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