464         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5465         u64 to_reserve = 0;
5466         u64 csum_bytes;
5467         unsigned nr_extents = 0;
5468         int extra_reserve = 0;
5469         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5470         int ret = 0;
5471         bool delalloc_lock = true;
5472         u64 to_free = 0;
5473         unsigned dropped;
5474
5475         /* If we are a free space inode we need to not flush since we will be in
5476          * the middle of a transaction commit.  We also don't need the delalloc
5477          * mutex since we won't race with anybody.  We need this mostly to make
5478          * lockdep shut its filthy mouth.
5479          */
5480         if (btrfs_is_free_space_inode(inode)) {
5481                 flush = BTRFS_RESERVE_NO_FLUSH;
5482                 delalloc_lock = false;
5483         }
5484
5485         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5486             btrfs_transaction_in_commit(root->fs_info))
5487                 schedule_timeout(1);
5488
5489         if (delalloc_lock)
5490                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5491
5492         num_bytes = ALIGN(num_bytes, root->sectorsize);
5493
5494         spin_lock(&BTRFS_I(inode)->lock);
5495         nr_extents = (unsigned)div64_u64(num_bytes +
5496                                          BTRFS_MAX_EXTENT_SIZE - 1,
5497                                          BTRFS_MAX_EXTENT_SIZE);
5498         BTRFS_I(inode)->outstanding_extents += nr_extents;
5499         nr_extents = 0;
5500
5501         if (BTRFS_I(inode)->outstanding_extents >
5502             BTRFS_I(inode)->reserved_extents)
5503                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5504                         BTRFS_I(inode)->reserved_extents;
5505
5506         /*
5507          * Add an item to reserve for updating the inode when we complete the
5508          * delalloc io.
5509          */
5510         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5511                       &BTRFS_I(inode)->runtime_flags)) {
5512                 nr_extents++;
5513                 extra_reserve = 1;
5514         }
5515
5516         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5517         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5518         csum_bytes = BTRFS_I(inode)->csum_bytes;
5519         spin_unlock(&BTRFS_I(inode)->lock);
5520
5521         if (root->fs_info->quota_enabled) {
5522                 ret = btrfs_qgroup_reserve(root, nr_extents * root->nodesize);
5523                 if (ret)
5524                         goto out_fail;
5525         }
5526
5527         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5528         if (unlikely(ret)) {
5529                 if (root->fs_info->quota_enabled)
5530                         btrfs_qgroup_free(root, nr_extents * root->nodesize);
5531                 goto out_fail;
5532         }
5533
5534         spin_lock(&BTRFS_I(inode)->lock);
5535         if (extra_reserve) {
5536                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5537                         &BTRFS_I(inode)->runtime_flags);
5538                 nr_extents--;
5539         }
5540         BTRFS_I(inode)->reserved_extents += nr_extents;
5541         spin_unlock(&BTRFS_I(inode)->lock);
5542
5543         if (delalloc_lock)
5544                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5545
5546         if (to_reserve)
5547                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5548                                               btrfs_ino(inode), to_reserve, 1);
5549         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5550
5551         return 0;
5552
5553 out_fail:
5554         spin_lock(&BTRFS_I(inode)->lock);
5555         dropped = drop_outstanding_extent(inode, num_bytes);
5556         /*
5557          * If the inodes csum_bytes is the same as the original
5558          * csum_bytes then we know we haven't raced with any free()ers
5559          * so we can just reduce our inodes csum bytes and carry on.
5560          */
5561         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5562                 calc_csum_metadata_size(inode, num_bytes, 0);
5563         } else {
5564                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5565                 u64 bytes;
5566
5567                 /*
5568                  * This is tricky, but first we need to figure out how much we
5569                  * free'd from any free-ers that occured during this
5570                  * reservation, so we reset ->csum_bytes to the csum_bytes
5571                  * before we dropped our lock, and then call the free for the
5572                  * number of bytes that were freed while we were trying our
5573                  * reservation.
5574                  */
5575                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5576                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5577                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5578
5579
5580                 /*
5581                  * Now we need to see how much we would have freed had we not
5582                  * been making this reservation and our ->csum_bytes were not
5583                  * artificially inflated.
5584                  */
5585                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5586                 bytes = csum_bytes - orig_csum_bytes;
5587                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5588
5589                 /*
5590                  * Now reset ->csum_bytes to what it should be.  If bytes is
5591                  * more than to_free then we would have free'd more space had we
5592                  * not had an artificially high ->csum_bytes, so we need to free
5593                  * the remainder.  If bytes is the same or less then we don't
5594                  * need to do anything, the other free-ers did the correct
5595                  * thing.
5596                  */
5597                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5598                 if (bytes > to_free)
5599                         to_free = bytes - to_free;
5600                 else
5601                         to_free = 0;
5602         }
5603         spin_unlock(&BTRFS_I(inode)->lock);
5604         if (dropped)
5605                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5606
5607         if (to_free) {
5608                 btrfs_block_rsv_release(root, block_rsv, to_free);
5609                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5610                                               btrfs_ino(inode), to_free, 0);
5611         }
5612         if (delalloc_lock)
5613                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5614         return ret;
5615 }
5616
5617 /**
5618  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5619  * @inode: the inode to release the reservation for
5620  * @num_bytes: the number of bytes we're releasing
5621  *
5622  * This will release the metadata reservation for an inode.  This can be called
5623  * once we complete IO for a given set of bytes to release their metadata
5624  * reservations.
5625  */
5626 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5627 {
5628         struct btrfs_root *root = BTRFS_I(inode)->root;
5629         u64 to_free = 0;
5630         unsigned dropped;
5631
5632         num_bytes = ALIGN(num_bytes, root->sectorsize);
5633         spin_lock(&BTRFS_I(inode)->lock);
5634         dropped = drop_outstanding_extent(inode, num_bytes);
5635
5636         if (num_bytes)
5637                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5638         spin_unlock(&BTRFS_I(inode)->lock);
5639         if (dropped > 0)
5640                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5641
5642         if (btrfs_test_is_dummy_root(root))
5643                 return;
5644
5645         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5646                                       btrfs_ino(inode), to_free, 0);
5647
5648         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5649                                 to_free);
5650 }
5651
5652 /**
5653  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5654  * @inode: inode we're writing to
5655  * @num_bytes: the number of bytes we want to allocate
5656  *
5657  * This will do the following things
5658  *
5659  * o reserve space in the data space info for num_bytes
5660  * o reserve space in the metadata space info based on number of outstanding
5661  *   extents and how much csums will be needed
5662  * o add to the inodes ->delalloc_bytes
5663  * o add it to the fs_info's delalloc inodes list.
5664  *
5665  * This will return 0 for success and -ENOSPC if there is no space left.
5666  */
5667 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5668 {
5669         int ret;
5670
5671         ret = btrfs_check_data_free_space(inode, num_bytes, num_bytes);
5672         if (ret)
5673                 return ret;
5674
5675         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5676         if (ret) {
5677                 btrfs_free_reserved_data_space(inode, num_bytes);
5678                 return ret;
5679         }
5680
5681         return 0;
5682 }
5683
5684 /**
5685  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5686  * @inode: inode we're releasing space for
5687  * @num_bytes: the number of bytes we want to free up
5688  *
5689  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5690  * called in the case that we don't need the metadata AND data reservations
5691  * anymore.  So if there is an error or we insert an inline extent.
5692  *
5693  * This function will release the metadata space that was not used and will
5694  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5695  * list if there are no delalloc bytes left.
5696  */
5697 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5698 {
5699         btrfs_delalloc_release_metadata(inode, num_bytes);
5700         btrfs_free_reserved_data_space(inode, num_bytes);
5701 }
5702
5703 static int update_block_group(struct btrfs_trans_handle *trans,
5704                               struct btrfs_root *root, u64 bytenr,
5705                               u64 num_bytes, int alloc)
5706 {
5707         struct btrfs_block_group_cache *cache = NULL;
5708         struct btrfs_fs_info *info = root->fs_info;
5709         u64 total = num_bytes;
5710         u64 old_val;
5711         u64 byte_in_group;
5712         int factor;
5713
5714         /* block accounting for super block */
5715         spin_lock(&info->delalloc_root_lock);
5716         old_val = btrfs_super_bytes_used(info->super_copy);
5717         if (alloc)
5718                 old_val += num_bytes;
5719         else
5720                 old_val -= num_bytes;
5721         btrfs_set_super_bytes_used(info->super_copy, old_val);
5722         spin_unlock(&info->delalloc_root_lock);
5723
5724         while (total) {
5725                 cache = btrfs_lookup_block_group(info, bytenr);
5726                 if (!cache)
5727                         return -ENOENT;
5728                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5729                                     BTRFS_BLOCK_GROUP_RAID1 |
5730                                     BTRFS_BLOCK_GROUP_RAID10))
5731                         factor = 2;
5732                 else
5733                         factor = 1;
5734                 /*
5735                  * If this block group has free space cache written out, we
5736                  * need to make sure to load it if we are removing space.  This
5737                  * is because we need the unpinning stage to actually add the
5738                  * space back to the block group, otherwise we will leak space.
5739                  */
5740                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5741                         cache_block_group(cache, 1);
5742
5743                 byte_in_group = bytenr - cache->key.objectid;
5744                 WARN_ON(byte_in_group > cache->key.offset);
5745
5746                 spin_lock(&cache->space_info->lock);
5747                 spin_lock(&cache->lock);
5748
5749                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5750                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5751                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5752
5753                 old_val = btrfs_block_group_used(&cache->item);
5754                 num_bytes = min(total, cache->key.offset - byte_in_group);
5755                 if (alloc) {
5756                         old_val += num_bytes;
5757                         btrfs_set_block_group_used(&cache->item, old_val);
5758                         cache->reserved -= num_bytes;
5759                         cache->space_info->bytes_reserved -= num_bytes;
5760                         cache->space_info->bytes_used += num_bytes;
5761                         cache->space_info->disk_used += num_bytes * factor;
5762                         spin_unlock(&cache->lock);
5763                         spin_unlock(&cache->space_info->lock);
5764                 } else {
5765                         old_val -= num_bytes;
5766                         btrfs_set_block_group_used(&cache->item, old_val);
5767                         cache->pinned += num_bytes;
5768                         cache->space_info->bytes_pinned += num_bytes;
5769                         cache->space_info->bytes_used -= num_bytes;
5770                         cache->space_info->disk_used -= num_bytes * factor;
5771                         spin_unlock(&cache->lock);
5772                         spin_unlock(&cache->space_info->lock);
5773
5774                         set_extent_dirty(info->pinned_extents,
5775                                          bytenr, bytenr + num_bytes - 1,
5776                                          GFP_NOFS | __GFP_NOFAIL);
5777                         /*
5778                          * No longer have used bytes in this block group, queue
5779                          * it for deletion.
5780                          */
5781                         if (old_val == 0) {
5782                                 spin_lock(&info->unused_bgs_lock);
5783                                 if (list_empty(&cache->bg_list)) {
5784                                         btrfs_get_block_group(cache);
5785                                         list_add_tail(&cache->bg_list,
5786                                                       &info->unused_bgs);
5787                                 }
5788                                 spin_unlock(&info->unused_bgs_lock);
5789                         }
5790                 }
5791
5792                 spin_lock(&trans->transaction->dirty_bgs_lock);
5793                 if (list_empty(&cache->dirty_list)) {
5794                         list_add_tail(&cache->dirty_list,
5795                                       &trans->transaction->dirty_bgs);
5796                                 trans->transaction->num_dirty_bgs++;
5797                         btrfs_get_block_group(cache);
5798                 }
5799                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5800
5801                 btrfs_put_block_group(cache);
5802                 total -= num_bytes;
5803                 bytenr += num_bytes;
5804         }
5805         return 0;
5806 }
5807
5808 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5809 {
5810         struct btrfs_block_group_cache *cache;
5811         u64 bytenr;
5812
5813         spin_lock(&root->fs_info->block_group_cache_lock);
5814         bytenr = root->fs_info->first_logical_byte;
5815         spin_unlock(&root->fs_info->block_group_cache_lock);
5816
5817         if (bytenr < (u64)-1)
5818                 return bytenr;
5819
5820         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5821         if (!cache)
5822                 return 0;
5823
5824         bytenr = cache->key.objectid;
5825         btrfs_put_block_group(cache);
5826
5827         return bytenr;
5828 }
5829
5830 static int pin_down_extent(struct btrfs_root *root,
5831                            struct btrfs_block_group_cache *cache,
5832                            u64 bytenr, u64 num_bytes, int reserved)
5833 {
5834         spin_lock(&cache->space_info->lock);
5835         spin_lock(&cache->lock);
5836         cache->pinned += num_bytes;
5837         cache->space_info->bytes_pinned += num_bytes;
5838         if (reserved) {
5839                 cache->reserved -= num_bytes;
5840                 cache->space_info->bytes_reserved -= num_bytes;
5841         }
5842         spin_unlock(&cache->lock);
5843         spin_unlock(&cache->space_info->lock);
5844
5845         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5846                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5847         if (reserved)
5848                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5849         return 0;
5850 }
5851
5852 /*
5853  * this function must be called within transaction
5854  */
5855 int btrfs_pin_extent(struct btrfs_root *root,
5856                      u64 bytenr, u64 num_bytes, int reserved)
5857 {
5858         struct btrfs_block_group_cache *cache;
5859
5860         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5861         BUG_ON(!cache); /* Logic error */
5862
5863         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5864
5865         btrfs_put_block_group(cache);
5866         return 0;
5867 }
5868
5869 /*
5870  * this function must be called within transaction
5871  */
5872 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5873                                     u64 bytenr, u64 num_bytes)
5874 {
5875         struct btrfs_block_group_cache *cache;
5876         int ret;
5877
5878         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5879         if (!cache)
5880                 return -EINVAL;
5881
5882         /*
5883          * pull in the free space cache (if any) so that our pin
5884          * removes the free space from the cache.  We have load_only set
5885          * to one because the slow code to read in the free extents does check
5886          * the pinned extents.
5887          */
5888         cache_block_group(cache, 1);
5889
5890         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5891
5892         /* remove us from the free space cache (if we're there at all) */
5893         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5894         btrfs_put_block_group(cache);
5895         return ret;
5896 }
5897
5898 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5899 {
5900         int ret;
5901         struct btrfs_block_group_cache *block_group;
5902         struct btrfs_caching_control *caching_ctl;
5903
5904         block_group = btrfs_lookup_block_group(root->fs_info, start);
5905         if (!block_group)
5906                 return -EINVAL;
5907
5908         cache_block_group(block_group, 0);
5909         caching_ctl = get_caching_control(block_group);
5910
5911         if (!caching_ctl) {
5912                 /* Logic error */
5913                 BUG_ON(!block_group_cache_done(block_group));
5914                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5915         } else {
5916                 mutex_lock(&caching_ctl->mutex);
5917
5918                 if (start >= caching_ctl->progress) {
5919                         ret = add_excluded_extent(root, start, num_bytes);
5920                 } else if (start + num_bytes <= caching_ctl->progress) {
5921                         ret = btrfs_remove_free_space(block_group,
5922                                                       start, num_bytes);
5923                 } else {
5924                         num_bytes = caching_ctl->progress - start;
5925                         ret = btrfs_remove_free_space(block_group,
5926                                                       start, num_bytes);
5927                         if (ret)
5928                                 goto out_lock;
5929
5930                         num_bytes = (start + num_bytes) -
5931                                 caching_ctl->progress;
5932                         start = caching_ctl->progress;
5933                         ret = add_excluded_extent(root, start, num_bytes);
5934                 }
5935 out_lock:
5936                 mutex_unlock(&caching_ctl->mutex);
5937                 put_caching_control(caching_ctl);
5938         }
5939         btrfs_put_block_group(block_group);
5940         return ret;
5941 }
5942
5943 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5944                                  struct extent_buffer *eb)
5945 {
5946         struct btrfs_file_extent_item *item;
5947         struct btrfs_key key;
5948         int found_type;
5949         int i;
5950
5951         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5952                 return 0;
5953
5954         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5955                 btrfs_item_key_to_cpu(eb, &key, i);
5956                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5957                         continue;
5958                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5959                 found_type = btrfs_file_extent_type(eb, item);
5960                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5961                         continue;
5962                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5963                         continue;
5964                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5965                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5966                 __exclude_logged_extent(log, key.objectid, key.offset);
5967         }
5968
5969         return 0;
5970 }
5971
5972 /**
5973  * btrfs_update_reserved_bytes - update the block_group and space info counters
5974  * @cache:      The cache we are manipulating
5975  * @num_bytes:  The number of bytes in question
5976  * @reserve:    One of the reservation enums
5977  * @delalloc:   The blocks are allocated for the delalloc write
5978  *
5979  * This is called by the allocator when it reserves space, or by somebody who is
5980  * freeing space that was never actually used on disk.  For example if you
5981  * reserve some space for a new leaf in transaction A and before transaction A
5982  * commits you free that leaf, you call this with reserve set to 0 in order to
5983  * clear the reservation.
5984  *
5985  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5986  * ENOSPC accounting.  For data we handle the reservation through clearing the
5987  * delalloc bits in the io_tree.  We have to do this since we could end up
5988  * allocating less disk space for the amount of data we have reserved in the
5989  * case of compression.
5990  *
5991  * If this is a reservation and the block group has become read only we cannot
5992  * make the reservation and return -EAGAIN, otherwise this function always
5993  * succeeds.
5994  */
5995 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5996                                        u64 num_bytes, int reserve, int delalloc)
5997 {
5998         struct btrfs_space_info *space_info = cache->space_info;
5999         int ret = 0;
6000
6001         spin_lock(&space_info->lock);
6002         spin_lock(&cache->lock);
6003         if (reserve != RESERVE_FREE) {
6004                 if (cache->ro) {
6005                         ret = -EAGAIN;
6006                 } else {
6007                         cache->reserved += num_bytes;
6008                         space_info->bytes_reserved += num_bytes;
6009                         if (reserve == RESERVE_ALLOC) {
6010                                 trace_btrfs_space_reservation(cache->fs_info,
6011                                                 "space_info", space_info->flags,
6012                                                 num_bytes, 0);
6013                                 space_info->bytes_may_use -= num_bytes;
6014                         }
6015
6016                         if (delalloc)
6017                                 cache->delalloc_bytes += num_bytes;
6018                 }
6019         } else {
6020                 if (cache->ro)
6021                         space_info->bytes_readonly += num_bytes;
6022                 cache->reserved -= num_bytes;
6023                 space_info->bytes_reserved -= num_bytes;
6024
6025                 if (delalloc)
6026                         cache->delalloc_bytes -= num_bytes;
6027         }
6028         spin_unlock(&cache->lock);
6029         spin_unlock(&space_info->lock);
6030         return ret;
6031 }
6032
6033 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6034                                 struct btrfs_root *root)
6035 {
6036         struct btrfs_fs_info *fs_info = root->fs_info;
6037         struct btrfs_caching_control *next;
6038         struct btrfs_caching_control *caching_ctl;
6039         struct btrfs_block_group_cache *cache;
6040
6041         down_write(&fs_info->commit_root_sem);
6042
6043         list_for_each_entry_safe(caching_ctl, next,
6044                                  &fs_info->caching_block_groups, list) {
6045                 cache = caching_ctl->block_group;
6046                 if (block_group_cache_done(cache)) {
6047                         cache->last_byte_to_unpin = (u64)-1;
6048                         list_del_init(&caching_ctl->list);
6049                         put_caching_control(caching_ctl);
6050                 } else {
6051                         cache->last_byte_to_unpin = caching_ctl->progress;
6052                 }
6053         }
6054
6055         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6056                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6057         else
6058                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6059
6060         up_write(&fs_info->commit_root_sem);
6061
6062         update_global_block_rsv(fs_info);
6063 }
6064
6065 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
6066                               const bool return_free_space)
6067 {
6068         struct btrfs_fs_info *fs_info = root->fs_info;
6069         struct btrfs_block_group_cache *cache = NULL;
6070         struct btrfs_space_info *space_info;
6071         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6072         u64 len;
6073         bool readonly;
6074
6075         while (start <= end) {
6076                 readonly = false;
6077                 if (!cache ||
6078                     start >= cache->key.objectid + cache->key.offset) {
6079                         if (cache)
6080                                 btrfs_put_block_group(cache);
6081                         cache = btrfs_lookup_block_group(fs_info, start);
6082                         BUG_ON(!cache); /* Logic error */
6083                 }
6084
6085                 len = cache->key.objectid + cache->key.offset - start;
6086                 len = min(len, end + 1 - start);
6087
6088                 if (start < cache->last_byte_to_unpin) {
6089                         len = min(len, cache->last_byte_to_unpin - start);
6090                         if (return_free_space)
6091                                 btrfs_add_free_space(cache, start, len);
6092                 }
6093
6094                 start += len;
6095                 space_info = cache->space_info;
6096
6097                 spin_lock(&space_info->lock);
6098                 spin_lock(&cache->lock);
6099                 cache->pinned -= len;
6100                 space_info->bytes_pinned -= len;
6101                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6102                 if (cache->ro) {
6103                         space_info->bytes_readonly += len;
6104                         readonly = true;
6105                 }
6106                 spin_unlock(&cache->lock);
6107                 if (!readonly && global_rsv->space_info == space_info) {
6108                         spin_lock(&global_rsv->lock);
6109                         if (!global_rsv->full) {
6110                                 len = min(len, global_rsv->size -
6111                                           global_rsv->reserved);
6112                                 global_rsv->reserved += len;
6113                                 space_info->bytes_may_use += len;
6114                                 if (global_rsv->reserved >= global_rsv->size)
6115                                         global_rsv->full = 1;
6116                         }
6117                         spin_unlock(&global_rsv->lock);
6118                 }
6119                 spin_unlock(&space_info->lock);
6120         }
6121
6122         if (cache)
6123                 btrfs_put_block_group(cache);
6124         return 0;
6125 }
6126
6127 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6128                                struct btrfs_root *root)
6129 {
6130         struct btrfs_fs_info *fs_info = root->fs_info;
6131         struct btrfs_block_group_cache *block_group, *tmp;
6132         struct list_head *deleted_bgs;
6133         struct extent_io_tree *unpin;
6134         u64 start;
6135         u64 end;
6136         int ret;
6137
6138         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6139                 unpin = &fs_info->freed_extents[1];
6140         else
6141                 unpin = &fs_info->freed_extents[0];
6142
6143         while (!trans->aborted) {
6144                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6145                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6146                                             EXTENT_DIRTY, NULL);
6147                 if (ret) {
6148                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6149                         break;
6150                 }
6151
6152                 if (btrfs_test_opt(root, DISCARD))
6153                         ret = btrfs_discard_extent(root, start,
6154                                                    end + 1 - start, NULL);
6155
6156                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
6157                 unpin_extent_range(root, start, end, true);
6158                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6159                 cond_resched();
6160         }
6161
6162         /*
6163          * Transaction is finished.  We don't need the lock anymore.  We
6164          * do need to clean up the block groups in case of a transaction
6165          * abort.
6166          */
6167         deleted_bgs = &trans->transaction->deleted_bgs;
6168         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6169                 u64 trimmed = 0;
6170
6171                 ret = -EROFS;
6172                 if (!trans->aborted)
6173                         ret = btrfs_discard_extent(root,
6174                                                    block_group->key.objectid,
6175                                                    block_group->key.offset,
6176                                                    &trimmed);
6177
6178                 list_del_init(&block_group->bg_list);
6179                 btrfs_put_block_group_trimming(block_group);
6180                 btrfs_put_block_group(block_group);
6181
6182                 if (ret) {
6183                         const char *errstr = btrfs_decode_error(ret);
6184                         btrfs_warn(fs_info,
6185                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6186                                    ret, errstr);
6187                 }
6188         }
6189
6190         return 0;
6191 }
6192
6193 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6194                              u64 owner, u64 root_objectid)
6195 {
6196         struct btrfs_space_info *space_info;
6197         u64 flags;
6198
6199         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6200                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6201                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6202                 else
6203                         flags = BTRFS_BLOCK_GROUP_METADATA;
6204         } else {
6205                 flags = BTRFS_BLOCK_GROUP_DATA;
6206         }
6207
6208         space_info = __find_space_info(fs_info, flags);
6209         BUG_ON(!space_info); /* Logic bug */
6210         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6211 }
6212
6213
6214 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6215                                 struct btrfs_root *root,
6216                                 struct btrfs_delayed_ref_node *node, u64 parent,
6217                                 u64 root_objectid, u64 owner_objectid,
6218                                 u64 owner_offset, int refs_to_drop,
6219                                 struct btrfs_delayed_extent_op *extent_op)
6220 {
6221         struct btrfs_key key;
6222         struct btrfs_path *path;
6223         struct btrfs_fs_info *info = root->fs_info;
6224         struct btrfs_root *extent_root = info->extent_root;
6225         struct extent_buffer *leaf;
6226         struct btrfs_extent_item *ei;
6227         struct btrfs_extent_inline_ref *iref;
6228         int ret;
6229         int is_data;
6230         int extent_slot = 0;
6231         int found_extent = 0;
6232         int num_to_del = 1;
6233         int no_quota = node->no_quota;
6234         u32 item_size;
6235         u64 refs;
6236         u64 bytenr = node->bytenr;
6237         u64 num_bytes = node->num_bytes;
6238         int last_ref = 0;
6239         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6240                                                  SKINNY_METADATA);
6241
6242         if (!info->quota_enabled || !is_fstree(root_objectid))
6243                 no_quota = 1;
6244
6245         path = btrfs_alloc_path();
6246         if (!path)
6247                 return -ENOMEM;
6248
6249         path->reada = 1;
6250         path->leave_spinning = 1;
6251
6252         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6253         BUG_ON(!is_data && refs_to_drop != 1);
6254
6255         if (is_data)
6256                 skinny_metadata = 0;
6257
6258         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6259                                     bytenr, num_bytes, parent,
6260                                     root_objectid, owner_objectid,
6261                                     owner_offset);
6262         if (ret == 0) {
6263                 extent_slot = path->slots[0];
6264                 while (extent_slot >= 0) {
6265                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6266                                               extent_slot);
6267                         if (key.objectid != bytenr)
6268                                 break;
6269                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6270                             key.offset == num_bytes) {
6271                                 found_extent = 1;
6272                                 break;
6273                         }
6274                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6275                             key.offset == owner_objectid) {
6276                                 found_extent = 1;
6277                                 break;
6278                         }
6279                         if (path->slots[0] - extent_slot > 5)
6280                                 break;
6281                         extent_slot--;
6282                 }
6283 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6284                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6285                 if (found_extent && item_size < sizeof(*ei))
6286                         found_extent = 0;
6287 #endif
6288                 if (!found_extent) {
6289                         BUG_ON(iref);
6290                         ret = remove_extent_backref(trans, extent_root, path,
6291                                                     NULL, refs_to_drop,
6292                                                     is_data, &last_ref);
6293                         if (ret) {
6294                                 btrfs_abort_transaction(trans, extent_root, ret);
6295                                 goto out;
6296                         }
6297                         btrfs_release_path(path);
6298                         path->leave_spinning = 1;
6299
6300                         key.objectid = bytenr;
6301                         key.type = BTRFS_EXTENT_ITEM_KEY;
6302                         key.offset = num_bytes;
6303
6304                         if (!is_data && skinny_metadata) {
6305                                 key.type = BTRFS_METADATA_ITEM_KEY;
6306                                 key.offset = owner_objectid;
6307                         }
6308
6309                         ret = btrfs_search_slot(trans, extent_root,
6310                                                 &key, path, -1, 1);
6311                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6312                                 /*
6313                                  * Couldn't find our skinny metadata item,
6314                                  * see if we have ye olde extent item.
6315                                  */
6316                                 path->slots[0]--;
6317                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6318                                                       path->slots[0]);
6319                                 if (key.objectid == bytenr &&
6320                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6321                                     key.offset == num_bytes)
6322                                         ret = 0;
6323                         }
6324
6325                         if (ret > 0 && skinny_metadata) {
6326                                 skinny_metadata = false;
6327                                 key.objectid = bytenr;
6328                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6329                                 key.offset = num_bytes;
6330                                 btrfs_release_path(path);
6331                                 ret = btrfs_search_slot(trans, extent_root,
6332                                                         &key, path, -1, 1);
6333                         }
6334
6335                         if (ret) {
6336                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6337                                         ret, bytenr);
6338                                 if (ret > 0)
6339                                         btrfs_print_leaf(extent_root,
6340                                                          path->nodes[0]);
6341                         }
6342                         if (ret < 0) {
6343                                 btrfs_abort_transaction(trans, extent_root, ret);
6344                                 goto out;
6345                         }
6346                         extent_slot = path->slots[0];
6347                 }
6348         } else if (WARN_ON(ret == -ENOENT)) {
6349                 btrfs_print_leaf(extent_root, path->nodes[0]);
6350                 btrfs_err(info,
6351                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6352                         bytenr, parent, root_objectid, owner_objectid,
6353                         owner_offset);
6354                 btrfs_abort_transaction(trans, extent_root, ret);
6355                 goto out;
6356         } else {
6357                 btrfs_abort_transaction(trans, extent_root, ret);
6358                 goto out;
6359         }
6360
6361         leaf = path->nodes[0];
6362         item_size = btrfs_item_size_nr(leaf, extent_slot);
6363 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6364         if (item_size < sizeof(*ei)) {
6365                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6366                 ret = convert_extent_item_v0(trans, extent_root, path,
6367                                              owner_objectid, 0);
6368                 if (ret < 0) {
6369                         btrfs_abort_transaction(trans, extent_root, ret);
6370                         goto out;
6371                 }
6372
6373                 btrfs_release_path(path);
6374                 path->leave_spinning = 1;
6375
6376                 key.objectid = bytenr;
6377                 key.type = BTRFS_EXTENT_ITEM_KEY;
6378                 key.offset = num_bytes;
6379
6380                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6381                                         -1, 1);
6382                 if (ret) {
6383                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6384                                 ret, bytenr);
6385                         btrfs_print_leaf(extent_root, path->nodes[0]);
6386                 }
6387                 if (ret < 0) {
6388                         btrfs_abort_transaction(trans, extent_root, ret);
6389                         goto out;
6390                 }
6391
6392                 extent_slot = path->slots[0];
6393                 leaf = path->nodes[0];
6394                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6395         }
6396 #endif
6397         BUG_ON(item_size < sizeof(*ei));
6398         ei = btrfs_item_ptr(leaf, extent_slot,
6399                             struct btrfs_extent_item);
6400         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6401             key.type == BTRFS_EXTENT_ITEM_KEY) {
6402                 struct btrfs_tree_block_info *bi;
6403                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6404                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6405                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6406         }
6407
6408         refs = btrfs_extent_refs(leaf, ei);
6409         if (refs < refs_to_drop) {
6410                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6411                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6412                 ret = -EINVAL;
6413                 btrfs_abort_transaction(trans, extent_root, ret);
6414                 goto out;
6415         }
6416         refs -= refs_to_drop;
6417
6418         if (refs > 0) {
6419                 if (extent_op)
6420                         __run_delayed_extent_op(extent_op, leaf, ei);
6421                 /*
6422                  * In the case of inline back ref, reference count will
6423                  * be updated by remove_extent_backref
6424                  */
6425                 if (iref) {
6426                         BUG_ON(!found_extent);
6427                 } else {
6428                         btrfs_set_extent_refs(leaf, ei, refs);
6429                         btrfs_mark_buffer_dirty(leaf);
6430                 }
6431                 if (found_extent) {
6432                         ret = remove_extent_backref(trans, extent_root, path,
6433                                                     iref, refs_to_drop,
6434                                                     is_data, &last_ref);
6435                         if (ret) {
6436                                 btrfs_abort_transaction(trans, extent_root, ret);
6437                                 goto out;
6438                         }
6439                 }
6440                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6441                                  root_objectid);
6442         } else {
6443                 if (found_extent) {
6444                         BUG_ON(is_data && refs_to_drop !=
6445                                extent_data_ref_count(path, iref));
6446                         if (iref) {
6447                                 BUG_ON(path->slots[0] != extent_slot);
6448                         } else {
6449                                 BUG_ON(path->slots[0] != extent_slot + 1);
6450                                 path->slots[0] = extent_slot;
6451                                 num_to_del = 2;
6452                         }
6453                 }
6454
6455                 last_ref = 1;
6456                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6457                                       num_to_del);
6458                 if (ret) {
6459                         btrfs_abort_transaction(trans, extent_root, ret);
6460                         goto out;
6461                 }
6462                 btrfs_release_path(path);
6463
6464                 if (is_data) {
6465                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6466                         if (ret) {
6467                                 btrfs_abort_transaction(trans, extent_root, ret);
6468                                 goto out;
6469                         }
6470                 }
6471
6472                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6473                 if (ret) {
6474                         btrfs_abort_transaction(trans, extent_root, ret);
6475                         goto out;
6476                 }
6477         }
6478         btrfs_release_path(path);
6479
6480 out:
6481         btrfs_free_path(path);
6482         return ret;
6483 }
6484
6485 /*
6486  * when we free an block, it is possible (and likely) that we free the last
6487  * delayed ref for that extent as well.  This searches the delayed ref tree for
6488  * a given extent, and if there are no other delayed refs to be processed, it
6489  * removes it from the tree.
6490  */
6491 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6492                                       struct btrfs_root *root, u64 bytenr)
6493 {
6494         struct btrfs_delayed_ref_head *head;
6495         struct btrfs_delayed_ref_root *delayed_refs;
6496         int ret = 0;
6497
6498         delayed_refs = &trans->transaction->delayed_refs;
6499         spin_lock(&delayed_refs->lock);
6500         head = btrfs_find_delayed_ref_head(trans, bytenr);
6501         if (!head)
6502                 goto out_delayed_unlock;
6503
6504         spin_lock(&head->lock);
6505         if (!list_empty(&head->ref_list))
6506                 goto out;
6507
6508         if (head->extent_op) {
6509                 if (!head->must_insert_reserved)
6510                         goto out;
6511                 btrfs_free_delayed_extent_op(head->extent_op);
6512                 head->extent_op = NULL;
6513         }
6514
6515         /*
6516          * waiting for the lock here would deadlock.  If someone else has it
6517          * locked they are already in the process of dropping it anyway
6518          */
6519         if (!mutex_trylock(&head->mutex))
6520                 goto out;
6521
6522         /*
6523          * at this point we have a head with no other entries.  Go
6524          * ahead and process it.
6525          */
6526         head->node.in_tree = 0;
6527         rb_erase(&head->href_node, &delayed_refs->href_root);
6528
6529         atomic_dec(&delayed_refs->num_entries);
6530
6531         /*
6532          * we don't take a ref on the node because we're removing it from the
6533          * tree, so we just steal the ref the tree was holding.
6534          */
6535         delayed_refs->num_heads--;
6536         if (head->processing == 0)
6537                 delayed_refs->num_heads_ready--;
6538         head->processing = 0;
6539         spin_unlock(&head->lock);
6540         spin_unlock(&delayed_refs->lock);
6541
6542         BUG_ON(head->extent_op);
6543         if (head->must_insert_reserved)
6544                 ret = 1;
6545
6546         mutex_unlock(&head->mutex);
6547         btrfs_put_delayed_ref(&head->node);
6548         return ret;
6549 out:
6550         spin_unlock(&head->lock);
6551
6552 out_delayed_unlock:
6553         spin_unlock(&delayed_refs->lock);
6554         return 0;
6555 }
6556
6557 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6558                            struct btrfs_root *root,
6559                            struct extent_buffer *buf,
6560                            u64 parent, int last_ref)
6561 {
6562         int pin = 1;
6563         int ret;
6564
6565         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6566                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6567                                         buf->start, buf->len,
6568                                         parent, root->root_key.objectid,
6569                                         btrfs_header_level(buf),
6570                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6571                 BUG_ON(ret); /* -ENOMEM */
6572         }
6573
6574         if (!last_ref)
6575                 return;
6576
6577         if (btrfs_header_generation(buf) == trans->transid) {
6578                 struct btrfs_block_group_cache *cache;
6579
6580                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6581                         ret = check_ref_cleanup(trans, root, buf->start);
6582                         if (!ret)
6583                                 goto out;
6584                 }
6585
6586                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6587
6588                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6589                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6590                         btrfs_put_block_group(cache);
6591                         goto out;
6592                 }
6593
6594                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6595
6596                 btrfs_add_free_space(cache, buf->start, buf->len);
6597                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6598                 btrfs_put_block_group(cache);
6599                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6600                 pin = 0;
6601         }
6602 out:
6603         if (pin)
6604                 add_pinned_bytes(root->fs_info, buf->len,
6605                                  btrfs_header_level(buf),
6606                                  root->root_key.objectid);
6607
6608         /*
6609          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6610          * anymore.
6611          */
6612         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6613 }
6614
6615 /* Can return -ENOMEM */
6616 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6617                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6618                       u64 owner, u64 offset, int no_quota)
6619 {
6620         int ret;
6621         struct btrfs_fs_info *fs_info = root->fs_info;
6622
6623         if (btrfs_test_is_dummy_root(root))
6624                 return 0;
6625
6626         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6627
6628         /*
6629          * tree log blocks never actually go into the extent allocation
6630          * tree, just update pinning info and exit early.
6631          */
6632         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6633                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6634                 /* unlocks the pinned mutex */
6635                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6636                 ret = 0;
6637         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6638                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6639                                         num_bytes,
6640                                         parent, root_objectid, (int)owner,
6641                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6642         } else {
6643                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6644                                                 num_bytes,
6645                                                 parent, root_objectid, owner,
6646                                                 offset, BTRFS_DROP_DELAYED_REF,
6647                                                 NULL, no_quota);
6648         }
6649         return ret;
6650 }
6651
6652 /*
6653  * when we wait for progress in the block group caching, its because
6654  * our allocation attempt failed at least once.  So, we must sleep
6655  * and let some progress happen before we try again.
6656  *
6657  * This function will sleep at least once waiting for new free space to
6658  * show up, and then it will check the block group free space numbers
6659  * for our min num_bytes.  Another option is to have it go ahead
6660  * and look in the rbtree for a free extent of a given size, but this
6661  * is a good start.
6662  *
6663  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6664  * any of the information in this block group.
6665  */
6666 static noinline void
6667 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6668                                 u64 num_bytes)
6669 {
6670         struct btrfs_caching_control *caching_ctl;
6671
6672         caching_ctl = get_caching_control(cache);
6673         if (!caching_ctl)
6674                 return;
6675
6676         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6677                    (cache->free_space_ctl->free_space >= num_bytes));
6678
6679         put_caching_control(caching_ctl);
6680 }
6681
6682 static noinline int
6683 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6684 {
6685         struct btrfs_caching_control *caching_ctl;
6686         int ret = 0;
6687
6688         caching_ctl = get_caching_control(cache);
6689         if (!caching_ctl)
6690                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6691
6692         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6693         if (cache->cached == BTRFS_CACHE_ERROR)
6694                 ret = -EIO;
6695         put_caching_control(caching_ctl);
6696         return ret;
6697 }
6698
6699 int __get_raid_index(u64 flags)
6700 {
6701         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6702                 return BTRFS_RAID_RAID10;
6703         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6704                 return BTRFS_RAID_RAID1;
6705         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6706                 return BTRFS_RAID_DUP;
6707         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6708                 return BTRFS_RAID_RAID0;
6709         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6710                 return BTRFS_RAID_RAID5;
6711         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6712                 return BTRFS_RAID_RAID6;
6713
6714         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6715 }
6716
6717 int get_block_group_index(struct btrfs_block_group_cache *cache)
6718 {
6719         return __get_raid_index(cache->flags);
6720 }
6721
6722 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6723         [BTRFS_RAID_RAID10]     = "raid10",
6724         [BTRFS_RAID_RAID1]      = "raid1",
6725         [BTRFS_RAID_DUP]        = "dup",
6726         [BTRFS_RAID_RAID0]      = "raid0",
6727         [BTRFS_RAID_SINGLE]     = "single",
6728         [BTRFS_RAID_RAID5]      = "raid5",
6729         [BTRFS_RAID_RAID6]      = "raid6",
6730 };
6731
6732 static const char *get_raid_name(enum btrfs_raid_types type)
6733 {
6734         if (type >= BTRFS_NR_RAID_TYPES)
6735                 return NULL;
6736
6737         return btrfs_raid_type_names[type];
6738 }
6739
6740 enum btrfs_loop_type {
6741         LOOP_CACHING_NOWAIT = 0,
6742         LOOP_CACHING_WAIT = 1,
6743         LOOP_ALLOC_CHUNK = 2,
6744         LOOP_NO_EMPTY_SIZE = 3,
6745 };
6746
6747 static inline void
6748 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6749                        int delalloc)
6750 {
6751         if (delalloc)
6752                 down_read(&cache->data_rwsem);
6753 }
6754
6755 static inline void
6756 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6757                        int delalloc)
6758 {
6759         btrfs_get_block_group(cache);
6760         if (delalloc)
6761                 down_read(&cache->data_rwsem);
6762 }
6763
6764 static struct btrfs_block_group_cache *
6765 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6766                    struct btrfs_free_cluster *cluster,
6767                    int delalloc)
6768 {
6769         struct btrfs_block_group_cache *used_bg;
6770         bool locked = false;
6771 again:
6772         spin_lock(&cluster->refill_lock);
6773         if (locked) {
6774                 if (used_bg == cluster->block_group)
6775                         return used_bg;
6776
6777                 up_read(&used_bg->data_rwsem);
6778                 btrfs_put_block_group(used_bg);
6779         }
6780
6781         used_bg = cluster->block_group;
6782         if (!used_bg)
6783                 return NULL;
6784
6785         if (used_bg == block_group)
6786                 return used_bg;
6787
6788         btrfs_get_block_group(used_bg);
6789
6790         if (!delalloc)
6791                 return used_bg;
6792
6793         if (down_read_trylock(&used_bg->data_rwsem))
6794                 return used_bg;
6795
6796         spin_unlock(&cluster->refill_lock);
6797         down_read(&used_bg->data_rwsem);
6798         locked = true;
6799         goto again;
6800 }
6801
6802 static inline void
6803 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6804                          int delalloc)
6805 {
6806         if (delalloc)
6807                 up_read(&cache->data_rwsem);
6808         btrfs_put_block_group(cache);
6809 }
6810
6811 /*
6812  * walks the btree of allocated extents and find a hole of a given size.
6813  * The key ins is changed to record the hole:
6814  * ins->objectid == start position
6815  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6816  * ins->offset == the size of the hole.
6817  * Any available blocks before search_start are skipped.
6818  *
6819  * If there is no suitable free space, we will record the max size of
6820  * the free space extent currently.
6821  */
6822 static noinline int find_free_extent(struct btrfs_root *orig_root,
6823                                      u64 num_bytes, u64 empty_size,
6824                                      u64 hint_byte, struct btrfs_key *ins,
6825                                      u64 flags, int delalloc)
6826 {
6827         int ret = 0;
6828         struct btrfs_root *root = orig_root->fs_info->extent_root;
6829         struct btrfs_free_cluster *last_ptr = NULL;
6830         struct btrfs_block_group_cache *block_group = NULL;
6831         u64 search_start = 0;
6832         u64 max_extent_size = 0;
6833         int empty_cluster = 2 * 1024 * 1024;
6834         struct btrfs_space_info *space_info;
6835         int loop = 0;
6836         int index = __get_raid_index(flags);
6837         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6838                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6839         bool failed_cluster_refill = false;
6840         bool failed_alloc = false;
6841         bool use_cluster = true;
6842         bool have_caching_bg = false;
6843
6844         WARN_ON(num_bytes < root->sectorsize);
6845         ins->type = BTRFS_EXTENT_ITEM_KEY;
6846         ins->objectid = 0;
6847         ins->offset = 0;
6848
6849         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6850
6851         space_info = __find_space_info(root->fs_info, flags);
6852         if (!space_info) {
6853                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6854                 return -ENOSPC;
6855         }
6856
6857         /*
6858          * If the space info is for both data and metadata it means we have a
6859          * small filesystem and we can't use the clustering stuff.
6860          */
6861         if (btrfs_mixed_space_info(space_info))
6862                 use_cluster = false;
6863
6864         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6865                 last_ptr = &root->fs_info->meta_alloc_cluster;
6866                 if (!btrfs_test_opt(root, SSD))
6867                         empty_cluster = 64 * 1024;
6868         }
6869
6870         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6871             btrfs_test_opt(root, SSD)) {
6872                 last_ptr = &root->fs_info->data_alloc_cluster;
6873         }
6874
6875         if (last_ptr) {
6876                 spin_lock(&last_ptr->lock);
6877                 if (last_ptr->block_group)
6878                         hint_byte = last_ptr->window_start;
6879                 spin_unlock(&last_ptr->lock);
6880         }
6881
6882         search_start = max(search_start, first_logical_byte(root, 0));
6883         search_start = max(search_start, hint_byte);
6884
6885         if (!last_ptr)
6886                 empty_cluster = 0;
6887
6888         if (search_start == hint_byte) {
6889                 block_group = btrfs_lookup_block_group(root->fs_info,
6890                                                        search_start);
6891                 /*
6892                  * we don't want to use the block group if it doesn't match our
6893                  * allocation bits, or if its not cached.
6894                  *
6895                  * However if we are re-searching with an ideal block group
6896                  * picked out then we don't care that the block group is cached.
6897                  */
6898                 if (block_group && block_group_bits(block_group, flags) &&
6899                     block_group->cached != BTRFS_CACHE_NO) {
6900                         down_read(&space_info->groups_sem);
6901                         if (list_empty(&block_group->list) ||
6902                             block_group->ro) {
6903                                 /*
6904                                  * someone is removing this block group,
6905                                  * we can't jump into the have_block_group
6906                                  * target because our list pointers are not
6907                                  * valid
6908                                  */
6909                                 btrfs_put_block_group(block_group);
6910                                 up_read(&space_info->groups_sem);
6911                         } else {
6912                                 index = get_block_group_index(block_group);
6913                                 btrfs_lock_block_group(block_group, delalloc);
6914                                 goto have_block_group;
6915                         }
6916                 } else if (block_group) {
6917                         btrfs_put_block_group(block_group);
6918                 }
6919         }
6920 search:
6921         have_caching_bg = false;
6922         down_read(&space_info->groups_sem);
6923         list_for_each_entry(block_group, &space_info->block_groups[index],
6924                             list) {
6925                 u64 offset;
6926                 int cached;
6927
6928                 btrfs_grab_block_group(block_group, delalloc);
6929                 search_start = block_group->key.objectid;
6930
6931                 /*
6932                  * this can happen if we end up cycling through all the
6933                  * raid types, but we want to make sure we only allocate
6934                  * for the proper type.
6935                  */
6936                 if (!block_group_bits(block_group, flags)) {
6937                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6938                                 BTRFS_BLOCK_GROUP_RAID1 |
6939                                 BTRFS_BLOCK_GROUP_RAID5 |
6940                                 BTRFS_BLOCK_GROUP_RAID6 |
6941                                 BTRFS_BLOCK_GROUP_RAID10;
6942
6943                         /*
6944                          * if they asked for extra copies and this block group
6945                          * doesn't provide them, bail.  This does allow us to
6946                          * fill raid0 from raid1.
6947                          */
6948                         if ((flags & extra) && !(block_group->flags & extra))
6949                                 goto loop;
6950                 }
6951
6952 have_block_group:
6953                 cached = block_group_cache_done(block_group);
6954                 if (unlikely(!cached)) {
6955                         ret = cache_block_group(block_group, 0);
6956                         BUG_ON(ret < 0);
6957                         ret = 0;
6958                 }
6959
6960                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6961                         goto loop;
6962                 if (unlikely(block_group->ro))
6963                         goto loop;
6964
6965                 /*
6966                  * Ok we want to try and use the cluster allocator, so
6967                  * lets look there
6968                  */
6969                 if (last_ptr) {
6970                         struct btrfs_block_group_cache *used_block_group;
6971                         unsigned long aligned_cluster;
6972                         /*
6973                          * the refill lock keeps out other
6974                          * people trying to start a new cluster
6975                          */
6976                         used_block_group = btrfs_lock_cluster(block_group,
6977                                                               last_ptr,
6978                                                               delalloc);
6979                         if (!used_block_group)
6980                                 goto refill_cluster;
6981
6982                         if (used_block_group != block_group &&
6983                             (used_block_group->ro ||
6984                              !block_group_bits(used_block_group, flags)))
6985                                 goto release_cluster;
6986
6987                         offset = btrfs_alloc_from_cluster(used_block_group,
6988                                                 last_ptr,
6989                                                 num_bytes,
6990                                                 used_block_group->key.objectid,
6991                                                 &max_extent_size);
6992                         if (offset) {
6993                                 /* we have a block, we're done */
6994                                 spin_unlock(&last_ptr->refill_lock);
6995                                 trace_btrfs_reserve_extent_cluster(root,
6996                                                 used_block_group,
6997                                                 search_start, num_bytes);
6998                                 if (used_block_group != block_group) {
6999                                         btrfs_release_block_group(block_group,
7000                                                                   delalloc);
7001                                         block_group = used_block_group;
7002                                 }
7003                                 goto checks;
7004                         }
7005
7006                         WARN_ON(last_ptr->block_group != used_block_group);
7007 release_cluster:
7008                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7009                          * set up a new clusters, so lets just skip it
7010                          * and let the allocator find whatever block
7011                          * it can find.  If we reach this point, we
7012                          * will have tried the cluster allocator
7013                          * plenty of times and not have found
7014                          * anything, so we are likely way too
7015                          * fragmented for the clustering stuff to find
7016                          * anything.
7017                          *
7018                          * However, if the cluster is taken from the
7019                          * current block group, release the cluster
7020                          * first, so that we stand a better chance of
7021                          * succeeding in the unclustered
7022                          * allocation.  */
7023                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7024                             used_block_group != block_group) {
7025                                 spin_unlock(&last_ptr->refill_lock);
7026                                 btrfs_release_block_group(used_block_group,
7027                                                           delalloc);
7028                                 goto unclustered_alloc;
7029                         }
7030
7031                         /*
7032                          * this cluster didn't work out, free it and
7033                          * start over
7034                          */
7035                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7036
7037                         if (used_block_group != block_group)
7038                                 btrfs_release_block_group(used_block_group,
7039                                                           delalloc);
7040 refill_cluster:
7041                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7042                                 spin_unlock(&last_ptr->refill_lock);
7043                                 goto unclustered_alloc;
7044                         }
7045
7046                         aligned_cluster = max_t(unsigned long,
7047                                                 empty_cluster + empty_size,
7048                                               block_group->full_stripe_len);
7049
7050                         /* allocate a cluster in this block group */
7051                         ret = btrfs_find_space_cluster(root, block_group,
7052                                                        last_ptr, search_start,
7053                                                        num_bytes,
7054                                                        aligned_cluster);
7055                         if (ret == 0) {
7056                                 /*
7057                                  * now pull our allocation out of this
7058                                  * cluster
7059                                  */
7060                                 offset = btrfs_alloc_from_cluster(block_group,
7061                                                         last_ptr,
7062                                                         num_bytes,
7063                                                         search_start,
7064                                                         &max_extent_size);
7065                                 if (offset) {
7066                                         /* we found one, proceed */
7067                                         spin_unlock(&last_ptr->refill_lock);
7068                                         trace_btrfs_reserve_extent_cluster(root,
7069                                                 block_group, search_start,
7070                                                 num_bytes);
7071                                         goto checks;
7072                                 }
7073                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7074                                    && !failed_cluster_refill) {
7075                                 spin_unlock(&last_ptr->refill_lock);
7076
7077                                 failed_cluster_refill = true;
7078                                 wait_block_group_cache_progress(block_group,
7079                                        num_bytes + empty_cluster + empty_size);
7080                                 goto have_block_group;
7081                         }
7082
7083                         /*
7084                          * at this point we either didn't find a cluster
7085                          * or we weren't able to allocate a block from our
7086                          * cluster.  Free the cluster we've been trying
7087                          * to use, and go to the next block group
7088                          */
7089                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7090                         spin_unlock(&last_ptr->refill_lock);
7091                         goto loop;
7092                 }
7093
7094 unclustered_alloc:
7095                 spin_lock(&block_group->free_space_ctl->tree_lock);
7096                 if (cached &&
7097                     block_group->free_space_ctl->free_space <
7098                     num_bytes + empty_cluster + empty_size) {
7099                         if (block_group->free_space_ctl->free_space >
7100                             max_extent_size)
7101                                 max_extent_size =
7102                                         block_group->free_space_ctl->free_space;
7103                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7104                         goto loop;
7105                 }
7106                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7107
7108                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7109                                                     num_bytes, empty_size,
7110                                                     &max_extent_size);
7111                 /*
7112                  * If we didn't find a chunk, and we haven't failed on this
7113                  * block group before, and this block group is in the middle of
7114                  * caching and we are ok with waiting, then go ahead and wait
7115                  * for progress to be made, and set failed_alloc to true.
7116                  *
7117                  * If failed_alloc is true then we've already waited on this
7118                  * block group once and should move on to the next block group.
7119                  */
7120                 if (!offset && !failed_alloc && !cached &&
7121                     loop > LOOP_CACHING_NOWAIT) {
7122                         wait_block_group_cache_progress(block_group,
7123                                                 num_bytes + empty_size);
7124                         failed_alloc = true;
7125                         goto have_block_group;
7126                 } else if (!offset) {
7127                         if (!cached)
7128                                 have_caching_bg = true;
7129                         goto loop;
7130                 }
7131 checks:
7132                 search_start = ALIGN(offset, root->stripesize);
7133
7134                 /* move on to the next group */
7135                 if (search_start + num_bytes >
7136                     block_group->key.objectid + block_group->key.offset) {
7137                         btrfs_add_free_space(block_group, offset, num_bytes);
7138                         goto loop;
7139                 }
7140
7141                 if (offset < search_start)
7142                         btrfs_add_free_space(block_group, offset,
7143                                              search_start - offset);
7144                 BUG_ON(offset > search_start);
7145
7146                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7147                                                   alloc_type, delalloc);
7148                 if (ret == -EAGAIN) {
7149                         btrfs_add_free_space(block_group, offset, num_bytes);
7150                         goto loop;
7151                 }
7152
7153                 /* we are all good, lets return */
7154                 ins->objectid = search_start;
7155                 ins->offset = num_bytes;
7156
7157                 trace_btrfs_reserve_extent(orig_root, block_group,
7158                                            search_start, num_bytes);
7159                 btrfs_release_block_group(block_group, delalloc);
7160                 break;
7161 loop:
7162                 failed_cluster_refill = false;
7163                 failed_alloc = false;
7164                 BUG_ON(index != get_block_group_index(block_group));
7165                 btrfs_release_block_group(block_group, delalloc);
7166         }
7167         up_read(&space_info->groups_sem);
7168
7169         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7170                 goto search;
7171
7172         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7173                 goto search;
7174
7175         /*
7176          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7177          *                      caching kthreads as we move along
7178          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7179          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7180          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7181          *                      again
7182          */
7183         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7184                 index = 0;
7185                 loop++;
7186                 if (loop == LOOP_ALLOC_CHUNK) {
7187                         struct btrfs_trans_handle *trans;
7188                         int exist = 0;
7189
7190                         trans = current->journal_info;
7191                         if (trans)
7192                                 exist = 1;
7193                         else
7194                                 trans = btrfs_join_transaction(root);
7195
7196                         if (IS_ERR(trans)) {
7197                                 ret = PTR_ERR(trans);
7198                                 goto out;
7199                         }
7200
7201                         ret = do_chunk_alloc(trans, root, flags,
7202                                              CHUNK_ALLOC_FORCE);
7203                         /*
7204                          * Do not bail out on ENOSPC since we
7205                          * can do more things.
7206                          */
7207                         if (ret < 0 && ret != -ENOSPC)
7208                                 btrfs_abort_transaction(trans,
7209                                                         root, ret);
7210                         else
7211                                 ret = 0;
7212                         if (!exist)
7213                                 btrfs_end_transaction(trans, root);
7214                         if (ret)
7215                                 goto out;
7216                 }
7217
7218                 if (loop == LOOP_NO_EMPTY_SIZE) {
7219                         empty_size = 0;
7220                         empty_cluster = 0;
7221                 }
7222
7223                 goto search;
7224         } else if (!ins->objectid) {
7225                 ret = -ENOSPC;
7226         } else if (ins->objectid) {
7227                 ret = 0;
7228         }
7229 out:
7230         if (ret == -ENOSPC)
7231                 ins->offset = max_extent_size;
7232         return ret;
7233 }
7234
7235 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7236                             int dump_block_groups)
7237 {
7238         struct btrfs_block_group_cache *cache;
7239         int index = 0;
7240
7241         spin_lock(&info->lock);
7242         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7243                info->flags,
7244                info->total_bytes - info->bytes_used - info->bytes_pinned -
7245                info->bytes_reserved - info->bytes_readonly,
7246                (info->full) ? "" : "not ");
7247         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7248                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7249                info->total_bytes, info->bytes_used, info->bytes_pinned,
7250                info->bytes_reserved, info->bytes_may_use,
7251                info->bytes_readonly);
7252         spin_unlock(&info->lock);
7253
7254         if (!dump_block_groups)
7255                 return;
7256
7257         down_read(&info->groups_sem);
7258 again:
7259         list_for_each_entry(cache, &info->block_groups[index], list) {
7260                 spin_lock(&cache->lock);
7261                 printk(KERN_INFO "BTRFS: "
7262                            "block group %llu has %llu bytes, "
7263                            "%llu used %llu pinned %llu reserved %s\n",
7264                        cache->key.objectid, cache->key.offset,
7265                        btrfs_block_group_used(&cache->item), cache->pinned,
7266                        cache->reserved, cache->ro ? "[readonly]" : "");
7267                 btrfs_dump_free_space(cache, bytes);
7268                 spin_unlock(&cache->lock);
7269         }
7270         if (++index < BTRFS_NR_RAID_TYPES)
7271                 goto again;
7272         up_read(&info->groups_sem);
7273 }
7274
7275 int btrfs_reserve_extent(struct btrfs_root *root,
7276                          u64 num_bytes, u64 min_alloc_size,
7277                          u64 empty_size, u64 hint_byte,
7278                          struct btrfs_key *ins, int is_data, int delalloc)
7279 {
7280         bool final_tried = false;
7281         u64 flags;
7282         int ret;
7283
7284         flags = btrfs_get_alloc_profile(root, is_data);
7285 again:
7286         WARN_ON(num_bytes < root->sectorsize);
7287         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7288                                flags, delalloc);
7289
7290         if (ret == -ENOSPC) {
7291                 if (!final_tried && ins->offset) {
7292                         num_bytes = min(num_bytes >> 1, ins->offset);
7293                         num_bytes = round_down(num_bytes, root->sectorsize);
7294                         num_bytes = max(num_bytes, min_alloc_size);
7295                         if (num_bytes == min_alloc_size)
7296                                 final_tried = true;
7297                         goto again;
7298                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7299                         struct btrfs_space_info *sinfo;
7300
7301                         sinfo = __find_space_info(root->fs_info, flags);
7302                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7303                                 flags, num_bytes);
7304                         if (sinfo)
7305                                 dump_space_info(sinfo, num_bytes, 1);
7306                 }
7307         }
7308
7309         return ret;
7310 }
7311
7312 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7313                                         u64 start, u64 len,
7314                                         int pin, int delalloc)
7315 {
7316         struct btrfs_block_group_cache *cache;
7317         int ret = 0;
7318
7319         cache = btrfs_lookup_block_group(root->fs_info, start);
7320         if (!cache) {
7321                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7322                         start);
7323                 return -ENOSPC;
7324         }
7325
7326         if (pin)
7327                 pin_down_extent(root, cache, start, len, 1);
7328         else {
7329                 if (btrfs_test_opt(root, DISCARD))
7330                         ret = btrfs_discard_extent(root, start, len, NULL);
7331                 btrfs_add_free_space(cache, start, len);
7332                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7333         }
7334
7335         btrfs_put_block_group(cache);
7336
7337         trace_btrfs_reserved_extent_free(root, start, len);
7338
7339         return ret;
7340 }
7341
7342 int btrfs_free_reserved_extent(struct btrfs_root *root,
7343                                u64 start, u64 len, int delalloc)
7344 {
7345         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7346 }
7347
7348 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7349                                        u64 start, u64 len)
7350 {
7351         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7352 }
7353
7354 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7355                                       struct btrfs_root *root,
7356                                       u64 parent, u64 root_objectid,
7357                                       u64 flags, u64 owner, u64 offset,
7358                                       struct btrfs_key *ins, int ref_mod)
7359 {
7360         int ret;
7361         struct btrfs_fs_info *fs_info = root->fs_info;
7362         struct btrfs_extent_item *extent_item;
7363         struct btrfs_extent_inline_ref *iref;
7364         struct btrfs_path *path;
7365         struct extent_buffer *leaf;
7366         int type;
7367         u32 size;
7368
7369         if (parent > 0)
7370                 type = BTRFS_SHARED_DATA_REF_KEY;
7371         else
7372                 type = BTRFS_EXTENT_DATA_REF_KEY;
7373
7374         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7375
7376         path = btrfs_alloc_path();
7377         if (!path)
7378                 return -ENOMEM;
7379
7380         path->leave_spinning = 1;
7381         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7382                                       ins, size);
7383         if (ret) {
7384                 btrfs_free_path(path);
7385                 return ret;
7386         }
7387
7388         leaf = path->nodes[0];
7389         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7390                                      struct btrfs_extent_item);
7391         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7392         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7393         btrfs_set_extent_flags(leaf, extent_item,
7394                                flags | BTRFS_EXTENT_FLAG_DATA);
7395
7396         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7397         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7398         if (parent > 0) {
7399                 struct btrfs_shared_data_ref *ref;
7400                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7401                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7402                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7403         } else {
7404                 struct btrfs_extent_data_ref *ref;
7405                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7406                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7407                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7408                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7409                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7410         }
7411
7412         btrfs_mark_buffer_dirty(path->nodes[0]);
7413         btrfs_free_path(path);
7414
7415         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7416         if (ret) { /* -ENOENT, logic error */
7417                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7418                         ins->objectid, ins->offset);
7419                 BUG();
7420         }
7421         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7422         return ret;
7423 }
7424
7425 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7426                                      struct btrfs_root *root,
7427                                      u64 parent, u64 root_objectid,
7428                                      u64 flags, struct btrfs_disk_key *key,
7429                                      int level, struct btrfs_key *ins,
7430                                      int no_quota)
7431 {
7432         int ret;
7433         struct btrfs_fs_info *fs_info = root->fs_info;
7434         struct btrfs_extent_item *extent_item;
7435         struct btrfs_tree_block_info *block_info;
7436         struct btrfs_extent_inline_ref *iref;
7437         struct btrfs_path *path;
7438         struct extent_buffer *leaf;
7439         u32 size = sizeof(*extent_item) + sizeof(*iref);
7440         u64 num_bytes = ins->offset;
7441         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7442                                                  SKINNY_METADATA);
7443
7444         if (!skinny_metadata)
7445                 size += sizeof(*block_info);
7446
7447         path = btrfs_alloc_path();
7448         if (!path) {
7449                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7450                                                    root->nodesize);
7451                 return -ENOMEM;
7452         }
7453
7454         path->leave_spinning = 1;
7455         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7456                                       ins, size);
7457         if (ret) {
7458                 btrfs_free_path(path);
7459                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7460                                                    root->nodesize);
7461                 return ret;
7462         }
7463
7464         leaf = path->nodes[0];
7465         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7466                                      struct btrfs_extent_item);
7467         btrfs_set_extent_refs(leaf, extent_item, 1);
7468         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7469         btrfs_set_extent_flags(leaf, extent_item,
7470                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7471
7472         if (skinny_metadata) {
7473                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7474                 num_bytes = root->nodesize;
7475         } else {
7476                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7477                 btrfs_set_tree_block_key(leaf, block_info, key);
7478                 btrfs_set_tree_block_level(leaf, block_info, level);
7479                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7480         }
7481
7482         if (parent > 0) {
7483                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7484                 btrfs_set_extent_inline_ref_type(leaf, iref,
7485                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7486                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7487         } else {
7488                 btrfs_set_extent_inline_ref_type(leaf, iref,
7489                                                  BTRFS_TREE_BLOCK_REF_KEY);
7490                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7491         }
7492
7493         btrfs_mark_buffer_dirty(leaf);
7494         btrfs_free_path(path);
7495
7496         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7497                                  1);
7498         if (ret) { /* -ENOENT, logic error */
7499                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7500                         ins->objectid, ins->offset);
7501                 BUG();
7502         }
7503
7504         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7505         return ret;
7506 }
7507
7508 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7509                                      struct btrfs_root *root,
7510                                      u64 root_objectid, u64 owner,
7511                                      u64 offset, struct btrfs_key *ins)
7512 {
7513         int ret;
7514
7515         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7516
7517         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7518                                          ins->offset, 0,
7519                                          root_objectid, owner, offset,
7520                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7521         return ret;
7522 }
7523
7524 /*
7525  * this is used by the tree logging recovery code.  It records that
7526  * an extent has been allocated and makes sure to clear the free
7527  * space cache bits as well
7528  */
7529 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7530                                    struct btrfs_root *root,
7531                                    u64 root_objectid, u64 owner, u64 offset,
7532                                    struct btrfs_key *ins)
7533 {
7534         int ret;
7535         struct btrfs_block_group_cache *block_group;
7536
7537         /*
7538          * Mixed block groups will exclude before processing the log so we only
7539          * need to do the exlude dance if this fs isn't mixed.
7540          */
7541         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7542                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7543                 if (ret)
7544                         return ret;
7545         }
7546
7547         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7548         if (!block_group)
7549                 return -EINVAL;
7550
7551         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7552                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7553         BUG_ON(ret); /* logic error */
7554         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7555                                          0, owner, offset, ins, 1);
7556         btrfs_put_block_group(block_group);
7557         return ret;
7558 }
7559
7560 static struct extent_buffer *
7561 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7562                       u64 bytenr, int level)
7563 {
7564         struct extent_buffer *buf;
7565
7566         buf = btrfs_find_create_tree_block(root, bytenr);
7567         if (!buf)
7568                 return ERR_PTR(-ENOMEM);
7569         btrfs_set_header_generation(buf, trans->transid);
7570         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7571         btrfs_tree_lock(buf);
7572         clean_tree_block(trans, root->fs_info, buf);
7573         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7574
7575         btrfs_set_lock_blocking(buf);
7576         btrfs_set_buffer_uptodate(buf);
7577
7578         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7579                 buf->log_index = root->log_transid % 2;
7580                 /*
7581                  * we allow two log transactions at a time, use different
7582                  * EXENT bit to differentiate dirty pages.
7583                  */
7584                 if (buf->log_index == 0)
7585                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7586                                         buf->start + buf->len - 1, GFP_NOFS);
7587                 else
7588                         set_extent_new(&root->dirty_log_pages, buf->start,
7589                                         buf->start + buf->len - 1, GFP_NOFS);
7590         } else {
7591                 buf->log_index = -1;
7592                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7593                          buf->start + buf->len - 1, GFP_NOFS);
7594         }
7595         trans->blocks_used++;
7596         /* this returns a buffer locked for blocking */
7597         return buf;
7598 }
7599
7600 static struct btrfs_block_rsv *
7601 use_block_rsv(struct btrfs_trans_handle *trans,
7602               struct btrfs_root *root, u32 blocksize)
7603 {
7604         struct btrfs_block_rsv *block_rsv;
7605         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7606         int ret;
7607         bool global_updated = false;
7608
7609         block_rsv = get_block_rsv(trans, root);
7610
7611         if (unlikely(block_rsv->size == 0))
7612                 goto try_reserve;
7613 again:
7614         ret = block_rsv_use_bytes(block_rsv, blocksize);
7615         if (!ret)
7616                 return block_rsv;
7617
7618         if (block_rsv->failfast)
7619                 return ERR_PTR(ret);
7620
7621         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7622                 global_updated = true;
7623                 update_global_block_rsv(root->fs_info);
7624                 goto again;
7625         }
7626
7627         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7628                 static DEFINE_RATELIMIT_STATE(_rs,
7629                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7630                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7631                 if (__ratelimit(&_rs))
7632                         WARN(1, KERN_DEBUG
7633                                 "BTRFS: block rsv returned %d\n", ret);
7634         }
7635 try_reserve:
7636         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7637                                      BTRFS_RESERVE_NO_FLUSH);
7638         if (!ret)
7639                 return block_rsv;
7640         /*
7641          * If we couldn't reserve metadata bytes try and use some from
7642          * the global reserve if its space type is the same as the global
7643          * reservation.
7644          */
7645         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7646             block_rsv->space_info == global_rsv->space_info) {
7647                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7648                 if (!ret)
7649                         return global_rsv;
7650         }
7651         return ERR_PTR(ret);
7652 }
7653
7654 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7655                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7656 {
7657         block_rsv_add_bytes(block_rsv, blocksize, 0);
7658         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7659 }
7660
7661 /*
7662  * finds a free extent and does all the dirty work required for allocation
7663  * returns the tree buffer or an ERR_PTR on error.
7664  */
7665 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7666                                         struct btrfs_root *root,
7667                                         u64 parent, u64 root_objectid,
7668                                         struct btrfs_disk_key *key, int level,
7669                                         u64 hint, u64 empty_size)
7670 {
7671         struct btrfs_key ins;
7672         struct btrfs_block_rsv *block_rsv;
7673         struct extent_buffer *buf;
7674         struct btrfs_delayed_extent_op *extent_op;
7675         u64 flags = 0;
7676         int ret;
7677         u32 blocksize = root->nodesize;
7678         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7679                                                  SKINNY_METADATA);
7680
7681         if (btrfs_test_is_dummy_root(root)) {
7682                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7683                                             level);
7684                 if (!IS_ERR(buf))
7685                         root->alloc_bytenr += blocksize;
7686                 return buf;
7687         }
7688
7689         block_rsv = use_block_rsv(trans, root, blocksize);
7690         if (IS_ERR(block_rsv))
7691                 return ERR_CAST(block_rsv);
7692
7693         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7694                                    empty_size, hint, &ins, 0, 0);
7695         if (ret)
7696                 goto out_unuse;
7697
7698         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
7699         if (IS_ERR(buf)) {
7700                 ret = PTR_ERR(buf);
7701                 goto out_free_reserved;
7702         }
7703
7704         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7705                 if (parent == 0)
7706                         parent = ins.objectid;
7707                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7708         } else
7709                 BUG_ON(parent > 0);
7710
7711         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7712                 extent_op = btrfs_alloc_delayed_extent_op();
7713                 if (!extent_op) {
7714                         ret = -ENOMEM;
7715                         goto out_free_buf;
7716                 }
7717                 if (key)
7718                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7719                 else
7720                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7721                 extent_op->flags_to_set = flags;
7722                 if (skinny_metadata)
7723                         extent_op->update_key = 0;
7724                 else
7725                         extent_op->update_key = 1;
7726                 extent_op->update_flags = 1;
7727                 extent_op->is_data = 0;
7728                 extent_op->level = level;
7729
7730                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7731                                                  ins.objectid, ins.offset,
7732                                                  parent, root_objectid, level,
7733                                                  BTRFS_ADD_DELAYED_EXTENT,
7734                                                  extent_op, 0);
7735                 if (ret)
7736                         goto out_free_delayed;
7737         }
7738         return buf;
7739
7740 out_free_delayed:
7741         btrfs_free_delayed_extent_op(extent_op);
7742 out_free_buf:
7743         free_extent_buffer(buf);
7744 out_free_reserved:
7745         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
7746 out_unuse:
7747         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7748         return ERR_PTR(ret);
7749 }
7750
7751 struct walk_control {
7752         u64 refs[BTRFS_MAX_LEVEL];
7753         u64 flags[BTRFS_MAX_LEVEL];
7754         struct btrfs_key update_progress;
7755         int stage;
7756         int level;
7757         int shared_level;
7758         int update_ref;
7759         int keep_locks;
7760         int reada_slot;
7761         int reada_count;
7762         int for_reloc;
7763 };
7764
7765 #define DROP_REFERENCE  1
7766 #define UPDATE_BACKREF  2
7767
7768 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7769                                      struct btrfs_root *root,
7770                                      struct walk_control *wc,
7771                                      struct btrfs_path *path)
7772 {
7773         u64 bytenr;
7774         u64 generation;
7775         u64 refs;
7776         u64 flags;
7777         u32 nritems;
7778         u32 blocksize;
7779         struct btrfs_key key;
7780         struct extent_buffer *eb;
7781         int ret;
7782         int slot;
7783         int nread = 0;
7784
7785         if (path->slots[wc->level] < wc->reada_slot) {
7786                 wc->reada_count = wc->reada_count * 2 / 3;
7787                 wc->reada_count = max(wc->reada_count, 2);
7788         } else {
7789                 wc->reada_count = wc->reada_count * 3 / 2;
7790                 wc->reada_count = min_t(int, wc->reada_count,
7791                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7792         }
7793
7794         eb = path->nodes[wc->level];
7795         nritems = btrfs_header_nritems(eb);
7796         blocksize = root->nodesize;
7797
7798         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7799                 if (nread >= wc->reada_count)
7800                         break;
7801
7802                 cond_resched();
7803                 bytenr = btrfs_node_blockptr(eb, slot);
7804                 generation = btrfs_node_ptr_generation(eb, slot);
7805
7806                 if (slot == path->slots[wc->level])
7807                         goto reada;
7808
7809                 if (wc->stage == UPDATE_BACKREF &&
7810                     generation <= root->root_key.offset)
7811                         continue;
7812
7813                 /* We don't lock the tree block, it's OK to be racy here */
7814                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7815                                                wc->level - 1, 1, &refs,
7816                                                &flags);
7817                 /* We don't care about errors in readahead. */
7818                 if (ret < 0)
7819                         continue;
7820                 BUG_ON(refs == 0);
7821
7822                 if (wc->stage == DROP_REFERENCE) {
7823                         if (refs == 1)
7824                                 goto reada;
7825
7826                         if (wc->level == 1 &&
7827                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7828                                 continue;
7829                         if (!wc->update_ref ||
7830                             generation <= root->root_key.offset)
7831                                 continue;
7832                         btrfs_node_key_to_cpu(eb, &key, slot);
7833                         ret = btrfs_comp_cpu_keys(&key,
7834                                                   &wc->update_progress);
7835                         if (ret < 0)
7836                                 continue;
7837                 } else {
7838                         if (wc->level == 1 &&
7839                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7840                                 continue;
7841                 }
7842 reada:
7843                 readahead_tree_block(root, bytenr);
7844                 nread++;
7845         }
7846         wc->reada_slot = slot;
7847 }
7848
7849 /*
7850  * TODO: Modify related function to add related node/leaf to dirty_extent_root,
7851  * for later qgroup accounting.
7852  *
7853  * Current, this function does nothing.
7854  */
7855 static int account_leaf_items(struct btrfs_trans_handle *trans,
7856                               struct btrfs_root *root,
7857                               struct extent_buffer *eb)
7858 {
7859         int nr = btrfs_header_nritems(eb);
7860         int i, extent_type;
7861         struct btrfs_key key;
7862         struct btrfs_file_extent_item *fi;
7863         u64 bytenr, num_bytes;
7864
7865         for (i = 0; i < nr; i++) {
7866                 btrfs_item_key_to_cpu(eb, &key, i);
7867
7868                 if (key.type != BTRFS_EXTENT_DATA_KEY)
7869                         continue;
7870
7871                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
7872                 /* filter out non qgroup-accountable extents  */
7873                 extent_type = btrfs_file_extent_type(eb, fi);
7874
7875                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
7876                         continue;
7877
7878                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
7879                 if (!bytenr)
7880                         continue;
7881
7882                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
7883         }
7884         return 0;
7885 }
7886
7887 /*
7888  * Walk up the tree from the bottom, freeing leaves and any interior
7889  * nodes which have had all slots visited. If a node (leaf or
7890  * interior) is freed, the node above it will have it's slot
7891  * incremented. The root node will never be freed.
7892  *
7893  * At the end of this function, we should have a path which has all
7894  * slots incremented to the next position for a search. If we need to
7895  * read a new node it will be NULL and the node above it will have the
7896  * correct slot selected for a later read.
7897  *
7898  * If we increment the root nodes slot counter past the number of
7899  * elements, 1 is returned to signal completion of the search.
7900  */
7901 static int adjust_slots_upwards(struct btrfs_root *root,
7902                                 struct btrfs_path *path, int root_level)
7903 {
7904         int level = 0;
7905         int nr, slot;
7906         struct extent_buffer *eb;
7907
7908         if (root_level == 0)
7909                 return 1;
7910
7911         while (level <= root_level) {
7912                 eb = path->nodes[level];
7913                 nr = btrfs_header_nritems(eb);
7914                 path->slots[level]++;
7915                 slot = path->slots[level];
7916                 if (slot >= nr || level == 0) {
7917                         /*
7918                          * Don't free the root -  we will detect this
7919                          * condition after our loop and return a
7920                          * positive value for caller to stop walking the tree.
7921                          */
7922                         if (level != root_level) {
7923                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7924                                 path->locks[level] = 0;
7925
7926                                 free_extent_buffer(eb);
7927                                 path->nodes[level] = NULL;
7928                                 path->slots[level] = 0;
7929                         }
7930                 } else {
7931                         /*
7932                          * We have a valid slot to walk back down
7933                          * from. Stop here so caller can process these
7934                          * new nodes.
7935                          */
7936                         break;
7937                 }
7938
7939                 level++;
7940         }
7941
7942         eb = path->nodes[root_level];
7943         if (path->slots[root_level] >= btrfs_header_nritems(eb))
7944                 return 1;
7945
7946         return 0;
7947 }
7948
7949 /*
7950  * root_eb is the subtree root and is locked before this function is called.
7951  * TODO: Modify this function to mark all (including complete shared node)
7952  * to dirty_extent_root to allow it get accounted in qgroup.
7953  */
7954 static int account_shared_subtree(struct btrfs_trans_handle *trans,
7955                                   struct btrfs_root *root,
7956                                   struct extent_buffer *root_eb,
7957                                   u64 root_gen,
7958                                   int root_level)
7959 {
7960         int ret = 0;
7961         int level;
7962         struct extent_buffer *eb = root_eb;
7963         struct btrfs_path *path = NULL;
7964
7965         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
7966         BUG_ON(root_eb == NULL);
7967
7968         if (!root->fs_info->quota_enabled)
7969                 return 0;
7970
7971         if (!extent_buffer_uptodate(root_eb)) {
7972                 ret = btrfs_read_buffer(root_eb, root_gen);
7973                 if (ret)
7974                         goto out;
7975         }
7976
7977         if (root_level == 0) {
7978                 ret = account_leaf_items(trans, root, root_eb);
7979                 goto out;
7980         }
7981
7982         path = btrfs_alloc_path();
7983         if (!path)
7984                 return -ENOMEM;
7985
7986         /*
7987          * Walk down the tree.  Missing extent blocks are filled in as
7988          * we go. Metadata is accounted every time we read a new
7989          * extent block.
7990          *
7991          * When we reach a leaf, we account for file extent items in it,
7992          * walk back up the tree (adjusting slot pointers as we go)
7993          * and restart the search process.
7994          */
7995         extent_buffer_get(root_eb); /* For path */
7996         path->nodes[root_level] = root_eb;
7997         path->slots[root_level] = 0;
7998         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
7999 walk_down:
8000         level = root_level;
8001         while (level >= 0) {
8002                 if (path->nodes[level] == NULL) {
8003                         int parent_slot;
8004                         u64 child_gen;
8005                         u64 child_bytenr;
8006
8007                         /* We need to get child blockptr/gen from
8008                          * parent before we can read it. */
8009                         eb = path->nodes[level + 1];
8010                         parent_slot = path->slots[level + 1];
8011                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8012                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8013
8014                         eb = read_tree_block(root, child_bytenr, child_gen);
8015                         if (IS_ERR(eb)) {
8016                                 ret = PTR_ERR(eb);
8017                                 goto out;
8018                         } else if (!extent_buffer_uptodate(eb)) {
8019                                 free_extent_buffer(eb);
8020                                 ret = -EIO;
8021                                 goto out;
8022                         }
8023
8024                         path->nodes[level] = eb;
8025                         path->slots[level] = 0;
8026
8027                         btrfs_tree_read_lock(eb);
8028                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8029                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8030                 }
8031
8032                 if (level == 0) {
8033                         ret = account_leaf_items(trans, root, path->nodes[level]);
8034                         if (ret)
8035                                 goto out;
8036
8037                         /* Nonzero return here means we completed our search */
8038                         ret = adjust_slots_upwards(root, path, root_level);
8039                         if (ret)
8040                                 break;
8041
8042                         /* Restart search with new slots */
8043                         goto walk_down;
8044                 }
8045
8046                 level--;
8047         }
8048
8049         ret = 0;
8050 out:
8051         btrfs_free_path(path);
8052
8053         return ret;
8054 }
8055
8056 /*
8057  * helper to process tree block while walking down the tree.
8058  *
8059  * when wc->stage == UPDATE_BACKREF, this function updates
8060  * back refs for pointers in the block.
8061  *
8062  * NOTE: return value 1 means we should stop walking down.
8063  */
8064 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8065                                    struct btrfs_root *root,
8066                                    struct btrfs_path *path,
8067                                    struct walk_control *wc, int lookup_info)
8068 {
8069         int level = wc->level;
8070         struct extent_buffer *eb = path->nodes[level];
8071         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8072         int ret;
8073
8074         if (wc->stage == UPDATE_BACKREF &&
8075             btrfs_header_owner(eb) != root->root_key.objectid)
8076                 return 1;
8077
8078         /*
8079          * when reference count of tree block is 1, it won't increase
8080          * again. once full backref flag is set, we never clear it.
8081          */
8082         if (lookup_info &&
8083             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8084              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8085                 BUG_ON(!path->locks[level]);
8086                 ret = btrfs_lookup_extent_info(trans, root,
8087                                                eb->start, level, 1,
8088                                                &wc->refs[level],
8089                                                &wc->flags[level]);
8090                 BUG_ON(ret == -ENOMEM);
8091                 if (ret)
8092                         return ret;
8093                 BUG_ON(wc->refs[level] == 0);
8094         }
8095
8096         if (wc->stage == DROP_REFERENCE) {
8097                 if (wc->refs[level] > 1)
8098                         return 1;
8099
8100                 if (path->locks[level] && !wc->keep_locks) {
8101                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8102                         path->locks[level] = 0;
8103                 }
8104                 return 0;
8105         }
8106
8107         /* wc->stage == UPDATE_BACKREF */
8108         if (!(wc->flags[level] & flag)) {
8109                 BUG_ON(!path->locks[level]);
8110                 ret = btrfs_inc_ref(trans, root, eb, 1);
8111                 BUG_ON(ret); /* -ENOMEM */
8112                 ret = btrfs_dec_ref(trans, root, eb, 0);
8113                 BUG_ON(ret); /* -ENOMEM */
8114                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8115                                                   eb->len, flag,
8116                                                   btrfs_header_level(eb), 0);
8117                 BUG_ON(ret); /* -ENOMEM */
8118                 wc->flags[level] |= flag;
8119         }
8120
8121         /*
8122          * the block is shared by multiple trees, so it's not good to
8123          * keep the tree lock
8124          */
8125         if (path->locks[level] && level > 0) {
8126                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8127                 path->locks[level] = 0;
8128         }
8129         return 0;
8130 }
8131
8132 /*
8133  * helper to process tree block pointer.
8134  *
8135  * when wc->stage == DROP_REFERENCE, this function checks
8136  * reference count of the block pointed to. if the block
8137  * is shared and we need update back refs for the subtree
8138  * rooted at the block, this function changes wc->stage to
8139  * UPDATE_BACKREF. if the block is shared and there is no
8140  * need to update back, this function drops the reference
8141  * to the block.
8142  *
8143  * NOTE: return value 1 means we should stop walking down.
8144  */
8145 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8146                                  struct btrfs_root *root,
8147                                  struct btrfs_path *path,
8148                                  struct walk_control *wc, int *lookup_info)
8149 {
8150         u64 bytenr;
8151         u64 generation;
8152         u64 parent;
8153         u32 blocksize;
8154         struct btrfs_key key;
8155         struct extent_buffer *next;
8156         int level = wc->level;
8157         int reada = 0;
8158         int ret = 0;
8159         bool need_account = false;
8160
8161         generation = btrfs_node_ptr_generation(path->nodes[level],
8162                                                path->slots[level]);
8163         /*
8164          * if the lower level block was created before the snapshot
8165          * was created, we know there is no need to update back refs
8166          * for the subtree
8167          */
8168         if (wc->stage == UPDATE_BACKREF &&
8169             generation <= root->root_key.offset) {
8170                 *lookup_info = 1;
8171                 return 1;
8172         }
8173
8174         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8175         blocksize = root->nodesize;
8176
8177         next = btrfs_find_tree_block(root->fs_info, bytenr);
8178         if (!next) {
8179                 next = btrfs_find_create_tree_block(root, bytenr);
8180                 if (!next)
8181                         return -ENOMEM;
8182                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8183                                                level - 1);
8184                 reada = 1;
8185         }
8186         btrfs_tree_lock(next);
8187         btrfs_set_lock_blocking(next);
8188
8189         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8190                                        &wc->refs[level - 1],
8191                                        &wc->flags[level - 1]);
8192         if (ret < 0) {
8193                 btrfs_tree_unlock(next);
8194                 return ret;
8195         }
8196
8197         if (unlikely(wc->refs[level - 1] == 0)) {
8198                 btrfs_err(root->fs_info, "Missing references.");
8199                 BUG();
8200         }
8201         *lookup_info = 0;
8202
8203         if (wc->stage == DROP_REFERENCE) {
8204                 if (wc->refs[level - 1] > 1) {
8205                         need_account = true;
8206                         if (level == 1 &&
8207                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8208                                 goto skip;
8209
8210                         if (!wc->update_ref ||
8211                             generation <= root->root_key.offset)
8212                                 goto skip;
8213
8214                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8215                                               path->slots[level]);
8216                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8217                         if (ret < 0)
8218                                 goto skip;
8219
8220                         wc->stage = UPDATE_BACKREF;
8221                         wc->shared_level = level - 1;
8222                 }
8223         } else {
8224                 if (level == 1 &&
8225                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8226                         goto skip;
8227         }
8228
8229         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8230                 btrfs_tree_unlock(next);
8231                 free_extent_buffer(next);
8232                 next = NULL;
8233                 *lookup_info = 1;
8234         }
8235
8236         if (!next) {
8237                 if (reada && level == 1)
8238                         reada_walk_down(trans, root, wc, path);
8239                 next = read_tree_block(root, bytenr, generation);
8240                 if (IS_ERR(next)) {
8241                         return PTR_ERR(next);
8242                 } else if (!extent_buffer_uptodate(next)) {
8243                         free_extent_buffer(next);
8244                         return -EIO;
8245                 }
8246                 btrfs_tree_lock(next);
8247                 btrfs_set_lock_blocking(next);
8248         }
8249
8250         level--;
8251         BUG_ON(level != btrfs_header_level(next));
8252         path->nodes[level] = next;
8253         path->slots[level] = 0;
8254         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8255         wc->level = level;
8256         if (wc->level == 1)
8257                 wc->reada_slot = 0;
8258         return 0;
8259 skip:
8260         wc->refs[level - 1] = 0;
8261         wc->flags[level - 1] = 0;
8262         if (wc->stage == DROP_REFERENCE) {
8263                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8264                         parent = path->nodes[level]->start;
8265                 } else {
8266                         BUG_ON(root->root_key.objectid !=
8267                                btrfs_header_owner(path->nodes[level]));
8268                         parent = 0;
8269                 }
8270
8271                 if (need_account) {
8272                         ret = account_shared_subtree(trans, root, next,
8273                                                      generation, level - 1);
8274                         if (ret) {
8275                                 btrfs_err_rl(root->fs_info,
8276                                         "Error "
8277                                         "%d accounting shared subtree. Quota "
8278                                         "is out of sync, rescan required.",
8279                                         ret);
8280                         }
8281                 }
8282                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8283                                 root->root_key.objectid, level - 1, 0, 0);
8284                 BUG_ON(ret); /* -ENOMEM */
8285         }
8286         btrfs_tree_unlock(next);
8287         free_extent_buffer(next);
8288         *lookup_info = 1;
8289         return 1;
8290 }
8291
8292 /*
8293  * helper to process tree block while walking up the tree.
8294  *
8295  * when wc->stage == DROP_REFERENCE, this function drops
8296  * reference count on the block.
8297  *
8298  * when wc->stage == UPDATE_BACKREF, this function changes
8299  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8300  * to UPDATE_BACKREF previously while processing the block.
8301  *
8302  * NOTE: return value 1 means we should stop walking up.
8303  */
8304 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8305                                  struct btrfs_root *root,
8306                                  struct btrfs_path *path,
8307                                  struct walk_control *wc)
8308 {
8309         int ret;
8310         int level = wc->level;
8311         struct extent_buffer *eb = path->nodes[level];
8312         u64 parent = 0;
8313
8314         if (wc->stage == UPDATE_BACKREF) {
8315                 BUG_ON(wc->shared_level < level);
8316                 if (level < wc->shared_level)
8317                         goto out;
8318
8319                 ret = find_next_key(path, level + 1, &wc->update_progress);
8320                 if (ret > 0)
8321                         wc->update_ref = 0;
8322
8323                 wc->stage = DROP_REFERENCE;
8324                 wc->shared_level = -1;
8325                 path->slots[level] = 0;
8326
8327                 /*
8328                  * check reference count again if the block isn't locked.
8329                  * we should start walking down the tree again if reference
8330                  * count is one.
8331                  */
8332                 if (!path->locks[level]) {
8333                         BUG_ON(level == 0);
8334                         btrfs_tree_lock(eb);
8335                         btrfs_set_lock_blocking(eb);
8336                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8337
8338                         ret = btrfs_lookup_extent_info(trans, root,
8339                                                        eb->start, level, 1,
8340                                                        &wc->refs[level],
8341                                                        &wc->flags[level]);
8342                         if (ret < 0) {
8343                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8344                                 path->locks[level] = 0;
8345                                 return ret;
8346                         }
8347                         BUG_ON(wc->refs[level] == 0);
8348                         if (wc->refs[level] == 1) {
8349                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8350                                 path->locks[level] = 0;
8351                                 return 1;
8352                         }
8353                 }
8354         }
8355
8356         /* wc->stage == DROP_REFERENCE */
8357         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8358
8359         if (wc->refs[level] == 1) {
8360                 if (level == 0) {
8361                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8362                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8363                         else
8364                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8365                         BUG_ON(ret); /* -ENOMEM */
8366                         ret = account_leaf_items(trans, root, eb);
8367                         if (ret) {
8368                                 btrfs_err_rl(root->fs_info,
8369                                         "error "
8370                                         "%d accounting leaf items. Quota "
8371                                         "is out of sync, rescan required.",
8372                                         ret);
8373                         }
8374                 }
8375                 /* make block locked assertion in clean_tree_block happy */
8376                 if (!path->locks[level] &&
8377                     btrfs_header_generation(eb) == trans->transid) {
8378                         btrfs_tree_lock(eb);
8379                         btrfs_set_lock_blocking(eb);
8380                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8381                 }
8382                 clean_tree_block(trans, root->fs_info, eb);
8383         }
8384
8385         if (eb == root->node) {
8386                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8387                         parent = eb->start;
8388                 else
8389                         BUG_ON(root->root_key.objectid !=
8390                                btrfs_header_owner(eb));
8391         } else {
8392                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8393                         parent = path->nodes[level + 1]->start;
8394                 else
8395                         BUG_ON(root->root_key.objectid !=
8396                                btrfs_header_owner(path->nodes[level + 1]));
8397         }
8398
8399         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8400 out:
8401         wc->refs[level] = 0;
8402         wc->flags[level] = 0;
8403         return 0;
8404 }
8405
8406 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8407                                    struct btrfs_root *root,
8408                                    struct btrfs_path *path,
8409                                    struct walk_control *wc)
8410 {
8411         int level = wc->level;
8412         int lookup_info = 1;
8413         int ret;
8414
8415         while (level >= 0) {
8416                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8417                 if (ret > 0)
8418                         break;
8419
8420                 if (level == 0)
8421                         break;
8422
8423                 if (path->slots[level] >=
8424                     btrfs_header_nritems(path->nodes[level]))
8425                         break;
8426
8427                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8428                 if (ret > 0) {
8429                         path->slots[level]++;
8430                         continue;
8431                 } else if (ret < 0)
8432                         return ret;
8433                 level = wc->level;
8434         }
8435         return 0;
8436 }
8437
8438 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8439                                  struct btrfs_root *root,
8440                                  struct btrfs_path *path,
8441                                  struct walk_control *wc, int max_level)
8442 {
8443         int level = wc->level;
8444         int ret;
8445
8446         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8447         while (level < max_level && path->nodes[level]) {
8448                 wc->level = level;
8449                 if (path->slots[level] + 1 <
8450                     btrfs_header_nritems(path->nodes[level])) {
8451                         path->slots[level]++;
8452                         return 0;
8453                 } else {
8454                         ret = walk_up_proc(trans, root, path, wc);
8455                         if (ret > 0)
8456                                 return 0;
8457
8458                         if (path->locks[level]) {
8459                                 btrfs_tree_unlock_rw(path->nodes[level],
8460                                                      path->locks[level]);
8461                                 path->locks[level] = 0;
8462                         }
8463                         free_extent_buffer(path->nodes[level]);
8464                         path->nodes[level] = NULL;
8465                         level++;
8466                 }
8467         }
8468         return 1;
8469 }
8470
8471 /*
8472  * drop a subvolume tree.
8473  *
8474  * this function traverses the tree freeing any blocks that only
8475  * referenced by the tree.
8476  *
8477  * when a shared tree block is found. this function decreases its
8478  * reference count by one. if update_ref is true, this function
8479  * also make sure backrefs for the shared block and all lower level
8480  * blocks are properly updated.
8481  *
8482  * If called with for_reloc == 0, may exit early with -EAGAIN
8483  */
8484 int btrfs_drop_snapshot(struct btrfs_root *root,
8485                          struct btrfs_block_rsv *block_rsv, int update_ref,
8486                          int for_reloc)
8487 {
8488         struct btrfs_path *path;
8489         struct btrfs_trans_handle *trans;
8490         struct btrfs_root *tree_root = root->fs_info->tree_root;
8491         struct btrfs_root_item *root_item = &root->root_item;
8492         struct walk_control *wc;
8493         struct btrfs_key key;
8494         int err = 0;
8495         int ret;
8496         int level;
8497         bool root_dropped = false;
8498
8499         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8500
8501         path = btrfs_alloc_path();
8502         if (!path) {
8503                 err = -ENOMEM;
8504                 goto out;
8505         }
8506
8507         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8508         if (!wc) {
8509                 btrfs_free_path(path);
8510                 err = -ENOMEM;
8511                 goto out;
8512         }
8513
8514         trans = btrfs_start_transaction(tree_root, 0);
8515         if (IS_ERR(trans)) {
8516                 err = PTR_ERR(trans);
8517                 goto out_free;
8518         }
8519
8520         if (block_rsv)
8521                 trans->block_rsv = block_rsv;
8522
8523         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8524                 level = btrfs_header_level(root->node);
8525                 path->nodes[level] = btrfs_lock_root_node(root);
8526                 btrfs_set_lock_blocking(path->nodes[level]);
8527                 path->slots[level] = 0;
8528                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8529                 memset(&wc->update_progress, 0,
8530                        sizeof(wc->update_progress));
8531         } else {
8532                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8533                 memcpy(&wc->update_progress, &key,
8534                        sizeof(wc->update_progress));
8535
8536                 level = root_item->drop_level;
8537                 BUG_ON(level == 0);
8538                 path->lowest_level = level;
8539                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8540                 path->lowest_level = 0;
8541                 if (ret < 0) {
8542                         err = ret;
8543                         goto out_end_trans;
8544                 }
8545                 WARN_ON(ret > 0);
8546
8547                 /*
8548                  * unlock our path, this is safe because only this
8549                  * function is allowed to delete this snapshot
8550                  */
8551                 btrfs_unlock_up_safe(path, 0);
8552
8553                 level = btrfs_header_level(root->node);
8554                 while (1) {
8555                         btrfs_tree_lock(path->nodes[level]);
8556                         btrfs_set_lock_blocking(path->nodes[level]);
8557                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8558
8559                         ret = btrfs_lookup_extent_info(trans, root,
8560                                                 path->nodes[level]->start,
8561                                                 level, 1, &wc->refs[level],
8562                                                 &wc->flags[level]);
8563                         if (ret < 0) {
8564                                 err = ret;
8565                                 goto out_end_trans;
8566                         }
8567                         BUG_ON(wc->refs[level] == 0);
8568
8569                         if (level == root_item->drop_level)
8570                                 break;
8571
8572                         btrfs_tree_unlock(path->nodes[level]);
8573                         path->locks[level] = 0;
8574                         WARN_ON(wc->refs[level] != 1);
8575                         level--;
8576                 }
8577         }
8578
8579         wc->level = level;
8580         wc->shared_level = -1;
8581         wc->stage = DROP_REFERENCE;
8582         wc->update_ref = update_ref;
8583         wc->keep_locks = 0;
8584         wc->for_reloc = for_reloc;
8585         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8586
8587         while (1) {
8588
8589                 ret = walk_down_tree(trans, root, path, wc);
8590                 if (ret < 0) {
8591                         err = ret;
8592                         break;
8593                 }
8594
8595                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8596                 if (ret < 0) {
8597                         err = ret;
8598                         break;
8599                 }
8600
8601                 if (ret > 0) {
8602                         BUG_ON(wc->stage != DROP_REFERENCE);
8603                         break;
8604                 }
8605
8606                 if (wc->stage == DROP_REFERENCE) {
8607                         level = wc->level;
8608                         btrfs_node_key(path->nodes[level],
8609                                        &root_item->drop_progress,
8610                                        path->slots[level]);
8611                         root_item->drop_level = level;
8612                 }
8613
8614                 BUG_ON(wc->level == 0);
8615                 if (btrfs_should_end_transaction(trans, tree_root) ||
8616                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8617                         ret = btrfs_update_root(trans, tree_root,
8618                                                 &root->root_key,
8619                                                 root_item);
8620                         if (ret) {
8621                                 btrfs_abort_transaction(trans, tree_root, ret);
8622                                 err = ret;
8623                                 goto out_end_trans;
8624                         }
8625
8626                         btrfs_end_transaction_throttle(trans, tree_root);
8627                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8628                                 pr_debug("BTRFS: drop snapshot early exit\n");
8629                                 err = -EAGAIN;
8630                                 goto out_free;
8631                         }
8632
8633                         trans = btrfs_start_transaction(tree_root, 0);
8634                         if (IS_ERR(trans)) {
8635                                 err = PTR_ERR(trans);
8636                                 goto out_free;
8637                         }
8638                         if (block_rsv)
8639                                 trans->block_rsv = block_rsv;
8640                 }
8641         }
8642         btrfs_release_path(path);
8643         if (err)
8644                 goto out_end_trans;
8645
8646         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8647         if (ret) {
8648                 btrfs_abort_transaction(trans, tree_root, ret);
8649                 goto out_end_trans;
8650         }
8651
8652         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8653                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8654                                       NULL, NULL);
8655                 if (ret < 0) {
8656                         btrfs_abort_transaction(trans, tree_root, ret);
8657                         err = ret;
8658                         goto out_end_trans;
8659                 } else if (ret > 0) {
8660                         /* if we fail to delete the orphan item this time
8661                          * around, it'll get picked up the next time.
8662                          *
8663                          * The most common failure here is just -ENOENT.
8664                          */
8665                         btrfs_del_orphan_item(trans, tree_root,
8666                                               root->root_key.objectid);
8667                 }
8668         }
8669
8670         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8671                 btrfs_add_dropped_root(trans, root);
8672         } else {
8673                 free_extent_buffer(root->node);
8674                 free_extent_buffer(root->commit_root);
8675                 btrfs_put_fs_root(root);
8676         }
8677         root_dropped = true;
8678 out_end_trans:
8679         btrfs_end_transaction_throttle(trans, tree_root);
8680 out_free:
8681         kfree(wc);
8682         btrfs_free_path(path);
8683 out:
8684         /*
8685          * So if we need to stop dropping the snapshot for whatever reason we
8686          * need to make sure to add it back to the dead root list so that we
8687          * keep trying to do the work later.  This also cleans up roots if we
8688          * don't have it in the radix (like when we recover after a power fail
8689          * or unmount) so we don't leak memory.
8690          */
8691         if (!for_reloc && root_dropped == false)
8692                 btrfs_add_dead_root(root);
8693         if (err && err != -EAGAIN)
8694                 btrfs_std_error(root->fs_info, err, NULL);
8695         return err;
8696 }
8697
8698 /*
8699  * drop subtree rooted at tree block 'node'.
8700  *
8701  * NOTE: this function will unlock and release tree block 'node'
8702  * only used by relocation code
8703  */
8704 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8705                         struct btrfs_root *root,
8706                         struct extent_buffer *node,
8707                         struct extent_buffer *parent)
8708 {
8709         struct btrfs_path *path;
8710         struct walk_control *wc;
8711         int level;
8712         int parent_level;
8713         int ret = 0;
8714         int wret;
8715
8716         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8717
8718         path = btrfs_alloc_path();
8719         if (!path)
8720                 return -ENOMEM;
8721
8722         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8723         if (!wc) {
8724                 btrfs_free_path(path);
8725                 return -ENOMEM;
8726         }
8727
8728         btrfs_assert_tree_locked(parent);
8729         parent_level = btrfs_header_level(parent);
8730         extent_buffer_get(parent);
8731         path->nodes[parent_level] = parent;
8732         path->slots[parent_level] = btrfs_header_nritems(parent);
8733
8734         btrfs_assert_tree_locked(node);
8735         level = btrfs_header_level(node);
8736         path->nodes[level] = node;
8737         path->slots[level] = 0;
8738         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8739
8740         wc->refs[parent_level] = 1;
8741         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8742         wc->level = level;
8743         wc->shared_level = -1;
8744         wc->stage = DROP_REFERENCE;
8745         wc->update_ref = 0;
8746         wc->keep_locks = 1;
8747         wc->for_reloc = 1;
8748         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8749
8750         while (1) {
8751                 wret = walk_down_tree(trans, root, path, wc);
8752                 if (wret < 0) {
8753                         ret = wret;
8754                         break;
8755                 }
8756
8757                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8758                 if (wret < 0)
8759                         ret = wret;
8760                 if (wret != 0)
8761                         break;
8762         }
8763
8764         kfree(wc);
8765         btrfs_free_path(path);
8766         return ret;
8767 }
8768
8769 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8770 {
8771         u64 num_devices;
8772         u64 stripped;
8773
8774         /*
8775          * if restripe for this chunk_type is on pick target profile and
8776          * return, otherwise do the usual balance
8777          */
8778         stripped = get_restripe_target(root->fs_info, flags);
8779         if (stripped)
8780                 return extended_to_chunk(stripped);
8781
8782         num_devices = root->fs_info->fs_devices->rw_devices;
8783
8784         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8785                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8786                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8787
8788         if (num_devices == 1) {
8789                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8790                 stripped = flags & ~stripped;
8791
8792                 /* turn raid0 into single device chunks */
8793                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8794                         return stripped;
8795
8796                 /* turn mirroring into duplication */
8797                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8798                              BTRFS_BLOCK_GROUP_RAID10))
8799                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8800         } else {
8801                 /* they already had raid on here, just return */
8802                 if (flags & stripped)
8803                         return flags;
8804
8805                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8806                 stripped = flags & ~stripped;
8807
8808                 /* switch duplicated blocks with raid1 */
8809                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8810                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8811
8812                 /* this is drive concat, leave it alone */
8813         }
8814
8815         return flags;
8816 }
8817
8818 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8819 {
8820         struct btrfs_space_info *sinfo = cache->space_info;
8821         u64 num_bytes;
8822         u64 min_allocable_bytes;
8823         int ret = -ENOSPC;
8824
8825         /*
8826          * We need some metadata space and system metadata space for
8827          * allocating chunks in some corner cases until we force to set
8828          * it to be readonly.
8829          */
8830         if ((sinfo->flags &
8831              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8832             !force)
8833                 min_allocable_bytes = 1 * 1024 * 1024;
8834         else
8835                 min_allocable_bytes = 0;
8836
8837         spin_lock(&sinfo->lock);
8838         spin_lock(&cache->lock);
8839
8840         if (cache->ro) {
8841                 cache->ro++;
8842                 ret = 0;
8843                 goto out;
8844         }
8845
8846         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8847                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8848
8849         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8850             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8851             min_allocable_bytes <= sinfo->total_bytes) {
8852                 sinfo->bytes_readonly += num_bytes;
8853                 cache->ro++;
8854                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
8855                 ret = 0;
8856         }
8857 out:
8858         spin_unlock(&cache->lock);
8859         spin_unlock(&sinfo->lock);
8860         return ret;
8861 }
8862
8863 int btrfs_inc_block_group_ro(struct btrfs_root *root,
8864                              struct btrfs_block_group_cache *cache)
8865
8866 {
8867         struct btrfs_trans_handle *trans;
8868         u64 alloc_flags;
8869         int ret;
8870
8871 again:
8872         trans = btrfs_join_transaction(root);
8873         if (IS_ERR(trans))
8874                 return PTR_ERR(trans);
8875
8876         /*
8877          * we're not allowed to set block groups readonly after the dirty
8878          * block groups cache has started writing.  If it already started,
8879          * back off and let this transaction commit
8880          */
8881         mutex_lock(&root->fs_info->ro_block_group_mutex);
8882         if (trans->transaction->dirty_bg_run) {
8883                 u64 transid = trans->transid;
8884
8885                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
8886                 btrfs_end_transaction(trans, root);
8887
8888                 ret = btrfs_wait_for_commit(root, transid);
8889                 if (ret)
8890                         return ret;
8891                 goto again;
8892         }
8893
8894         /*
8895          * if we are changing raid levels, try to allocate a corresponding
8896          * block group with the new raid level.
8897          */
8898         alloc_flags = update_block_group_flags(root, cache->flags);
8899         if (alloc_flags != cache->flags) {
8900                 ret = do_chunk_alloc(trans, root, alloc_flags,
8901                                      CHUNK_ALLOC_FORCE);
8902                 /*
8903                  * ENOSPC is allowed here, we may have enough space
8904                  * already allocated at the new raid level to
8905                  * carry on
8906                  */
8907                 if (ret == -ENOSPC)
8908                         ret = 0;
8909                 if (ret < 0)
8910                         goto out;
8911         }
8912
8913         ret = inc_block_group_ro(cache, 0);
8914         if (!ret)
8915                 goto out;
8916         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8917         ret = do_chunk_alloc(trans, root, alloc_flags,
8918                              CHUNK_ALLOC_FORCE);
8919         if (ret < 0)
8920                 goto out;
8921         ret = inc_block_group_ro(cache, 0);
8922 out:
8923         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
8924                 alloc_flags = update_block_group_flags(root, cache->flags);
8925                 lock_chunks(root->fs_info->chunk_root);
8926                 check_system_chunk(trans, root, alloc_flags);
8927                 unlock_chunks(root->fs_info->chunk_root);
8928         }
8929         mutex_unlock(&root->fs_info->ro_block_group_mutex);
8930
8931         btrfs_end_transaction(trans, root);
8932         return ret;
8933 }
8934
8935 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8936                             struct btrfs_root *root, u64 type)
8937 {
8938         u64 alloc_flags = get_alloc_profile(root, type);
8939         return do_chunk_alloc(trans, root, alloc_flags,
8940                               CHUNK_ALLOC_FORCE);
8941 }
8942
8943 /*
8944  * helper to account the unused space of all the readonly block group in the
8945  * space_info. takes mirrors into account.
8946  */
8947 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8948 {
8949         struct btrfs_block_group_cache *block_group;
8950         u64 free_bytes = 0;
8951         int factor;
8952
8953         /* It's df, we don't care if it's racey */
8954         if (list_empty(&sinfo->ro_bgs))
8955                 return 0;
8956
8957         spin_lock(&sinfo->lock);
8958         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
8959                 spin_lock(&block_group->lock);
8960
8961                 if (!block_group->ro) {
8962                         spin_unlock(&block_group->lock);
8963                         continue;
8964                 }
8965
8966                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8967                                           BTRFS_BLOCK_GROUP_RAID10 |
8968                                           BTRFS_BLOCK_GROUP_DUP))
8969                         factor = 2;
8970                 else
8971                         factor = 1;
8972
8973                 free_bytes += (block_group->key.offset -
8974                                btrfs_block_group_used(&block_group->item)) *
8975                                factor;
8976
8977                 spin_unlock(&block_group->lock);
8978         }
8979         spin_unlock(&sinfo->lock);
8980
8981         return free_bytes;
8982 }
8983
8984 void btrfs_dec_block_group_ro(struct btrfs_root *root,
8985                               struct btrfs_block_group_cache *cache)
8986 {
8987         struct btrfs_space_info *sinfo = cache->space_info;
8988         u64 num_bytes;
8989
8990         BUG_ON(!cache->ro);
8991
8992         spin_lock(&sinfo->lock);
8993         spin_lock(&cache->lock);
8994         if (!--cache->ro) {
8995                 num_bytes = cache->key.offset - cache->reserved -
8996                             cache->pinned - cache->bytes_super -
8997                             btrfs_block_group_used(&cache->item);
8998                 sinfo->bytes_readonly -= num_bytes;
8999                 list_del_init(&cache->ro_list);
9000         }
9001         spin_unlock(&cache->lock);
9002         spin_unlock(&sinfo->lock);
9003 }
9004
9005 /*
9006  * checks to see if its even possible to relocate this block group.
9007  *
9008  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9009  * ok to go ahead and try.
9010  */
9011 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9012 {
9013         struct btrfs_block_group_cache *block_group;
9014         struct btrfs_space_info *space_info;
9015         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9016         struct btrfs_device *device;
9017         struct btrfs_trans_handle *trans;
9018         u64 min_free;
9019         u64 dev_min = 1;
9020         u64 dev_nr = 0;
9021         u64 target;
9022         int index;
9023         int full = 0;
9024         int ret = 0;
9025
9026         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9027
9028         /* odd, couldn't find the block group, leave it alone */
9029         if (!block_group)
9030                 return -1;
9031
9032         min_free = btrfs_block_group_used(&block_group->item);
9033
9034         /* no bytes used, we're good */
9035         if (!min_free)
9036                 goto out;
9037
9038         space_info = block_group->space_info;
9039         spin_lock(&space_info->lock);
9040
9041         full = space_info->full;
9042
9043         /*
9044          * if this is the last block group we have in this space, we can't
9045          * relocate it unless we're able to allocate a new chunk below.
9046          *
9047          * Otherwise, we need to make sure we have room in the space to handle
9048          * all of the extents from this block group.  If we can, we're good
9049          */
9050         if ((space_info->total_bytes != block_group->key.offset) &&
9051             (space_info->bytes_used + space_info->bytes_reserved +
9052              space_info->bytes_pinned + space_info->bytes_readonly +
9053              min_free < space_info->total_bytes)) {
9054                 spin_unlock(&space_info->lock);
9055                 goto out;
9056         }
9057         spin_unlock(&space_info->lock);
9058
9059         /*
9060          * ok we don't have enough space, but maybe we have free space on our
9061          * devices to allocate new chunks for relocation, so loop through our
9062          * alloc devices and guess if we have enough space.  if this block
9063          * group is going to be restriped, run checks against the target
9064          * profile instead of the current one.
9065          */
9066         ret = -1;
9067
9068         /*
9069          * index:
9070          *      0: raid10
9071          *      1: raid1
9072          *      2: dup
9073          *      3: raid0
9074          *      4: single
9075          */
9076         target = get_restripe_target(root->fs_info, block_group->flags);
9077         if (target) {
9078                 index = __get_raid_index(extended_to_chunk(target));
9079         } else {
9080                 /*
9081                  * this is just a balance, so if we were marked as full
9082                  * we know there is no space for a new chunk
9083                  */
9084                 if (full)
9085                         goto out;
9086
9087                 index = get_block_group_index(block_group);
9088         }
9089
9090         if (index == BTRFS_RAID_RAID10) {
9091                 dev_min = 4;
9092                 /* Divide by 2 */
9093                 min_free >>= 1;
9094         } else if (index == BTRFS_RAID_RAID1) {
9095                 dev_min = 2;
9096         } else if (index == BTRFS_RAID_DUP) {
9097                 /* Multiply by 2 */
9098                 min_free <<= 1;
9099         } else if (index == BTRFS_RAID_RAID0) {
9100                 dev_min = fs_devices->rw_devices;
9101                 min_free = div64_u64(min_free, dev_min);
9102         }
9103
9104         /* We need to do this so that we can look at pending chunks */
9105         trans = btrfs_join_transaction(root);
9106         if (IS_ERR(trans)) {
9107                 ret = PTR_ERR(trans);
9108                 goto out;
9109         }
9110
9111         mutex_lock(&root->fs_info->chunk_mutex);
9112         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9113                 u64 dev_offset;
9114
9115                 /*
9116                  * check to make sure we can actually find a chunk with enough
9117                  * space to fit our block group in.
9118                  */
9119                 if (device->total_bytes > device->bytes_used + min_free &&
9120                     !device->is_tgtdev_for_dev_replace) {
9121                         ret = find_free_dev_extent(trans, device, min_free,
9122                                                    &dev_offset, NULL);
9123                         if (!ret)
9124                                 dev_nr++;
9125
9126                         if (dev_nr >= dev_min)
9127                                 break;
9128
9129                         ret = -1;
9130                 }
9131         }
9132         mutex_unlock(&root->fs_info->chunk_mutex);
9133         btrfs_end_transaction(trans, root);
9134 out:
9135         btrfs_put_block_group(block_group);
9136         return ret;
9137 }
9138
9139 static int find_first_block_group(struct btrfs_root *root,
9140                 struct btrfs_path *path, struct btrfs_key *key)
9141 {
9142         int ret = 0;
9143         struct btrfs_key found_key;
9144         struct extent_buffer *leaf;
9145         int slot;
9146
9147         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9148         if (ret < 0)
9149                 goto out;
9150
9151         while (1) {
9152                 slot = path->slots[0];
9153                 leaf = path->nodes[0];
9154                 if (slot >= btrfs_header_nritems(leaf)) {
9155                         ret = btrfs_next_leaf(root, path);
9156                         if (ret == 0)
9157                                 continue;
9158                         if (ret < 0)
9159                                 goto out;
9160                         break;
9161                 }
9162                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9163
9164                 if (found_key.objectid >= key->objectid &&
9165                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9166                         ret = 0;
9167                         goto out;
9168                 }
9169                 path->slots[0]++;
9170         }
9171 out:
9172         return ret;
9173 }
9174
9175 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9176 {
9177         struct btrfs_block_group_cache *block_group;
9178         u64 last = 0;
9179
9180         while (1) {
9181                 struct inode *inode;
9182
9183                 block_group = btrfs_lookup_first_block_group(info, last);
9184                 while (block_group) {
9185                         spin_lock(&block_group->lock);
9186                         if (block_group->iref)
9187                                 break;
9188                         spin_unlock(&block_group->lock);
9189                         block_group = next_block_group(info->tree_root,
9190                                                        block_group);
9191                 }
9192                 if (!block_group) {
9193                         if (last == 0)
9194                                 break;
9195                         last = 0;
9196                         continue;
9197                 }
9198
9199                 inode = block_group->inode;
9200                 block_group->iref = 0;
9201                 block_group->inode = NULL;
9202                 spin_unlock(&block_group->lock);
9203                 iput(inode);
9204                 last = block_group->key.objectid + block_group->key.offset;
9205                 btrfs_put_block_group(block_group);
9206         }
9207 }
9208
9209 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9210 {
9211         struct btrfs_block_group_cache *block_group;
9212         struct btrfs_space_info *space_info;
9213         struct btrfs_caching_control *caching_ctl;
9214         struct rb_node *n;
9215
9216         down_write(&info->commit_root_sem);
9217         while (!list_empty(&info->caching_block_groups)) {
9218                 caching_ctl = list_entry(info->caching_block_groups.next,
9219                                          struct btrfs_caching_control, list);
9220                 list_del(&caching_ctl->list);
9221                 put_caching_control(caching_ctl);
9222         }
9223         up_write(&info->commit_root_sem);
9224
9225         spin_lock(&info->unused_bgs_lock);
9226         while (!list_empty(&info->unused_bgs)) {
9227                 block_group = list_first_entry(&info->unused_bgs,
9228                                                struct btrfs_block_group_cache,
9229                                                bg_list);
9230                 list_del_init(&block_group->bg_list);
9231                 btrfs_put_block_group(block_group);
9232         }
9233         spin_unlock(&info->unused_bgs_lock);
9234
9235         spin_lock(&info->block_group_cache_lock);
9236         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9237                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9238                                        cache_node);
9239                 rb_erase(&block_group->cache_node,
9240                          &info->block_group_cache_tree);
9241                 RB_CLEAR_NODE(&block_group->cache_node);
9242                 spin_unlock(&info->block_group_cache_lock);
9243
9244                 down_write(&block_group->space_info->groups_sem);
9245                 list_del(&block_group->list);
9246                 up_write(&block_group->space_info->groups_sem);
9247
9248                 if (block_group->cached == BTRFS_CACHE_STARTED)
9249                         wait_block_group_cache_done(block_group);
9250
9251                 /*
9252                  * We haven't cached this block group, which means we could
9253                  * possibly have excluded extents on this block group.
9254                  */
9255                 if (block_group->cached == BTRFS_CACHE_NO ||
9256                     block_group->cached == BTRFS_CACHE_ERROR)
9257                         free_excluded_extents(info->extent_root, block_group);
9258
9259                 btrfs_remove_free_space_cache(block_group);
9260                 btrfs_put_block_group(block_group);
9261
9262                 spin_lock(&info->block_group_cache_lock);
9263         }
9264         spin_unlock(&info->block_group_cache_lock);
9265
9266         /* now that all the block groups are freed, go through and
9267          * free all the space_info structs.  This is only called during
9268          * the final stages of unmount, and so we know nobody is
9269          * using them.  We call synchronize_rcu() once before we start,
9270          * just to be on the safe side.
9271          */
9272         synchronize_rcu();
9273
9274         release_global_block_rsv(info);
9275
9276         while (!list_empty(&info->space_info)) {
9277                 int i;
9278
9279                 space_info = list_entry(info->space_info.next,
9280                                         struct btrfs_space_info,
9281                                         list);
9282                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9283                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9284                             space_info->bytes_reserved > 0 ||
9285                             space_info->bytes_may_use > 0)) {
9286                                 dump_space_info(space_info, 0, 0);
9287                         }
9288                 }
9289                 list_del(&space_info->list);
9290                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9291                         struct kobject *kobj;
9292                         kobj = space_info->block_group_kobjs[i];
9293                         space_info->block_group_kobjs[i] = NULL;
9294                         if (kobj) {
9295                                 kobject_del(kobj);
9296                                 kobject_put(kobj);
9297                         }
9298                 }
9299                 kobject_del(&space_info->kobj);
9300                 kobject_put(&space_info->kobj);
9301         }
9302         return 0;
9303 }
9304
9305 static void __link_block_group(struct btrfs_space_info *space_info,
9306                                struct btrfs_block_group_cache *cache)
9307 {
9308         int index = get_block_group_index(cache);
9309         bool first = false;
9310
9311         down_write(&space_info->groups_sem);
9312         if (list_empty(&space_info->block_groups[index]))
9313                 first = true;
9314         list_add_tail(&cache->list, &space_info->block_groups[index]);
9315         up_write(&space_info->groups_sem);
9316
9317         if (first) {
9318                 struct raid_kobject *rkobj;
9319                 int ret;
9320
9321                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9322                 if (!rkobj)
9323                         goto out_err;
9324                 rkobj->raid_type = index;
9325                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9326                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9327                                   "%s", get_raid_name(index));
9328                 if (ret) {
9329                         kobject_put(&rkobj->kobj);
9330                         goto out_err;
9331                 }
9332                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9333         }
9334
9335         return;
9336 out_err:
9337         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9338 }
9339
9340 static struct btrfs_block_group_cache *
9341 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9342 {
9343         struct btrfs_block_group_cache *cache;
9344
9345         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9346         if (!cache)
9347                 return NULL;
9348
9349         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9350                                         GFP_NOFS);
9351         if (!cache->free_space_ctl) {
9352                 kfree(cache);
9353                 return NULL;
9354         }
9355
9356         cache->key.objectid = start;
9357         cache->key.offset = size;
9358         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9359
9360         cache->sectorsize = root->sectorsize;
9361         cache->fs_info = root->fs_info;
9362         cache->full_stripe_len = btrfs_full_stripe_len(root,
9363                                                &root->fs_info->mapping_tree,
9364                                                start);
9365         atomic_set(&cache->count, 1);
9366         spin_lock_init(&cache->lock);
9367         init_rwsem(&cache->data_rwsem);
9368         INIT_LIST_HEAD(&cache->list);
9369         INIT_LIST_HEAD(&cache->cluster_list);
9370         INIT_LIST_HEAD(&cache->bg_list);
9371         INIT_LIST_HEAD(&cache->ro_list);
9372         INIT_LIST_HEAD(&cache->dirty_list);
9373         INIT_LIST_HEAD(&cache->io_list);
9374         btrfs_init_free_space_ctl(cache);
9375         atomic_set(&cache->trimming, 0);
9376
9377         return cache;
9378 }
9379
9380 int btrfs_read_block_groups(struct btrfs_root *root)
9381 {
9382         struct btrfs_path *path;
9383         int ret;
9384         struct btrfs_block_group_cache *cache;
9385         struct btrfs_fs_info *info = root->fs_info;
9386         struct btrfs_space_info *space_info;
9387         struct btrfs_key key;
9388         struct btrfs_key found_key;
9389         struct extent_buffer *leaf;
9390         int need_clear = 0;
9391         u64 cache_gen;
9392
9393         root = info->extent_root;
9394         key.objectid = 0;
9395         key.offset = 0;
9396         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9397         path = btrfs_alloc_path();
9398         if (!path)
9399                 return -ENOMEM;
9400         path->reada = 1;
9401
9402         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9403         if (btrfs_test_opt(root, SPACE_CACHE) &&
9404             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9405                 need_clear = 1;
9406         if (btrfs_test_opt(root, CLEAR_CACHE))
9407                 need_clear = 1;
9408
9409         while (1) {
9410                 ret = find_first_block_group(root, path, &key);
9411                 if (ret > 0)
9412                         break;
9413                 if (ret != 0)
9414                         goto error;
9415
9416                 leaf = path->nodes[0];
9417                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9418
9419                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9420                                                        found_key.offset);
9421                 if (!cache) {
9422                         ret = -ENOMEM;
9423                         goto error;
9424                 }
9425
9426                 if (need_clear) {
9427                         /*
9428                          * When we mount with old space cache, we need to
9429                          * set BTRFS_DC_CLEAR and set dirty flag.
9430                          *
9431                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9432                          *    truncate the old free space cache inode and
9433                          *    setup a new one.
9434                          * b) Setting 'dirty flag' makes sure that we flush
9435                          *    the new space cache info onto disk.
9436                          */
9437                         if (btrfs_test_opt(root, SPACE_CACHE))
9438                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9439                 }
9440
9441                 read_extent_buffer(leaf, &cache->item,
9442                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9443                                    sizeof(cache->item));
9444                 cache->flags = btrfs_block_group_flags(&cache->item);
9445
9446                 key.objectid = found_key.objectid + found_key.offset;
9447                 btrfs_release_path(path);
9448
9449                 /*
9450                  * We need to exclude the super stripes now so that the space
9451                  * info has super bytes accounted for, otherwise we'll think
9452                  * we have more space than we actually do.
9453                  */
9454                 ret = exclude_super_stripes(root, cache);
9455                 if (ret) {
9456                         /*
9457                          * We may have excluded something, so call this just in
9458                          * case.
9459                          */
9460                         free_excluded_extents(root, cache);
9461                         btrfs_put_block_group(cache);
9462                         goto error;
9463                 }
9464
9465                 /*
9466                  * check for two cases, either we are full, and therefore
9467                  * don't need to bother with the caching work since we won't
9468                  * find any space, or we are empty, and we can just add all
9469                  * the space in and be done with it.  This saves us _alot_ of
9470                  * time, particularly in the full case.
9471                  */
9472                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9473                         cache->last_byte_to_unpin = (u64)-1;
9474                         cache->cached = BTRFS_CACHE_FINISHED;
9475                         free_excluded_extents(root, cache);
9476                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9477                         cache->last_byte_to_unpin = (u64)-1;
9478                         cache->cached = BTRFS_CACHE_FINISHED;
9479                         add_new_free_space(cache, root->fs_info,
9480                                            found_key.objectid,
9481                                            found_key.objectid +
9482                                            found_key.offset);
9483                         free_excluded_extents(root, cache);
9484                 }
9485
9486                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9487                 if (ret) {
9488                         btrfs_remove_free_space_cache(cache);
9489                         btrfs_put_block_group(cache);
9490                         goto error;
9491                 }
9492
9493                 ret = update_space_info(info, cache->flags, found_key.offset,
9494                                         btrfs_block_group_used(&cache->item),
9495                                         &space_info);
9496                 if (ret) {
9497                         btrfs_remove_free_space_cache(cache);
9498                         spin_lock(&info->block_group_cache_lock);
9499                         rb_erase(&cache->cache_node,
9500                                  &info->block_group_cache_tree);
9501                         RB_CLEAR_NODE(&cache->cache_node);
9502                         spin_unlock(&info->block_group_cache_lock);
9503                         btrfs_put_block_group(cache);
9504                         goto error;
9505                 }
9506
9507                 cache->space_info = space_info;
9508                 spin_lock(&cache->space_info->lock);
9509                 cache->space_info->bytes_readonly += cache->bytes_super;
9510                 spin_unlock(&cache->space_info->lock);
9511
9512                 __link_block_group(space_info, cache);
9513
9514                 set_avail_alloc_bits(root->fs_info, cache->flags);
9515                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9516                         inc_block_group_ro(cache, 1);
9517                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9518                         spin_lock(&info->unused_bgs_lock);
9519                         /* Should always be true but just in case. */
9520                         if (list_empty(&cache->bg_list)) {
9521                                 btrfs_get_block_group(cache);
9522                                 list_add_tail(&cache->bg_list,
9523                                               &info->unused_bgs);
9524                         }
9525                         spin_unlock(&info->unused_bgs_lock);
9526                 }
9527         }
9528
9529         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9530                 if (!(get_alloc_profile(root, space_info->flags) &
9531                       (BTRFS_BLOCK_GROUP_RAID10 |
9532                        BTRFS_BLOCK_GROUP_RAID1 |
9533                        BTRFS_BLOCK_GROUP_RAID5 |
9534                        BTRFS_BLOCK_GROUP_RAID6 |
9535                        BTRFS_BLOCK_GROUP_DUP)))
9536                         continue;
9537                 /*
9538                  * avoid allocating from un-mirrored block group if there are
9539                  * mirrored block groups.
9540                  */
9541                 list_for_each_entry(cache,
9542                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9543                                 list)
9544                         inc_block_group_ro(cache, 1);
9545                 list_for_each_entry(cache,
9546                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9547                                 list)
9548                         inc_block_group_ro(cache, 1);
9549         }
9550
9551         init_global_block_rsv(info);
9552         ret = 0;
9553 error:
9554         btrfs_free_path(path);
9555         return ret;
9556 }
9557
9558 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9559                                        struct btrfs_root *root)
9560 {
9561         struct btrfs_block_group_cache *block_group, *tmp;
9562         struct btrfs_root *extent_root = root->fs_info->extent_root;
9563         struct btrfs_block_group_item item;
9564         struct btrfs_key key;
9565         int ret = 0;
9566         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
9567
9568         trans->can_flush_pending_bgs = false;
9569         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9570                 if (ret)
9571                         goto next;
9572
9573                 spin_lock(&block_group->lock);
9574                 memcpy(&item, &block_group->item, sizeof(item));
9575                 memcpy(&key, &block_group->key, sizeof(key));
9576                 spin_unlock(&block_group->lock);
9577
9578                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9579                                         sizeof(item));
9580                 if (ret)
9581                         btrfs_abort_transaction(trans, extent_root, ret);
9582                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9583                                                key.objectid, key.offset);
9584                 if (ret)
9585                         btrfs_abort_transaction(trans, extent_root, ret);
9586 next:
9587                 list_del_init(&block_group->bg_list);
9588         }
9589         trans->can_flush_pending_bgs = can_flush_pending_bgs;
9590 }
9591
9592 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9593                            struct btrfs_root *root, u64 bytes_used,
9594                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9595                            u64 size)
9596 {
9597         int ret;
9598         struct btrfs_root *extent_root;
9599         struct btrfs_block_group_cache *cache;
9600
9601         extent_root = root->fs_info->extent_root;
9602
9603         btrfs_set_log_full_commit(root->fs_info, trans);
9604
9605         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9606         if (!cache)
9607                 return -ENOMEM;
9608
9609         btrfs_set_block_group_used(&cache->item, bytes_used);
9610         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9611         btrfs_set_block_group_flags(&cache->item, type);
9612
9613         cache->flags = type;
9614         cache->last_byte_to_unpin = (u64)-1;
9615         cache->cached = BTRFS_CACHE_FINISHED;
9616         ret = exclude_super_stripes(root, cache);
9617         if (ret) {
9618                 /*
9619                  * We may have excluded something, so call this just in
9620                  * case.
9621                  */
9622                 free_excluded_extents(root, cache);
9623                 btrfs_put_block_group(cache);
9624                 return ret;
9625         }
9626
9627         add_new_free_space(cache, root->fs_info, chunk_offset,
9628                            chunk_offset + size);
9629
9630         free_excluded_extents(root, cache);
9631
9632         /*
9633          * Call to ensure the corresponding space_info object is created and
9634          * assigned to our block group, but don't update its counters just yet.
9635          * We want our bg to be added to the rbtree with its ->space_info set.
9636          */
9637         ret = update_space_info(root->fs_info, cache->flags, 0, 0,
9638                                 &cache->space_info);
9639         if (ret) {
9640                 btrfs_remove_free_space_cache(cache);
9641                 btrfs_put_block_group(cache);
9642                 return ret;
9643         }
9644
9645         ret = btrfs_add_block_group_cache(root->fs_info, cache);
9646         if (ret) {
9647                 btrfs_remove_free_space_cache(cache);
9648                 btrfs_put_block_group(cache);
9649                 return ret;
9650         }
9651
9652         /*
9653          * Now that our block group has its ->space_info set and is inserted in
9654          * the rbtree, update the space info's counters.
9655          */
9656         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
9657                                 &cache->space_info);
9658         if (ret) {
9659                 btrfs_remove_free_space_cache(cache);
9660                 spin_lock(&root->fs_info->block_group_cache_lock);
9661                 rb_erase(&cache->cache_node,
9662                          &root->fs_info->block_group_cache_tree);
9663                 RB_CLEAR_NODE(&cache->cache_node);
9664                 spin_unlock(&root->fs_info->block_group_cache_lock);
9665                 btrfs_put_block_group(cache);
9666                 return ret;
9667         }
9668         update_global_block_rsv(root->fs_info);
9669
9670         spin_lock(&cache->space_info->lock);
9671         cache->space_info->bytes_readonly += cache->bytes_super;
9672         spin_unlock(&cache->space_info->lock);
9673
9674         __link_block_group(cache->space_info, cache);
9675
9676         list_add_tail(&cache->bg_list, &trans->new_bgs);
9677
9678         set_avail_alloc_bits(extent_root->fs_info, type);
9679
9680         return 0;
9681 }
9682
9683 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9684 {
9685         u64 extra_flags = chunk_to_extended(flags) &
9686                                 BTRFS_EXTENDED_PROFILE_MASK;
9687
9688         write_seqlock(&fs_info->profiles_lock);
9689         if (flags & BTRFS_BLOCK_GROUP_DATA)
9690                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9691         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9692                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9693         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9694                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9695         write_sequnlock(&fs_info->profiles_lock);
9696 }
9697
9698 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9699                              struct btrfs_root *root, u64 group_start,
9700                              struct extent_map *em)
9701 {
9702         struct btrfs_path *path;
9703         struct btrfs_block_group_cache *block_group;
9704         struct btrfs_free_cluster *cluster;
9705         struct btrfs_root *tree_root = root->fs_info->tree_root;
9706         struct btrfs_key key;
9707         struct inode *inode;
9708         struct kobject *kobj = NULL;
9709         int ret;
9710         int index;
9711         int factor;
9712         struct btrfs_caching_control *caching_ctl = NULL;
9713         bool remove_em;
9714
9715         root = root->fs_info->extent_root;
9716
9717         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9718         BUG_ON(!block_group);
9719         BUG_ON(!block_group->ro);
9720
9721         /*
9722          * Free the reserved super bytes from this block group before
9723          * remove it.
9724          */
9725         free_excluded_extents(root, block_group);
9726
9727         memcpy(&key, &block_group->key, sizeof(key));
9728         index = get_block_group_index(block_group);
9729         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9730                                   BTRFS_BLOCK_GROUP_RAID1 |
9731                                   BTRFS_BLOCK_GROUP_RAID10))
9732                 factor = 2;
9733         else
9734                 factor = 1;
9735
9736         /* make sure this block group isn't part of an allocation cluster */
9737         cluster = &root->fs_info->data_alloc_cluster;
9738         spin_lock(&cluster->refill_lock);
9739         btrfs_return_cluster_to_free_space(block_group, cluster);
9740         spin_unlock(&cluster->refill_lock);
9741
9742         /*
9743          * make sure this block group isn't part of a metadata
9744          * allocation cluster
9745          */
9746         cluster = &root->fs_info->meta_alloc_cluster;
9747         spin_lock(&cluster->refill_lock);
9748         btrfs_return_cluster_to_free_space(block_group, cluster);
9749         spin_unlock(&cluster->refill_lock);
9750
9751         path = btrfs_alloc_path();
9752         if (!path) {
9753                 ret = -ENOMEM;
9754                 goto out;
9755         }
9756
9757         /*
9758          * get the inode first so any iput calls done for the io_list
9759          * aren't the final iput (no unlinks allowed now)
9760          */
9761         inode = lookup_free_space_inode(tree_root, block_group, path);
9762
9763         mutex_lock(&trans->transaction->cache_write_mutex);
9764         /*
9765          * make sure our free spache cache IO is done before remove the
9766          * free space inode
9767          */
9768         spin_lock(&trans->transaction->dirty_bgs_lock);
9769         if (!list_empty(&block_group->io_list)) {
9770                 list_del_init(&block_group->io_list);
9771
9772                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
9773
9774                 spin_unlock(&trans->transaction->dirty_bgs_lock);
9775                 btrfs_wait_cache_io(root, trans, block_group,
9776                                     &block_group->io_ctl, path,
9777                                     block_group->key.objectid);
9778                 btrfs_put_block_group(block_group);
9779                 spin_lock(&trans->transaction->dirty_bgs_lock);
9780         }
9781
9782         if (!list_empty(&block_group->dirty_list)) {
9783                 list_del_init(&block_group->dirty_list);
9784                 btrfs_put_block_group(block_group);
9785         }
9786         spin_unlock(&trans->transaction->dirty_bgs_lock);
9787         mutex_unlock(&trans->transaction->cache_write_mutex);
9788
9789         if (!IS_ERR(inode)) {
9790                 ret = btrfs_orphan_add(trans, inode);
9791                 if (ret) {
9792                         btrfs_add_delayed_iput(inode);
9793                         goto out;
9794                 }
9795                 clear_nlink(inode);
9796                 /* One for the block groups ref */
9797                 spin_lock(&block_group->lock);
9798                 if (block_group->iref) {
9799                         block_group->iref = 0;
9800                         block_group->inode = NULL;
9801                         spin_unlock(&block_group->lock);
9802                         iput(inode);
9803                 } else {
9804                         spin_unlock(&block_group->lock);
9805                 }
9806                 /* One for our lookup ref */
9807                 btrfs_add_delayed_iput(inode);
9808         }
9809
9810         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9811         key.offset = block_group->key.objectid;
9812         key.type = 0;
9813
9814         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9815         if (ret < 0)
9816                 goto out;
9817         if (ret > 0)
9818                 btrfs_release_path(path);
9819         if (ret == 0) {
9820                 ret = btrfs_del_item(trans, tree_root, path);
9821                 if (ret)
9822                         goto out;
9823                 btrfs_release_path(path);
9824         }
9825
9826         spin_lock(&root->fs_info->block_group_cache_lock);
9827         rb_erase(&block_group->cache_node,
9828                  &root->fs_info->block_group_cache_tree);
9829         RB_CLEAR_NODE(&block_group->cache_node);
9830
9831         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9832                 root->fs_info->first_logical_byte = (u64)-1;
9833         spin_unlock(&root->fs_info->block_group_cache_lock);
9834
9835         down_write(&block_group->space_info->groups_sem);
9836         /*
9837          * we must use list_del_init so people can check to see if they
9838          * are still on the list after taking the semaphore
9839          */
9840         list_del_init(&block_group->list);
9841         if (list_empty(&block_group->space_info->block_groups[index])) {
9842                 kobj = block_group->space_info->block_group_kobjs[index];
9843                 block_group->space_info->block_group_kobjs[index] = NULL;
9844                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9845         }
9846         up_write(&block_group->space_info->groups_sem);
9847         if (kobj) {
9848                 kobject_del(kobj);
9849                 kobject_put(kobj);
9850         }
9851
9852         if (block_group->has_caching_ctl)
9853                 caching_ctl = get_caching_control(block_group);
9854         if (block_group->cached == BTRFS_CACHE_STARTED)
9855                 wait_block_group_cache_done(block_group);
9856         if (block_group->has_caching_ctl) {
9857                 down_write(&root->fs_info->commit_root_sem);
9858                 if (!caching_ctl) {
9859                         struct btrfs_caching_control *ctl;
9860
9861                         list_for_each_entry(ctl,
9862                                     &root->fs_info->caching_block_groups, list)
9863                                 if (ctl->block_group == block_group) {
9864                                         caching_ctl = ctl;
9865                                         atomic_inc(&caching_ctl->count);
9866                                         break;
9867                                 }
9868                 }
9869                 if (caching_ctl)
9870                         list_del_init(&caching_ctl->list);
9871                 up_write(&root->fs_info->commit_root_sem);
9872                 if (caching_ctl) {
9873                         /* Once for the caching bgs list and once for us. */
9874                         put_caching_control(caching_ctl);
9875                         put_caching_control(caching_ctl);
9876                 }
9877         }
9878
9879         spin_lock(&trans->transaction->dirty_bgs_lock);
9880         if (!list_empty(&block_group->dirty_list)) {
9881                 WARN_ON(1);
9882         }
9883         if (!list_empty(&block_group->io_list)) {
9884                 WARN_ON(1);
9885         }
9886         spin_unlock(&trans->transaction->dirty_bgs_lock);
9887         btrfs_remove_free_space_cache(block_group);
9888
9889         spin_lock(&block_group->space_info->lock);
9890         list_del_init(&block_group->ro_list);
9891
9892         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
9893                 WARN_ON(block_group->space_info->total_bytes
9894                         < block_group->key.offset);
9895                 WARN_ON(block_group->space_info->bytes_readonly
9896                         < block_group->key.offset);
9897                 WARN_ON(block_group->space_info->disk_total
9898                         < block_group->key.offset * factor);
9899         }
9900         block_group->space_info->total_bytes -= block_group->key.offset;
9901         block_group->space_info->bytes_readonly -= block_group->key.offset;
9902         block_group->space_info->disk_total -= block_group->key.offset * factor;
9903
9904         spin_unlock(&block_group->space_info->lock);
9905
9906         memcpy(&key, &block_group->key, sizeof(key));
9907
9908         lock_chunks(root);
9909         if (!list_empty(&em->list)) {
9910                 /* We're in the transaction->pending_chunks list. */
9911                 free_extent_map(em);
9912         }
9913         spin_lock(&block_group->lock);
9914         block_group->removed = 1;
9915         /*
9916          * At this point trimming can't start on this block group, because we
9917          * removed the block group from the tree fs_info->block_group_cache_tree
9918          * so no one can't find it anymore and even if someone already got this
9919          * block group before we removed it from the rbtree, they have already
9920          * incremented block_group->trimming - if they didn't, they won't find
9921          * any free space entries because we already removed them all when we
9922          * called btrfs_remove_free_space_cache().
9923          *
9924          * And we must not remove the extent map from the fs_info->mapping_tree
9925          * to prevent the same logical address range and physical device space
9926          * ranges from being reused for a new block group. This is because our
9927          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
9928          * completely transactionless, so while it is trimming a range the
9929          * currently running transaction might finish and a new one start,
9930          * allowing for new block groups to be created that can reuse the same
9931          * physical device locations unless we take this special care.
9932          *
9933          * There may also be an implicit trim operation if the file system
9934          * is mounted with -odiscard. The same protections must remain
9935          * in place until the extents have been discarded completely when
9936          * the transaction commit has completed.
9937          */
9938         remove_em = (atomic_read(&block_group->trimming) == 0);
9939         /*
9940          * Make sure a trimmer task always sees the em in the pinned_chunks list
9941          * if it sees block_group->removed == 1 (needs to lock block_group->lock
9942          * before checking block_group->removed).
9943          */
9944         if (!remove_em) {
9945                 /*
9946                  * Our em might be in trans->transaction->pending_chunks which
9947                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
9948                  * and so is the fs_info->pinned_chunks list.
9949                  *
9950                  * So at this point we must be holding the chunk_mutex to avoid
9951                  * any races with chunk allocation (more specifically at
9952                  * volumes.c:contains_pending_extent()), to ensure it always
9953                  * sees the em, either in the pending_chunks list or in the
9954                  * pinned_chunks list.
9955                  */
9956                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
9957         }
9958         spin_unlock(&block_group->lock);
9959
9960         if (remove_em) {
9961                 struct extent_map_tree *em_tree;
9962
9963                 em_tree = &root->fs_info->mapping_tree.map_tree;
9964                 write_lock(&em_tree->lock);
9965                 /*
9966                  * The em might be in the pending_chunks list, so make sure the
9967                  * chunk mutex is locked, since remove_extent_mapping() will
9968                  * delete us from that list.
9969                  */
9970                 remove_extent_mapping(em_tree, em);
9971                 write_unlock(&em_tree->lock);
9972                 /* once for the tree */
9973                 free_extent_map(em);
9974         }
9975
9976         unlock_chunks(root);
9977
9978         btrfs_put_block_group(block_group);
9979         btrfs_put_block_group(block_group);
9980
9981         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9982         if (ret > 0)
9983                 ret = -EIO;
9984         if (ret < 0)
9985                 goto out;
9986
9987         ret = btrfs_del_item(trans, root, path);
9988 out:
9989         btrfs_free_path(path);
9990         return ret;
9991 }
9992
9993 /*
9994  * Process the unused_bgs list and remove any that don't have any allocated
9995  * space inside of them.
9996  */
9997 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
9998 {
9999         struct btrfs_block_group_cache *block_group;
10000         struct btrfs_space_info *space_info;
10001         struct btrfs_root *root = fs_info->extent_root;
10002         struct btrfs_trans_handle *trans;
10003         int ret = 0;
10004
10005         if (!fs_info->open)
10006                 return;
10007
10008         spin_lock(&fs_info->unused_bgs_lock);
10009         while (!list_empty(&fs_info->unused_bgs)) {
10010                 u64 start, end;
10011                 int trimming;
10012
10013                 block_group = list_first_entry(&fs_info->unused_bgs,
10014                                                struct btrfs_block_group_cache,
10015                                                bg_list);
10016                 space_info = block_group->space_info;
10017                 list_del_init(&block_group->bg_list);
10018                 if (ret || btrfs_mixed_space_info(space_info)) {
10019                         btrfs_put_block_group(block_group);
10020                         continue;
10021                 }
10022                 spin_unlock(&fs_info->unused_bgs_lock);
10023
10024                 mutex_lock(&root->fs_info->delete_unused_bgs_mutex);
10025
10026                 /* Don't want to race with allocators so take the groups_sem */
10027                 down_write(&space_info->groups_sem);
10028                 spin_lock(&block_group->lock);
10029                 if (block_group->reserved ||
10030                     btrfs_block_group_used(&block_group->item) ||
10031                     block_group->ro) {
10032                         /*
10033                          * We want to bail if we made new allocations or have
10034                          * outstanding allocations in this block group.  We do
10035                          * the ro check in case balance is currently acting on
10036                          * this block group.
10037                          */
10038                         spin_unlock(&block_group->lock);
10039                         up_write(&space_info->groups_sem);
10040                         goto next;
10041                 }
10042                 spin_unlock(&block_group->lock);
10043
10044                 /* We don't want to force the issue, only flip if it's ok. */
10045                 ret = inc_block_group_ro(block_group, 0);
10046                 up_write(&space_info->groups_sem);
10047                 if (ret < 0) {
10048                         ret = 0;
10049                         goto next;
10050                 }
10051
10052                 /*
10053                  * Want to do this before we do anything else so we can recover
10054                  * properly if we fail to join the transaction.
10055                  */
10056                 /* 1 for btrfs_orphan_reserve_metadata() */
10057                 trans = btrfs_start_transaction(root, 1);
10058                 if (IS_ERR(trans)) {
10059                         btrfs_dec_block_group_ro(root, block_group);
10060                         ret = PTR_ERR(trans);
10061                         goto next;
10062                 }
10063
10064                 /*
10065                  * We could have pending pinned extents for this block group,
10066                  * just delete them, we don't care about them anymore.
10067                  */
10068                 start = block_group->key.objectid;
10069                 end = start + block_group->key.offset - 1;
10070                 /*
10071                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10072                  * btrfs_finish_extent_commit(). If we are at transaction N,
10073                  * another task might be running finish_extent_commit() for the
10074                  * previous transaction N - 1, and have seen a range belonging
10075                  * to the block group in freed_extents[] before we were able to
10076                  * clear the whole block group range from freed_extents[]. This
10077                  * means that task can lookup for the block group after we
10078                  * unpinned it from freed_extents[] and removed it, leading to
10079                  * a BUG_ON() at btrfs_unpin_extent_range().
10080                  */
10081                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10082                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10083                                   EXTENT_DIRTY, GFP_NOFS);
10084                 if (ret) {
10085                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10086                         btrfs_dec_block_group_ro(root, block_group);
10087                         goto end_trans;
10088                 }
10089                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10090                                   EXTENT_DIRTY, GFP_NOFS);
10091                 if (ret) {
10092                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10093                         btrfs_dec_block_group_ro(root, block_group);
10094                         goto end_trans;
10095                 }
10096                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10097
10098                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10099                 spin_lock(&space_info->lock);
10100                 spin_lock(&block_group->lock);
10101
10102                 space_info->bytes_pinned -= block_group->pinned;
10103                 space_info->bytes_readonly += block_group->pinned;
10104                 percpu_counter_add(&space_info->total_bytes_pinned,
10105                                    -block_group->pinned);
10106                 block_group->pinned = 0;
10107
10108                 spin_unlock(&block_group->lock);
10109                 spin_unlock(&space_info->lock);
10110
10111                 /* DISCARD can flip during remount */
10112                 trimming = btrfs_test_opt(root, DISCARD);
10113
10114                 /* Implicit trim during transaction commit. */
10115                 if (trimming)
10116                         btrfs_get_block_group_trimming(block_group);
10117
10118                 /*
10119                  * Btrfs_remove_chunk will abort the transaction if things go
10120                  * horribly wrong.
10121                  */
10122                 ret = btrfs_remove_chunk(trans, root,
10123                                          block_group->key.objectid);
10124
10125                 if (ret) {
10126                         if (trimming)
10127                                 btrfs_put_block_group_trimming(block_group);
10128                         goto end_trans;
10129                 }
10130
10131                 /*
10132                  * If we're not mounted with -odiscard, we can just forget
10133                  * about this block group. Otherwise we'll need to wait
10134                  * until transaction commit to do the actual discard.
10135                  */
10136                 if (trimming) {
10137                         WARN_ON(!list_empty(&block_group->bg_list));
10138                         spin_lock(&trans->transaction->deleted_bgs_lock);
10139                         list_move(&block_group->bg_list,
10140                                   &trans->transaction->deleted_bgs);
10141                         spin_unlock(&trans->transaction->deleted_bgs_lock);
10142                         btrfs_get_block_group(block_group);
10143                 }
10144 end_trans:
10145                 btrfs_end_transaction(trans, root);
10146 next:
10147                 mutex_unlock(&root->fs_info->delete_unused_bgs_mutex);
10148                 btrfs_put_block_group(block_group);
10149                 spin_lock(&fs_info->unused_bgs_lock);
10150         }
10151         spin_unlock(&fs_info->unused_bgs_lock);
10152 }
10153
10154 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10155 {
10156         struct btrfs_space_info *space_info;
10157         struct btrfs_super_block *disk_super;
10158         u64 features;
10159         u64 flags;
10160         int mixed = 0;
10161         int ret;
10162
10163         disk_super = fs_info->super_copy;
10164         if (!btrfs_super_root(disk_super))
10165                 return 1;
10166
10167         features = btrfs_super_incompat_flags(disk_super);
10168         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10169                 mixed = 1;
10170
10171         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10172         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10173         if (ret)
10174                 goto out;
10175
10176         if (mixed) {
10177                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10178                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10179         } else {
10180                 flags = BTRFS_BLOCK_GROUP_METADATA;
10181                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10182                 if (ret)
10183                         goto out;
10184
10185                 flags = BTRFS_BLOCK_GROUP_DATA;
10186                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10187         }
10188 out:
10189         return ret;
10190 }
10191
10192 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10193 {
10194         return unpin_extent_range(root, start, end, false);
10195 }
10196
10197 /*
10198  * It used to be that old block groups would be left around forever.
10199  * Iterating over them would be enough to trim unused space.  Since we
10200  * now automatically remove them, we also need to iterate over unallocated
10201  * space.
10202  *
10203  * We don't want a transaction for this since the discard may take a
10204  * substantial amount of time.  We don't require that a transaction be
10205  * running, but we do need to take a running transaction into account
10206  * to ensure that we're not discarding chunks that were released in
10207  * the current transaction.
10208  *
10209  * Holding the chunks lock will prevent other threads from allocating
10210  * or releasing chunks, but it won't prevent a running transaction
10211  * from committing and releasing the memory that the pending chunks
10212  * list head uses.  For that, we need to take a reference to the
10213  * transaction.
10214  */
10215 static int btrfs_trim_free_extents(struct btrfs_device *device,
10216                                    u64 minlen, u64 *trimmed)
10217 {
10218         u64 start = 0, len = 0;
10219         int ret;
10220
10221         *trimmed = 0;
10222
10223         /* Not writeable = nothing to do. */
10224         if (!device->writeable)
10225                 return 0;
10226
10227         /* No free space = nothing to do. */
10228         if (device->total_bytes <= device->bytes_used)
10229                 return 0;
10230
10231         ret = 0;
10232
10233         while (1) {
10234                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
10235                 struct btrfs_transaction *trans;
10236                 u64 bytes;
10237
10238                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10239                 if (ret)
10240                         return ret;
10241
10242                 down_read(&fs_info->commit_root_sem);
10243
10244                 spin_lock(&fs_info->trans_lock);
10245                 trans = fs_info->running_transaction;
10246                 if (trans)
10247                         atomic_inc(&trans->use_count);
10248                 spin_unlock(&fs_info->trans_lock);
10249
10250                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10251                                                  &start, &len);
10252                 if (trans)
10253                         btrfs_put_transaction(trans);
10254
10255                 if (ret) {
10256                         up_read(&fs_info->commit_root_sem);
10257                         mutex_unlock(&fs_info->chunk_mutex);
10258                         if (ret == -ENOSPC)
10259                                 ret = 0;
10260                         break;
10261                 }
10262
10263                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10264                 up_read(&fs_info->commit_root_sem);
10265                 mutex_unlock(&fs_info->chunk_mutex);
10266
10267                 if (ret)
10268                         break;
10269
10270                 start += len;
10271                 *trimmed += bytes;
10272
10273                 if (fatal_signal_pending(current)) {
10274                         ret = -ERESTARTSYS;
10275                         break;
10276                 }
10277
10278                 cond_resched();
10279         }
10280
10281         return ret;
10282 }
10283
10284 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10285 {
10286         struct btrfs_fs_info *fs_info = root->fs_info;
10287         struct btrfs_block_group_cache *cache = NULL;
10288         struct btrfs_device *device;
10289         struct list_head *devices;
10290         u64 group_trimmed;
10291         u64 start;
10292         u64 end;
10293         u64 trimmed = 0;
10294         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10295         int ret = 0;
10296
10297         /*
10298          * try to trim all FS space, our block group may start from non-zero.
10299          */
10300         if (range->len == total_bytes)
10301                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10302         else
10303                 cache = btrfs_lookup_block_group(fs_info, range->start);
10304
10305         while (cache) {
10306                 if (cache->key.objectid >= (range->start + range->len)) {
10307                         btrfs_put_block_group(cache);
10308                         break;
10309                 }
10310
10311                 start = max(range->start, cache->key.objectid);
10312                 end = min(range->start + range->len,
10313                                 cache->key.objectid + cache->key.offset);
10314
10315                 if (end - start >= range->minlen) {
10316                         if (!block_group_cache_done(cache)) {
10317                                 ret = cache_block_group(cache, 0);
10318                                 if (ret) {
10319                                         btrfs_put_block_group(cache);
10320                                         break;
10321                                 }
10322                                 ret = wait_block_group_cache_done(cache);
10323                                 if (ret) {
10324                                         btrfs_put_block_group(cache);
10325                                         break;
10326                                 }
10327                         }
10328                         ret = btrfs_trim_block_group(cache,
10329                                                      &group_trimmed,
10330                                                      start,
10331                                                      end,
10332                                                      range->minlen);
10333
10334                         trimmed += group_trimmed;
10335                         if (ret) {
10336                                 btrfs_put_block_group(cache);
10337                                 break;
10338                         }
10339                 }
10340
10341                 cache = next_block_group(fs_info->tree_root, cache);
10342         }
10343
10344         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
10345         devices = &root->fs_info->fs_devices->alloc_list;
10346         list_for_each_entry(device, devices, dev_alloc_list) {
10347                 ret = btrfs_trim_free_extents(device, range->minlen,
10348                                               &group_trimmed);
10349                 if (ret)
10350                         break;
10351
10352                 trimmed += group_trimmed;
10353         }
10354         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
10355
10356         range->len = trimmed;
10357         return ret;
10358 }
10359
10360 /*
10361  * btrfs_{start,end}_write_no_snapshoting() are similar to
10362  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10363  * data into the page cache through nocow before the subvolume is snapshoted,
10364  * but flush the data into disk after the snapshot creation, or to prevent
10365  * operations while snapshoting is ongoing and that cause the snapshot to be
10366  * inconsistent (writes followed by expanding truncates for example).
10367  */
10368 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10369 {
10370         percpu_counter_dec(&root->subv_writers->counter);
10371         /*
10372          * Make sure counter is updated before we wake up waiters.
10373          */
10374         smp_mb();
10375         if (waitqueue_active(&root->subv_writers->wait))
10376                 wake_up(&root->subv_writers->wait);
10377 }
10378
10379 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10380 {
10381         if (atomic_read(&root->will_be_snapshoted))
10382                 return 0;
10383
10384         percpu_counter_inc(&root->subv_writers->counter);
10385         /*
10386          * Make sure counter is updated before we check for snapshot creation.
10387          */
10388         smp_mb();
10389         if (atomic_read(&root->will_be_snapshoted)) {
10390                 btrfs_end_write_no_snapshoting(root);
10391                 return 0;
10392         }
10393         return 1;
10394 }