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 "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "raid56.h"
  35 #include "locking.h"
  36 #include "free-space-cache.h"
  37 #include "math.h"
  38 #include "sysfs.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 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  86                                     struct extent_buffer *leaf,
  87                                     struct btrfs_extent_item *ei);
  88 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  89                                       struct btrfs_root *root,
  90                                       u64 parent, u64 root_objectid,
  91                                       u64 flags, u64 owner, u64 offset,
  92                                       struct btrfs_key *ins, int ref_mod);
  93 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  94                                      struct btrfs_root *root,
  95                                      u64 parent, u64 root_objectid,
  96                                      u64 flags, struct btrfs_disk_key *key,
  97                                      int level, struct btrfs_key *ins);
  98 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  99                           struct btrfs_root *extent_root, u64 flags,
 100                           int force);
 101 static int find_next_key(struct btrfs_path *path, int level,
 102                          struct btrfs_key *key);
 103 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 104                             int dump_block_groups);
 105 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 106                                        u64 num_bytes, int reserve);
 107 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 108                                u64 num_bytes);
 109 int btrfs_pin_extent(struct btrfs_root *root,
 110                      u64 bytenr, u64 num_bytes, int reserved);
 111
 112 static noinline int
 113 block_group_cache_done(struct btrfs_block_group_cache *cache)
 114 {
 115         smp_mb();
 116         return cache->cached == BTRFS_CACHE_FINISHED ||
 117                 cache->cached == BTRFS_CACHE_ERROR;
 118 }
 119
 120 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 121 {
 122         return (cache->flags & bits) == bits;
 123 }
 124
 125 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 126 {
 127         atomic_inc(&cache->count);
 128 }
 129
 130 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 131 {
 132         if (atomic_dec_and_test(&cache->count)) {
 133                 WARN_ON(cache->pinned > 0);
 134                 WARN_ON(cache->reserved > 0);
 135                 kfree(cache->free_space_ctl);
 136                 kfree(cache);
 137         }
 138 }
 139
 140 /*
 141  * this adds the block group to the fs_info rb tree for the block group
 142  * cache
 143  */
 144 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 145                                 struct btrfs_block_group_cache *block_group)
 146 {
 147         struct rb_node **p;
 148         struct rb_node *parent = NULL;
 149         struct btrfs_block_group_cache *cache;
 150
 151         spin_lock(&info->block_group_cache_lock);
 152         p = &info->block_group_cache_tree.rb_node;
 153
 154         while (*p) {
 155                 parent = *p;
 156                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 157                                  cache_node);
 158                 if (block_group->key.objectid < cache->key.objectid) {
 159                         p = &(*p)->rb_left;
 160                 } else if (block_group->key.objectid > cache->key.objectid) {
 161                         p = &(*p)->rb_right;
 162                 } else {
 163                         spin_unlock(&info->block_group_cache_lock);
 164                         return -EEXIST;
 165                 }
 166         }
 167
 168         rb_link_node(&block_group->cache_node, parent, p);
 169         rb_insert_color(&block_group->cache_node,
 170                         &info->block_group_cache_tree);
 171
 172         if (info->first_logical_byte > block_group->key.objectid)
 173                 info->first_logical_byte = block_group->key.objectid;
 174
 175         spin_unlock(&info->block_group_cache_lock);
 176
 177         return 0;
 178 }
 179
 180 /*
 181  * This will return the block group at or after bytenr if contains is 0, else
 182  * it will return the block group that contains the bytenr
 183  */
 184 static struct btrfs_block_group_cache *
 185 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 186                               int contains)
 187 {
 188         struct btrfs_block_group_cache *cache, *ret = NULL;
 189         struct rb_node *n;
 190         u64 end, start;
 191
 192         spin_lock(&info->block_group_cache_lock);
 193         n = info->block_group_cache_tree.rb_node;
 194
 195         while (n) {
 196                 cache = rb_entry(n, struct btrfs_block_group_cache,
 197                                  cache_node);
 198                 end = cache->key.objectid + cache->key.offset - 1;
 199                 start = cache->key.objectid;
 200
 201                 if (bytenr < start) {
 202                         if (!contains && (!ret || start < ret->key.objectid))
 203                                 ret = cache;
 204                         n = n->rb_left;
 205                 } else if (bytenr > start) {
 206                         if (contains && bytenr <= end) {
 207                                 ret = cache;
 208                                 break;
 209                         }
 210                         n = n->rb_right;
 211                 } else {
 212                         ret = cache;
 213                         break;
 214                 }
 215         }
 216         if (ret) {
 217                 btrfs_get_block_group(ret);
 218                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 219                         info->first_logical_byte = ret->key.objectid;
 220         }
 221         spin_unlock(&info->block_group_cache_lock);
 222
 223         return ret;
 224 }
 225
 226 static int add_excluded_extent(struct btrfs_root *root,
 227                                u64 start, u64 num_bytes)
 228 {
 229         u64 end = start + num_bytes - 1;
 230         set_extent_bits(&root->fs_info->freed_extents[0],
 231                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 232         set_extent_bits(&root->fs_info->freed_extents[1],
 233                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 234         return 0;
 235 }
 236
 237 static void free_excluded_extents(struct btrfs_root *root,
 238                                   struct btrfs_block_group_cache *cache)
 239 {
 240         u64 start, end;
 241
 242         start = cache->key.objectid;
 243         end = start + cache->key.offset - 1;
 244
 245         clear_extent_bits(&root->fs_info->freed_extents[0],
 246                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 247         clear_extent_bits(&root->fs_info->freed_extents[1],
 248                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 249 }
 250
 251 static int exclude_super_stripes(struct btrfs_root *root,
 252                                  struct btrfs_block_group_cache *cache)
 253 {
 254         u64 bytenr;
 255         u64 *logical;
 256         int stripe_len;
 257         int i, nr, ret;
 258
 259         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 260                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 261                 cache->bytes_super += stripe_len;
 262                 ret = add_excluded_extent(root, cache->key.objectid,
 263                                           stripe_len);
 264                 if (ret)
 265                         return ret;
 266         }
 267
 268         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 269                 bytenr = btrfs_sb_offset(i);
 270                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 271                                        cache->key.objectid, bytenr,
 272                                        0, &logical, &nr, &stripe_len);
 273                 if (ret)
 274                         return ret;
 275
 276                 while (nr--) {
 277                         u64 start, len;
 278
 279                         if (logical[nr] > cache->key.objectid +
 280                             cache->key.offset)
 281                                 continue;
 282
 283                         if (logical[nr] + stripe_len <= cache->key.objectid)
 284                                 continue;
 285
 286                         start = logical[nr];
 287                         if (start < cache->key.objectid) {
 288                                 start = cache->key.objectid;
 289                                 len = (logical[nr] + stripe_len) - start;
 290                         } else {
 291                                 len = min_t(u64, stripe_len,
 292                                             cache->key.objectid +
 293                                             cache->key.offset - start);
 294                         }
 295
 296                         cache->bytes_super += len;
 297                         ret = add_excluded_extent(root, start, len);
 298                         if (ret) {
 299                                 kfree(logical);
 300                                 return ret;
 301                         }
 302                 }
 303
 304                 kfree(logical);
 305         }
 306         return 0;
 307 }
 308
 309 static struct btrfs_caching_control *
 310 get_caching_control(struct btrfs_block_group_cache *cache)
 311 {
 312         struct btrfs_caching_control *ctl;
 313
 314         spin_lock(&cache->lock);
 315         if (cache->cached != BTRFS_CACHE_STARTED) {
 316                 spin_unlock(&cache->lock);
 317                 return NULL;
 318         }
 319
 320         /* We're loading it the fast way, so we don't have a caching_ctl. */
 321         if (!cache->caching_ctl) {
 322                 spin_unlock(&cache->lock);
 323                 return NULL;
 324         }
 325
 326         ctl = cache->caching_ctl;
 327         atomic_inc(&ctl->count);
 328         spin_unlock(&cache->lock);
 329         return ctl;
 330 }
 331
 332 static void put_caching_control(struct btrfs_caching_control *ctl)
 333 {
 334         if (atomic_dec_and_test(&ctl->count))
 335                 kfree(ctl);
 336 }
 337
 338 /*
 339  * this is only called by cache_block_group, since we could have freed extents
 340  * we need to check the pinned_extents for any extents that can't be used yet
 341  * since their free space will be released as soon as the transaction commits.
 342  */
 343 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 344                               struct btrfs_fs_info *info, u64 start, u64 end)
 345 {
 346         u64 extent_start, extent_end, size, total_added = 0;
 347         int ret;
 348
 349         while (start < end) {
 350                 ret = find_first_extent_bit(info->pinned_extents, start,
 351                                             &extent_start, &extent_end,
 352                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 353                                             NULL);
 354                 if (ret)
 355                         break;
 356
 357                 if (extent_start <= start) {
 358                         start = extent_end + 1;
 359                 } else if (extent_start > start && extent_start < end) {
 360                         size = extent_start - start;
 361                         total_added += size;
 362                         ret = btrfs_add_free_space(block_group, start,
 363                                                    size);
 364                         BUG_ON(ret); /* -ENOMEM or logic error */
 365                         start = extent_end + 1;
 366                 } else {
 367                         break;
 368                 }
 369         }
 370
 371         if (start < end) {
 372                 size = end - start;
 373                 total_added += size;
 374                 ret = btrfs_add_free_space(block_group, start, size);
 375                 BUG_ON(ret); /* -ENOMEM or logic error */
 376         }
 377
 378         return total_added;
 379 }
 380
 381 static noinline void caching_thread(struct btrfs_work *work)
 382 {
 383         struct btrfs_block_group_cache *block_group;
 384         struct btrfs_fs_info *fs_info;
 385         struct btrfs_caching_control *caching_ctl;
 386         struct btrfs_root *extent_root;
 387         struct btrfs_path *path;
 388         struct extent_buffer *leaf;
 389         struct btrfs_key key;
 390         u64 total_found = 0;
 391         u64 last = 0;
 392         u32 nritems;
 393         int ret = -ENOMEM;
 394
 395         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 396         block_group = caching_ctl->block_group;
 397         fs_info = block_group->fs_info;
 398         extent_root = fs_info->extent_root;
 399
 400         path = btrfs_alloc_path();
 401         if (!path)
 402                 goto out;
 403
 404         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 405
 406         /*
 407          * We don't want to deadlock with somebody trying to allocate a new
 408          * extent for the extent root while also trying to search the extent
 409          * root to add free space.  So we skip locking and search the commit
 410          * root, since its read-only
 411          */
 412         path->skip_locking = 1;
 413         path->search_commit_root = 1;
 414         path->reada = 1;
 415
 416         key.objectid = last;
 417         key.offset = 0;
 418         key.type = BTRFS_EXTENT_ITEM_KEY;
 419 again:
 420         mutex_lock(&caching_ctl->mutex);
 421         /* need to make sure the commit_root doesn't disappear */
 422         down_read(&fs_info->extent_commit_sem);
 423
 424 next:
 425         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 426         if (ret < 0)
 427                 goto err;
 428
 429         leaf = path->nodes[0];
 430         nritems = btrfs_header_nritems(leaf);
 431
 432         while (1) {
 433                 if (btrfs_fs_closing(fs_info) > 1) {
 434                         last = (u64)-1;
 435                         break;
 436                 }
 437
 438                 if (path->slots[0] < nritems) {
 439                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 440                 } else {
 441                         ret = find_next_key(path, 0, &key);
 442                         if (ret)
 443                                 break;
 444
 445                         if (need_resched()) {
 446                                 caching_ctl->progress = last;
 447                                 btrfs_release_path(path);
 448                                 up_read(&fs_info->extent_commit_sem);
 449                                 mutex_unlock(&caching_ctl->mutex);
 450                                 cond_resched();
 451                                 goto again;
 452                         }
 453
 454                         ret = btrfs_next_leaf(extent_root, path);
 455                         if (ret < 0)
 456                                 goto err;
 457                         if (ret)
 458                                 break;
 459                         leaf = path->nodes[0];
 460                         nritems = btrfs_header_nritems(leaf);
 461                         continue;
 462                 }
 463
 464                 if (key.objectid < last) {
 465                         key.objectid = last;
 466                         key.offset = 0;
 467                         key.type = BTRFS_EXTENT_ITEM_KEY;
 468
 469                         caching_ctl->progress = last;
 470                         btrfs_release_path(path);
 471                         goto next;
 472                 }
 473
 474                 if (key.objectid < block_group->key.objectid) {
 475                         path->slots[0]++;
 476                         continue;
 477                 }
 478
 479                 if (key.objectid >= block_group->key.objectid +
 480                     block_group->key.offset)
 481                         break;
 482
 483                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 484                     key.type == BTRFS_METADATA_ITEM_KEY) {
 485                         total_found += add_new_free_space(block_group,
 486                                                           fs_info, last,
 487                                                           key.objectid);
 488                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 489                                 last = key.objectid +
 490                                         fs_info->tree_root->leafsize;
 491                         else
 492                                 last = key.objectid + key.offset;
 493
 494                         if (total_found > (1024 * 1024 * 2)) {
 495                                 total_found = 0;
 496                                 wake_up(&caching_ctl->wait);
 497                         }
 498                 }
 499                 path->slots[0]++;
 500         }
 501         ret = 0;
 502
 503         total_found += add_new_free_space(block_group, fs_info, last,
 504                                           block_group->key.objectid +
 505                                           block_group->key.offset);
 506         caching_ctl->progress = (u64)-1;
 507
 508         spin_lock(&block_group->lock);
 509         block_group->caching_ctl = NULL;
 510         block_group->cached = BTRFS_CACHE_FINISHED;
 511         spin_unlock(&block_group->lock);
 512
 513 err:
 514         btrfs_free_path(path);
 515         up_read(&fs_info->extent_commit_sem);
 516
 517         free_excluded_extents(extent_root, block_group);
 518
 519         mutex_unlock(&caching_ctl->mutex);
 520 out:
 521         if (ret) {
 522                 spin_lock(&block_group->lock);
 523                 block_group->caching_ctl = NULL;
 524                 block_group->cached = BTRFS_CACHE_ERROR;
 525                 spin_unlock(&block_group->lock);
 526         }
 527         wake_up(&caching_ctl->wait);
 528
 529         put_caching_control(caching_ctl);
 530         btrfs_put_block_group(block_group);
 531 }
 532
 533 static int cache_block_group(struct btrfs_block_group_cache *cache,
 534                              int load_cache_only)
 535 {
 536         DEFINE_WAIT(wait);
 537         struct btrfs_fs_info *fs_info = cache->fs_info;
 538         struct btrfs_caching_control *caching_ctl;
 539         int ret = 0;
 540
 541         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 542         if (!caching_ctl)
 543                 return -ENOMEM;
 544
 545         INIT_LIST_HEAD(&caching_ctl->list);
 546         mutex_init(&caching_ctl->mutex);
 547         init_waitqueue_head(&caching_ctl->wait);
 548         caching_ctl->block_group = cache;
 549         caching_ctl->progress = cache->key.objectid;
 550         atomic_set(&caching_ctl->count, 1);
 551         caching_ctl->work.func = caching_thread;
 552
 553         spin_lock(&cache->lock);
 554         /*
 555          * This should be a rare occasion, but this could happen I think in the
 556          * case where one thread starts to load the space cache info, and then
 557          * some other thread starts a transaction commit which tries to do an
 558          * allocation while the other thread is still loading the space cache
 559          * info.  The previous loop should have kept us from choosing this block
 560          * group, but if we've moved to the state where we will wait on caching
 561          * block groups we need to first check if we're doing a fast load here,
 562          * so we can wait for it to finish, otherwise we could end up allocating
 563          * from a block group who's cache gets evicted for one reason or
 564          * another.
 565          */
 566         while (cache->cached == BTRFS_CACHE_FAST) {
 567                 struct btrfs_caching_control *ctl;
 568
 569                 ctl = cache->caching_ctl;
 570                 atomic_inc(&ctl->count);
 571                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 572                 spin_unlock(&cache->lock);
 573
 574                 schedule();
 575
 576                 finish_wait(&ctl->wait, &wait);
 577                 put_caching_control(ctl);
 578                 spin_lock(&cache->lock);
 579         }
 580
 581         if (cache->cached != BTRFS_CACHE_NO) {
 582                 spin_unlock(&cache->lock);
 583                 kfree(caching_ctl);
 584                 return 0;
 585         }
 586         WARN_ON(cache->caching_ctl);
 587         cache->caching_ctl = caching_ctl;
 588         cache->cached = BTRFS_CACHE_FAST;
 589         spin_unlock(&cache->lock);
 590
 591         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 592                 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                 } else {
 600                         if (load_cache_only) {
 601                                 cache->caching_ctl = NULL;
 602                                 cache->cached = BTRFS_CACHE_NO;
 603                         } else {
 604                                 cache->cached = BTRFS_CACHE_STARTED;
 605                         }
 606                 }
 607                 spin_unlock(&cache->lock);
 608                 wake_up(&caching_ctl->wait);
 609                 if (ret == 1) {
 610                         put_caching_control(caching_ctl);
 611                         free_excluded_extents(fs_info->extent_root, cache);
 612                         return 0;
 613                 }
 614         } else {
 615                 /*
 616                  * We are not going to do the fast caching, set cached to the
 617                  * appropriate value and wakeup any waiters.
 618                  */
 619                 spin_lock(&cache->lock);
 620                 if (load_cache_only) {
 621                         cache->caching_ctl = NULL;
 622                         cache->cached = BTRFS_CACHE_NO;
 623                 } else {
 624                         cache->cached = BTRFS_CACHE_STARTED;
 625                 }
 626                 spin_unlock(&cache->lock);
 627                 wake_up(&caching_ctl->wait);
 628         }
 629
 630         if (load_cache_only) {
 631                 put_caching_control(caching_ctl);
 632                 return 0;
 633         }
 634
 635         down_write(&fs_info->extent_commit_sem);
 636         atomic_inc(&caching_ctl->count);
 637         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 638         up_write(&fs_info->extent_commit_sem);
 639
 640         btrfs_get_block_group(cache);
 641
 642         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 643
 644         return ret;
 645 }
 646
 647 /*
 648  * return the block group that starts at or after bytenr
 649  */
 650 static struct btrfs_block_group_cache *
 651 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 652 {
 653         struct btrfs_block_group_cache *cache;
 654
 655         cache = block_group_cache_tree_search(info, bytenr, 0);
 656
 657         return cache;
 658 }
 659
 660 /*
 661  * return the block group that contains the given bytenr
 662  */
 663 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 664                                                  struct btrfs_fs_info *info,
 665                                                  u64 bytenr)
 666 {
 667         struct btrfs_block_group_cache *cache;
 668
 669         cache = block_group_cache_tree_search(info, bytenr, 1);
 670
 671         return cache;
 672 }
 673
 674 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 675                                                   u64 flags)
 676 {
 677         struct list_head *head = &info->space_info;
 678         struct btrfs_space_info *found;
 679
 680         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 681
 682         rcu_read_lock();
 683         list_for_each_entry_rcu(found, head, list) {
 684                 if (found->flags & flags) {
 685                         rcu_read_unlock();
 686                         return found;
 687                 }
 688         }
 689         rcu_read_unlock();
 690         return NULL;
 691 }
 692
 693 /*
 694  * after adding space to the filesystem, we need to clear the full flags
 695  * on all the space infos.
 696  */
 697 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 698 {
 699         struct list_head *head = &info->space_info;
 700         struct btrfs_space_info *found;
 701
 702         rcu_read_lock();
 703         list_for_each_entry_rcu(found, head, list)
 704                 found->full = 0;
 705         rcu_read_unlock();
 706 }
 707
 708 /* simple helper to search for an existing extent at a given offset */
 709 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 710 {
 711         int ret;
 712         struct btrfs_key key;
 713         struct btrfs_path *path;
 714
 715         path = btrfs_alloc_path();
 716         if (!path)
 717                 return -ENOMEM;
 718
 719         key.objectid = start;
 720         key.offset = len;
 721         key.type = BTRFS_EXTENT_ITEM_KEY;
 722         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 723                                 0, 0);
 724         if (ret > 0) {
 725                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 726                 if (key.objectid == start &&
 727                     key.type == BTRFS_METADATA_ITEM_KEY)
 728                         ret = 0;
 729         }
 730         btrfs_free_path(path);
 731         return ret;
 732 }
 733
 734 /*
 735  * helper function to lookup reference count and flags of a tree block.
 736  *
 737  * the head node for delayed ref is used to store the sum of all the
 738  * reference count modifications queued up in the rbtree. the head
 739  * node may also store the extent flags to set. This way you can check
 740  * to see what the reference count and extent flags would be if all of
 741  * the delayed refs are not processed.
 742  */
 743 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 744                              struct btrfs_root *root, u64 bytenr,
 745                              u64 offset, int metadata, u64 *refs, u64 *flags)
 746 {
 747         struct btrfs_delayed_ref_head *head;
 748         struct btrfs_delayed_ref_root *delayed_refs;
 749         struct btrfs_path *path;
 750         struct btrfs_extent_item *ei;
 751         struct extent_buffer *leaf;
 752         struct btrfs_key key;
 753         u32 item_size;
 754         u64 num_refs;
 755         u64 extent_flags;
 756         int ret;
 757
 758         /*
 759          * If we don't have skinny metadata, don't bother doing anything
 760          * different
 761          */
 762         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 763                 offset = root->leafsize;
 764                 metadata = 0;
 765         }
 766
 767         path = btrfs_alloc_path();
 768         if (!path)
 769                 return -ENOMEM;
 770
 771         if (!trans) {
 772                 path->skip_locking = 1;
 773                 path->search_commit_root = 1;
 774         }
 775
 776 search_again:
 777         key.objectid = bytenr;
 778         key.offset = offset;
 779         if (metadata)
 780                 key.type = BTRFS_METADATA_ITEM_KEY;
 781         else
 782                 key.type = BTRFS_EXTENT_ITEM_KEY;
 783
 784 again:
 785         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 786                                 &key, path, 0, 0);
 787         if (ret < 0)
 788                 goto out_free;
 789
 790         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 791                 if (path->slots[0]) {
 792                         path->slots[0]--;
 793                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 794                                               path->slots[0]);
 795                         if (key.objectid == bytenr &&
 796                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 797                             key.offset == root->leafsize)
 798                                 ret = 0;
 799                 }
 800                 if (ret) {
 801                         key.objectid = bytenr;
 802                         key.type = BTRFS_EXTENT_ITEM_KEY;
 803                         key.offset = root->leafsize;
 804                         btrfs_release_path(path);
 805                         goto again;
 806                 }
 807         }
 808
 809         if (ret == 0) {
 810                 leaf = path->nodes[0];
 811                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 812                 if (item_size >= sizeof(*ei)) {
 813                         ei = btrfs_item_ptr(leaf, path->slots[0],
 814                                             struct btrfs_extent_item);
 815                         num_refs = btrfs_extent_refs(leaf, ei);
 816                         extent_flags = btrfs_extent_flags(leaf, ei);
 817                 } else {
 818 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 819                         struct btrfs_extent_item_v0 *ei0;
 820                         BUG_ON(item_size != sizeof(*ei0));
 821                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 822                                              struct btrfs_extent_item_v0);
 823                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 824                         /* FIXME: this isn't correct for data */
 825                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 826 #else
 827                         BUG();
 828 #endif
 829                 }
 830                 BUG_ON(num_refs == 0);
 831         } else {
 832                 num_refs = 0;
 833                 extent_flags = 0;
 834                 ret = 0;
 835         }
 836
 837         if (!trans)
 838                 goto out;
 839
 840         delayed_refs = &trans->transaction->delayed_refs;
 841         spin_lock(&delayed_refs->lock);
 842         head = btrfs_find_delayed_ref_head(trans, bytenr);
 843         if (head) {
 844                 if (!mutex_trylock(&head->mutex)) {
 845                         atomic_inc(&head->node.refs);
 846                         spin_unlock(&delayed_refs->lock);
 847
 848                         btrfs_release_path(path);
 849
 850                         /*
 851                          * Mutex was contended, block until it's released and try
 852                          * again
 853                          */
 854                         mutex_lock(&head->mutex);
 855                         mutex_unlock(&head->mutex);
 856                         btrfs_put_delayed_ref(&head->node);
 857                         goto search_again;
 858                 }
 859                 if (head->extent_op && head->extent_op->update_flags)
 860                         extent_flags |= head->extent_op->flags_to_set;
 861                 else
 862                         BUG_ON(num_refs == 0);
 863
 864                 num_refs += head->node.ref_mod;
 865                 mutex_unlock(&head->mutex);
 866         }
 867         spin_unlock(&delayed_refs->lock);
 868 out:
 869         WARN_ON(num_refs == 0);
 870         if (refs)
 871                 *refs = num_refs;
 872         if (flags)
 873                 *flags = extent_flags;
 874 out_free:
 875         btrfs_free_path(path);
 876         return ret;
 877 }
 878
 879 /*
 880  * Back reference rules.  Back refs have three main goals:
 881  *
 882  * 1) differentiate between all holders of references to an extent so that
 883  *    when a reference is dropped we can make sure it was a valid reference
 884  *    before freeing the extent.
 885  *
 886  * 2) Provide enough information to quickly find the holders of an extent
 887  *    if we notice a given block is corrupted or bad.
 888  *
 889  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 890  *    maintenance.  This is actually the same as #2, but with a slightly
 891  *    different use case.
 892  *
 893  * There are two kinds of back refs. The implicit back refs is optimized
 894  * for pointers in non-shared tree blocks. For a given pointer in a block,
 895  * back refs of this kind provide information about the block's owner tree
 896  * and the pointer's key. These information allow us to find the block by
 897  * b-tree searching. The full back refs is for pointers in tree blocks not
 898  * referenced by their owner trees. The location of tree block is recorded
 899  * in the back refs. Actually the full back refs is generic, and can be
 900  * used in all cases the implicit back refs is used. The major shortcoming
 901  * of the full back refs is its overhead. Every time a tree block gets
 902  * COWed, we have to update back refs entry for all pointers in it.
 903  *
 904  * For a newly allocated tree block, we use implicit back refs for
 905  * pointers in it. This means most tree related operations only involve
 906  * implicit back refs. For a tree block created in old transaction, the
 907  * only way to drop a reference to it is COW it. So we can detect the
 908  * event that tree block loses its owner tree's reference and do the
 909  * back refs conversion.
 910  *
 911  * When a tree block is COW'd through a tree, there are four cases:
 912  *
 913  * The reference count of the block is one and the tree is the block's
 914  * owner tree. Nothing to do in this case.
 915  *
 916  * The reference count of the block is one and the tree is not the
 917  * block's owner tree. In this case, full back refs is used for pointers
 918  * in the block. Remove these full back refs, add implicit back refs for
 919  * every pointers in the new block.
 920  *
 921  * The reference count of the block is greater than one and the tree is
 922  * the block's owner tree. In this case, implicit back refs is used for
 923  * pointers in the block. Add full back refs for every pointers in the
 924  * block, increase lower level extents' reference counts. The original
 925  * implicit back refs are entailed to the new block.
 926  *
 927  * The reference count of the block is greater than one and the tree is
 928  * not the block's owner tree. Add implicit back refs for every pointer in
 929  * the new block, increase lower level extents' reference count.
 930  *
 931  * Back Reference Key composing:
 932  *
 933  * The key objectid corresponds to the first byte in the extent,
 934  * The key type is used to differentiate between types of back refs.
 935  * There are different meanings of the key offset for different types
 936  * of back refs.
 937  *
 938  * File extents can be referenced by:
 939  *
 940  * - multiple snapshots, subvolumes, or different generations in one subvol
 941  * - different files inside a single subvolume
 942  * - different offsets inside a file (bookend extents in file.c)
 943  *
 944  * The extent ref structure for the implicit back refs has fields for:
 945  *
 946  * - Objectid of the subvolume root
 947  * - objectid of the file holding the reference
 948  * - original offset in the file
 949  * - how many bookend extents
 950  *
 951  * The key offset for the implicit back refs is hash of the first
 952  * three fields.
 953  *
 954  * The extent ref structure for the full back refs has field for:
 955  *
 956  * - number of pointers in the tree leaf
 957  *
 958  * The key offset for the implicit back refs is the first byte of
 959  * the tree leaf
 960  *
 961  * When a file extent is allocated, The implicit back refs is used.
 962  * the fields are filled in:
 963  *
 964  *     (root_key.objectid, inode objectid, offset in file, 1)
 965  *
 966  * When a file extent is removed file truncation, we find the
 967  * corresponding implicit back refs and check the following fields:
 968  *
 969  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 970  *
 971  * Btree extents can be referenced by:
 972  *
 973  * - Different subvolumes
 974  *
 975  * Both the implicit back refs and the full back refs for tree blocks
 976  * only consist of key. The key offset for the implicit back refs is
 977  * objectid of block's owner tree. The key offset for the full back refs
 978  * is the first byte of parent block.
 979  *
 980  * When implicit back refs is used, information about the lowest key and
 981  * level of the tree block are required. These information are stored in
 982  * tree block info structure.
 983  */
 984
 985 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 986 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 987                                   struct btrfs_root *root,
 988                                   struct btrfs_path *path,
 989                                   u64 owner, u32 extra_size)
 990 {
 991         struct btrfs_extent_item *item;
 992         struct btrfs_extent_item_v0 *ei0;
 993         struct btrfs_extent_ref_v0 *ref0;
 994         struct btrfs_tree_block_info *bi;
 995         struct extent_buffer *leaf;
 996         struct btrfs_key key;
 997         struct btrfs_key found_key;
 998         u32 new_size = sizeof(*item);
 999         u64 refs;
1000         int ret;
1001
1002         leaf = path->nodes[0];
1003         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1004
1005         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1006         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1007                              struct btrfs_extent_item_v0);
1008         refs = btrfs_extent_refs_v0(leaf, ei0);
1009
1010         if (owner == (u64)-1) {
1011                 while (1) {
1012                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1013                                 ret = btrfs_next_leaf(root, path);
1014                                 if (ret < 0)
1015                                         return ret;
1016                                 BUG_ON(ret > 0); /* Corruption */
1017                                 leaf = path->nodes[0];
1018                         }
1019                         btrfs_item_key_to_cpu(leaf, &found_key,
1020                                               path->slots[0]);
1021                         BUG_ON(key.objectid != found_key.objectid);
1022                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1023                                 path->slots[0]++;
1024                                 continue;
1025                         }
1026                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1027                                               struct btrfs_extent_ref_v0);
1028                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1029                         break;
1030                 }
1031         }
1032         btrfs_release_path(path);
1033
1034         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1035                 new_size += sizeof(*bi);
1036
1037         new_size -= sizeof(*ei0);
1038         ret = btrfs_search_slot(trans, root, &key, path,
1039                                 new_size + extra_size, 1);
1040         if (ret < 0)
1041                 return ret;
1042         BUG_ON(ret); /* Corruption */
1043
1044         btrfs_extend_item(root, path, new_size);
1045
1046         leaf = path->nodes[0];
1047         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1048         btrfs_set_extent_refs(leaf, item, refs);
1049         /* FIXME: get real generation */
1050         btrfs_set_extent_generation(leaf, item, 0);
1051         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1052                 btrfs_set_extent_flags(leaf, item,
1053                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1054                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1055                 bi = (struct btrfs_tree_block_info *)(item + 1);
1056                 /* FIXME: get first key of the block */
1057                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1058                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1059         } else {
1060                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1061         }
1062         btrfs_mark_buffer_dirty(leaf);
1063         return 0;
1064 }
1065 #endif
1066
1067 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1068 {
1069         u32 high_crc = ~(u32)0;
1070         u32 low_crc = ~(u32)0;
1071         __le64 lenum;
1072
1073         lenum = cpu_to_le64(root_objectid);
1074         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1075         lenum = cpu_to_le64(owner);
1076         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1077         lenum = cpu_to_le64(offset);
1078         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1079
1080         return ((u64)high_crc << 31) ^ (u64)low_crc;
1081 }
1082
1083 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1084                                      struct btrfs_extent_data_ref *ref)
1085 {
1086         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1087                                     btrfs_extent_data_ref_objectid(leaf, ref),
1088                                     btrfs_extent_data_ref_offset(leaf, ref));
1089 }
1090
1091 static int match_extent_data_ref(struct extent_buffer *leaf,
1092                                  struct btrfs_extent_data_ref *ref,
1093                                  u64 root_objectid, u64 owner, u64 offset)
1094 {
1095         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1096             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1097             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1098                 return 0;
1099         return 1;
1100 }
1101
1102 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1103                                            struct btrfs_root *root,
1104                                            struct btrfs_path *path,
1105                                            u64 bytenr, u64 parent,
1106                                            u64 root_objectid,
1107                                            u64 owner, u64 offset)
1108 {
1109         struct btrfs_key key;
1110         struct btrfs_extent_data_ref *ref;
1111         struct extent_buffer *leaf;
1112         u32 nritems;
1113         int ret;
1114         int recow;
1115         int err = -ENOENT;
1116
1117         key.objectid = bytenr;
1118         if (parent) {
1119                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1120                 key.offset = parent;
1121         } else {
1122                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1123                 key.offset = hash_extent_data_ref(root_objectid,
1124                                                   owner, offset);
1125         }
1126 again:
1127         recow = 0;
1128         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1129         if (ret < 0) {
1130                 err = ret;
1131                 goto fail;
1132         }
1133
1134         if (parent) {
1135                 if (!ret)
1136                         return 0;
1137 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1138                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1139                 btrfs_release_path(path);
1140                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1141                 if (ret < 0) {
1142                         err = ret;
1143                         goto fail;
1144                 }
1145                 if (!ret)
1146                         return 0;
1147 #endif
1148                 goto fail;
1149         }
1150
1151         leaf = path->nodes[0];
1152         nritems = btrfs_header_nritems(leaf);
1153         while (1) {
1154                 if (path->slots[0] >= nritems) {
1155                         ret = btrfs_next_leaf(root, path);
1156                         if (ret < 0)
1157                                 err = ret;
1158                         if (ret)
1159                                 goto fail;
1160
1161                         leaf = path->nodes[0];
1162                         nritems = btrfs_header_nritems(leaf);
1163                         recow = 1;
1164                 }
1165
1166                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1167                 if (key.objectid != bytenr ||
1168                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1169                         goto fail;
1170
1171                 ref = btrfs_item_ptr(leaf, path->slots[0],
1172                                      struct btrfs_extent_data_ref);
1173
1174                 if (match_extent_data_ref(leaf, ref, root_objectid,
1175                                           owner, offset)) {
1176                         if (recow) {
1177                                 btrfs_release_path(path);
1178                                 goto again;
1179                         }
1180                         err = 0;
1181                         break;
1182                 }
1183                 path->slots[0]++;
1184         }
1185 fail:
1186         return err;
1187 }
1188
1189 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1190                                            struct btrfs_root *root,
1191                                            struct btrfs_path *path,
1192                                            u64 bytenr, u64 parent,
1193                                            u64 root_objectid, u64 owner,
1194                                            u64 offset, int refs_to_add)
1195 {
1196         struct btrfs_key key;
1197         struct extent_buffer *leaf;
1198         u32 size;
1199         u32 num_refs;
1200         int ret;
1201
1202         key.objectid = bytenr;
1203         if (parent) {
1204                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1205                 key.offset = parent;
1206                 size = sizeof(struct btrfs_shared_data_ref);
1207         } else {
1208                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1209                 key.offset = hash_extent_data_ref(root_objectid,
1210                                                   owner, offset);
1211                 size = sizeof(struct btrfs_extent_data_ref);
1212         }
1213
1214         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1215         if (ret && ret != -EEXIST)
1216                 goto fail;
1217
1218         leaf = path->nodes[0];
1219         if (parent) {
1220                 struct btrfs_shared_data_ref *ref;
1221                 ref = btrfs_item_ptr(leaf, path->slots[0],
1222                                      struct btrfs_shared_data_ref);
1223                 if (ret == 0) {
1224                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1225                 } else {
1226                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1227                         num_refs += refs_to_add;
1228                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1229                 }
1230         } else {
1231                 struct btrfs_extent_data_ref *ref;
1232                 while (ret == -EEXIST) {
1233                         ref = btrfs_item_ptr(leaf, path->slots[0],
1234                                              struct btrfs_extent_data_ref);
1235                         if (match_extent_data_ref(leaf, ref, root_objectid,
1236                                                   owner, offset))
1237                                 break;
1238                         btrfs_release_path(path);
1239                         key.offset++;
1240                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1241                                                       size);
1242                         if (ret && ret != -EEXIST)
1243                                 goto fail;
1244
1245                         leaf = path->nodes[0];
1246                 }
1247                 ref = btrfs_item_ptr(leaf, path->slots[0],
1248                                      struct btrfs_extent_data_ref);
1249                 if (ret == 0) {
1250                         btrfs_set_extent_data_ref_root(leaf, ref,
1251                                                        root_objectid);
1252                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1253                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1254                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1255                 } else {
1256                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1257                         num_refs += refs_to_add;
1258                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1259                 }
1260         }
1261         btrfs_mark_buffer_dirty(leaf);
1262         ret = 0;
1263 fail:
1264         btrfs_release_path(path);
1265         return ret;
1266 }
1267
1268 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1269                                            struct btrfs_root *root,
1270                                            struct btrfs_path *path,
1271                                            int refs_to_drop)
1272 {
1273         struct btrfs_key key;
1274         struct btrfs_extent_data_ref *ref1 = NULL;
1275         struct btrfs_shared_data_ref *ref2 = NULL;
1276         struct extent_buffer *leaf;
1277         u32 num_refs = 0;
1278         int ret = 0;
1279
1280         leaf = path->nodes[0];
1281         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1282
1283         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1284                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1285                                       struct btrfs_extent_data_ref);
1286                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1287         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1288                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1289                                       struct btrfs_shared_data_ref);
1290                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1291 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1292         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1293                 struct btrfs_extent_ref_v0 *ref0;
1294                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1295                                       struct btrfs_extent_ref_v0);
1296                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1297 #endif
1298         } else {
1299                 BUG();
1300         }
1301
1302         BUG_ON(num_refs < refs_to_drop);
1303         num_refs -= refs_to_drop;
1304
1305         if (num_refs == 0) {
1306                 ret = btrfs_del_item(trans, root, path);
1307         } else {
1308                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1309                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1310                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1311                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1312 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1313                 else {
1314                         struct btrfs_extent_ref_v0 *ref0;
1315                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1316                                         struct btrfs_extent_ref_v0);
1317                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1318                 }
1319 #endif
1320                 btrfs_mark_buffer_dirty(leaf);
1321         }
1322         return ret;
1323 }
1324
1325 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1326                                           struct btrfs_path *path,
1327                                           struct btrfs_extent_inline_ref *iref)
1328 {
1329         struct btrfs_key key;
1330         struct extent_buffer *leaf;
1331         struct btrfs_extent_data_ref *ref1;
1332         struct btrfs_shared_data_ref *ref2;
1333         u32 num_refs = 0;
1334
1335         leaf = path->nodes[0];
1336         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1337         if (iref) {
1338                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1339                     BTRFS_EXTENT_DATA_REF_KEY) {
1340                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1341                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1342                 } else {
1343                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1344                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1345                 }
1346         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1347                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1348                                       struct btrfs_extent_data_ref);
1349                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1350         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1351                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1352                                       struct btrfs_shared_data_ref);
1353                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1354 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1355         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1356                 struct btrfs_extent_ref_v0 *ref0;
1357                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1358                                       struct btrfs_extent_ref_v0);
1359                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1360 #endif
1361         } else {
1362                 WARN_ON(1);
1363         }
1364         return num_refs;
1365 }
1366
1367 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1368                                           struct btrfs_root *root,
1369                                           struct btrfs_path *path,
1370                                           u64 bytenr, u64 parent,
1371                                           u64 root_objectid)
1372 {
1373         struct btrfs_key key;
1374         int ret;
1375
1376         key.objectid = bytenr;
1377         if (parent) {
1378                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1379                 key.offset = parent;
1380         } else {
1381                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1382                 key.offset = root_objectid;
1383         }
1384
1385         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1386         if (ret > 0)
1387                 ret = -ENOENT;
1388 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1389         if (ret == -ENOENT && parent) {
1390                 btrfs_release_path(path);
1391                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1392                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1393                 if (ret > 0)
1394                         ret = -ENOENT;
1395         }
1396 #endif
1397         return ret;
1398 }
1399
1400 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1401                                           struct btrfs_root *root,
1402                                           struct btrfs_path *path,
1403                                           u64 bytenr, u64 parent,
1404                                           u64 root_objectid)
1405 {
1406         struct btrfs_key key;
1407         int ret;
1408
1409         key.objectid = bytenr;
1410         if (parent) {
1411                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1412                 key.offset = parent;
1413         } else {
1414                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1415                 key.offset = root_objectid;
1416         }
1417
1418         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1419         btrfs_release_path(path);
1420         return ret;
1421 }
1422
1423 static inline int extent_ref_type(u64 parent, u64 owner)
1424 {
1425         int type;
1426         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1427                 if (parent > 0)
1428                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1429                 else
1430                         type = BTRFS_TREE_BLOCK_REF_KEY;
1431         } else {
1432                 if (parent > 0)
1433                         type = BTRFS_SHARED_DATA_REF_KEY;
1434                 else
1435                         type = BTRFS_EXTENT_DATA_REF_KEY;
1436         }
1437         return type;
1438 }
1439
1440 static int find_next_key(struct btrfs_path *path, int level,
1441                          struct btrfs_key *key)
1442
1443 {
1444         for (; level < BTRFS_MAX_LEVEL; level++) {
1445                 if (!path->nodes[level])
1446                         break;
1447                 if (path->slots[level] + 1 >=
1448                     btrfs_header_nritems(path->nodes[level]))
1449                         continue;
1450                 if (level == 0)
1451                         btrfs_item_key_to_cpu(path->nodes[level], key,
1452                                               path->slots[level] + 1);
1453                 else
1454                         btrfs_node_key_to_cpu(path->nodes[level], key,
1455                                               path->slots[level] + 1);
1456                 return 0;
1457         }
1458         return 1;
1459 }
1460
1461 /*
1462  * look for inline back ref. if back ref is found, *ref_ret is set
1463  * to the address of inline back ref, and 0 is returned.
1464  *
1465  * if back ref isn't found, *ref_ret is set to the address where it
1466  * should be inserted, and -ENOENT is returned.
1467  *
1468  * if insert is true and there are too many inline back refs, the path
1469  * points to the extent item, and -EAGAIN is returned.
1470  *
1471  * NOTE: inline back refs are ordered in the same way that back ref
1472  *       items in the tree are ordered.
1473  */
1474 static noinline_for_stack
1475 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1476                                  struct btrfs_root *root,
1477                                  struct btrfs_path *path,
1478                                  struct btrfs_extent_inline_ref **ref_ret,
1479                                  u64 bytenr, u64 num_bytes,
1480                                  u64 parent, u64 root_objectid,
1481                                  u64 owner, u64 offset, int insert)
1482 {
1483         struct btrfs_key key;
1484         struct extent_buffer *leaf;
1485         struct btrfs_extent_item *ei;
1486         struct btrfs_extent_inline_ref *iref;
1487         u64 flags;
1488         u64 item_size;
1489         unsigned long ptr;
1490         unsigned long end;
1491         int extra_size;
1492         int type;
1493         int want;
1494         int ret;
1495         int err = 0;
1496         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1497                                                  SKINNY_METADATA);
1498
1499         key.objectid = bytenr;
1500         key.type = BTRFS_EXTENT_ITEM_KEY;
1501         key.offset = num_bytes;
1502
1503         want = extent_ref_type(parent, owner);
1504         if (insert) {
1505                 extra_size = btrfs_extent_inline_ref_size(want);
1506                 path->keep_locks = 1;
1507         } else
1508                 extra_size = -1;
1509
1510         /*
1511          * Owner is our parent level, so we can just add one to get the level
1512          * for the block we are interested in.
1513          */
1514         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1515                 key.type = BTRFS_METADATA_ITEM_KEY;
1516                 key.offset = owner;
1517         }
1518
1519 again:
1520         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1521         if (ret < 0) {
1522                 err = ret;
1523                 goto out;
1524         }
1525
1526         /*
1527          * We may be a newly converted file system which still has the old fat
1528          * extent entries for metadata, so try and see if we have one of those.
1529          */
1530         if (ret > 0 && skinny_metadata) {
1531                 skinny_metadata = false;
1532                 if (path->slots[0]) {
1533                         path->slots[0]--;
1534                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1535                                               path->slots[0]);
1536                         if (key.objectid == bytenr &&
1537                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1538                             key.offset == num_bytes)
1539                                 ret = 0;
1540                 }
1541                 if (ret) {
1542                         key.type = BTRFS_EXTENT_ITEM_KEY;
1543                         key.offset = num_bytes;
1544                         btrfs_release_path(path);
1545                         goto again;
1546                 }
1547         }
1548
1549         if (ret && !insert) {
1550                 err = -ENOENT;
1551                 goto out;
1552         } else if (WARN_ON(ret)) {
1553                 err = -EIO;
1554                 goto out;
1555         }
1556
1557         leaf = path->nodes[0];
1558         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1559 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1560         if (item_size < sizeof(*ei)) {
1561                 if (!insert) {
1562                         err = -ENOENT;
1563                         goto out;
1564                 }
1565                 ret = convert_extent_item_v0(trans, root, path, owner,
1566                                              extra_size);
1567                 if (ret < 0) {
1568                         err = ret;
1569                         goto out;
1570                 }
1571                 leaf = path->nodes[0];
1572                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1573         }
1574 #endif
1575         BUG_ON(item_size < sizeof(*ei));
1576
1577         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1578         flags = btrfs_extent_flags(leaf, ei);
1579
1580         ptr = (unsigned long)(ei + 1);
1581         end = (unsigned long)ei + item_size;
1582
1583         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1584                 ptr += sizeof(struct btrfs_tree_block_info);
1585                 BUG_ON(ptr > end);
1586         }
1587
1588         err = -ENOENT;
1589         while (1) {
1590                 if (ptr >= end) {
1591                         WARN_ON(ptr > end);
1592                         break;
1593                 }
1594                 iref = (struct btrfs_extent_inline_ref *)ptr;
1595                 type = btrfs_extent_inline_ref_type(leaf, iref);
1596                 if (want < type)
1597                         break;
1598                 if (want > type) {
1599                         ptr += btrfs_extent_inline_ref_size(type);
1600                         continue;
1601                 }
1602
1603                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1604                         struct btrfs_extent_data_ref *dref;
1605                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1606                         if (match_extent_data_ref(leaf, dref, root_objectid,
1607                                                   owner, offset)) {
1608                                 err = 0;
1609                                 break;
1610                         }
1611                         if (hash_extent_data_ref_item(leaf, dref) <
1612                             hash_extent_data_ref(root_objectid, owner, offset))
1613                                 break;
1614                 } else {
1615                         u64 ref_offset;
1616                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1617                         if (parent > 0) {
1618                                 if (parent == ref_offset) {
1619                                         err = 0;
1620                                         break;
1621                                 }
1622                                 if (ref_offset < parent)
1623                                         break;
1624                         } else {
1625                                 if (root_objectid == ref_offset) {
1626                                         err = 0;
1627                                         break;
1628                                 }
1629                                 if (ref_offset < root_objectid)
1630                                         break;
1631                         }
1632                 }
1633                 ptr += btrfs_extent_inline_ref_size(type);
1634         }
1635         if (err == -ENOENT && insert) {
1636                 if (item_size + extra_size >=
1637                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1638                         err = -EAGAIN;
1639                         goto out;
1640                 }
1641                 /*
1642                  * To add new inline back ref, we have to make sure
1643                  * there is no corresponding back ref item.
1644                  * For simplicity, we just do not add new inline back
1645                  * ref if there is any kind of item for this block
1646                  */
1647                 if (find_next_key(path, 0, &key) == 0 &&
1648                     key.objectid == bytenr &&
1649                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1650                         err = -EAGAIN;
1651                         goto out;
1652                 }
1653         }
1654         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1655 out:
1656         if (insert) {
1657                 path->keep_locks = 0;
1658                 btrfs_unlock_up_safe(path, 1);
1659         }
1660         return err;
1661 }
1662
1663 /*
1664  * helper to add new inline back ref
1665  */
1666 static noinline_for_stack
1667 void setup_inline_extent_backref(struct btrfs_root *root,
1668                                  struct btrfs_path *path,
1669                                  struct btrfs_extent_inline_ref *iref,
1670                                  u64 parent, u64 root_objectid,
1671                                  u64 owner, u64 offset, int refs_to_add,
1672                                  struct btrfs_delayed_extent_op *extent_op)
1673 {
1674         struct extent_buffer *leaf;
1675         struct btrfs_extent_item *ei;
1676         unsigned long ptr;
1677         unsigned long end;
1678         unsigned long item_offset;
1679         u64 refs;
1680         int size;
1681         int type;
1682
1683         leaf = path->nodes[0];
1684         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1685         item_offset = (unsigned long)iref - (unsigned long)ei;
1686
1687         type = extent_ref_type(parent, owner);
1688         size = btrfs_extent_inline_ref_size(type);
1689
1690         btrfs_extend_item(root, path, size);
1691
1692         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1693         refs = btrfs_extent_refs(leaf, ei);
1694         refs += refs_to_add;
1695         btrfs_set_extent_refs(leaf, ei, refs);
1696         if (extent_op)
1697                 __run_delayed_extent_op(extent_op, leaf, ei);
1698
1699         ptr = (unsigned long)ei + item_offset;
1700         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1701         if (ptr < end - size)
1702                 memmove_extent_buffer(leaf, ptr + size, ptr,
1703                                       end - size - ptr);
1704
1705         iref = (struct btrfs_extent_inline_ref *)ptr;
1706         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1707         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1708                 struct btrfs_extent_data_ref *dref;
1709                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1710                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1711                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1712                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1713                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1714         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1715                 struct btrfs_shared_data_ref *sref;
1716                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1717                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1718                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1719         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1720                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1721         } else {
1722                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1723         }
1724         btrfs_mark_buffer_dirty(leaf);
1725 }
1726
1727 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1728                                  struct btrfs_root *root,
1729                                  struct btrfs_path *path,
1730                                  struct btrfs_extent_inline_ref **ref_ret,
1731                                  u64 bytenr, u64 num_bytes, u64 parent,
1732                                  u64 root_objectid, u64 owner, u64 offset)
1733 {
1734         int ret;
1735
1736         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1737                                            bytenr, num_bytes, parent,
1738                                            root_objectid, owner, offset, 0);
1739         if (ret != -ENOENT)
1740                 return ret;
1741
1742         btrfs_release_path(path);
1743         *ref_ret = NULL;
1744
1745         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1746                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1747                                             root_objectid);
1748         } else {
1749                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1750                                              root_objectid, owner, offset);
1751         }
1752         return ret;
1753 }
1754
1755 /*
1756  * helper to update/remove inline back ref
1757  */
1758 static noinline_for_stack
1759 void update_inline_extent_backref(struct btrfs_root *root,
1760                                   struct btrfs_path *path,
1761                                   struct btrfs_extent_inline_ref *iref,
1762                                   int refs_to_mod,
1763                                   struct btrfs_delayed_extent_op *extent_op)
1764 {
1765         struct extent_buffer *leaf;
1766         struct btrfs_extent_item *ei;
1767         struct btrfs_extent_data_ref *dref = NULL;
1768         struct btrfs_shared_data_ref *sref = NULL;
1769         unsigned long ptr;
1770         unsigned long end;
1771         u32 item_size;
1772         int size;
1773         int type;
1774         u64 refs;
1775
1776         leaf = path->nodes[0];
1777         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1778         refs = btrfs_extent_refs(leaf, ei);
1779         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1780         refs += refs_to_mod;
1781         btrfs_set_extent_refs(leaf, ei, refs);
1782         if (extent_op)
1783                 __run_delayed_extent_op(extent_op, leaf, ei);
1784
1785         type = btrfs_extent_inline_ref_type(leaf, iref);
1786
1787         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1788                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1789                 refs = btrfs_extent_data_ref_count(leaf, dref);
1790         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1791                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1792                 refs = btrfs_shared_data_ref_count(leaf, sref);
1793         } else {
1794                 refs = 1;
1795                 BUG_ON(refs_to_mod != -1);
1796         }
1797
1798         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1799         refs += refs_to_mod;
1800
1801         if (refs > 0) {
1802                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1803                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1804                 else
1805                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1806         } else {
1807                 size =  btrfs_extent_inline_ref_size(type);
1808                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1809                 ptr = (unsigned long)iref;
1810                 end = (unsigned long)ei + item_size;
1811                 if (ptr + size < end)
1812                         memmove_extent_buffer(leaf, ptr, ptr + size,
1813                                               end - ptr - size);
1814                 item_size -= size;
1815                 btrfs_truncate_item(root, path, item_size, 1);
1816         }
1817         btrfs_mark_buffer_dirty(leaf);
1818 }
1819
1820 static noinline_for_stack
1821 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1822                                  struct btrfs_root *root,
1823                                  struct btrfs_path *path,
1824                                  u64 bytenr, u64 num_bytes, u64 parent,
1825                                  u64 root_objectid, u64 owner,
1826                                  u64 offset, int refs_to_add,
1827                                  struct btrfs_delayed_extent_op *extent_op)
1828 {
1829         struct btrfs_extent_inline_ref *iref;
1830         int ret;
1831
1832         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1833                                            bytenr, num_bytes, parent,
1834                                            root_objectid, owner, offset, 1);
1835         if (ret == 0) {
1836                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1837                 update_inline_extent_backref(root, path, iref,
1838                                              refs_to_add, extent_op);
1839         } else if (ret == -ENOENT) {
1840                 setup_inline_extent_backref(root, path, iref, parent,
1841                                             root_objectid, owner, offset,
1842                                             refs_to_add, extent_op);
1843                 ret = 0;
1844         }
1845         return ret;
1846 }
1847
1848 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1849                                  struct btrfs_root *root,
1850                                  struct btrfs_path *path,
1851                                  u64 bytenr, u64 parent, u64 root_objectid,
1852                                  u64 owner, u64 offset, int refs_to_add)
1853 {
1854         int ret;
1855         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1856                 BUG_ON(refs_to_add != 1);
1857                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1858                                             parent, root_objectid);
1859         } else {
1860                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1861                                              parent, root_objectid,
1862                                              owner, offset, refs_to_add);
1863         }
1864         return ret;
1865 }
1866
1867 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1868                                  struct btrfs_root *root,
1869                                  struct btrfs_path *path,
1870                                  struct btrfs_extent_inline_ref *iref,
1871                                  int refs_to_drop, int is_data)
1872 {
1873         int ret = 0;
1874
1875         BUG_ON(!is_data && refs_to_drop != 1);
1876         if (iref) {
1877                 update_inline_extent_backref(root, path, iref,
1878                                              -refs_to_drop, NULL);
1879         } else if (is_data) {
1880                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1881         } else {
1882                 ret = btrfs_del_item(trans, root, path);
1883         }
1884         return ret;
1885 }
1886
1887 static int btrfs_issue_discard(struct block_device *bdev,
1888                                 u64 start, u64 len)
1889 {
1890         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1891 }
1892
1893 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1894                                 u64 num_bytes, u64 *actual_bytes)
1895 {
1896         int ret;
1897         u64 discarded_bytes = 0;
1898         struct btrfs_bio *bbio = NULL;
1899
1900
1901         /* Tell the block device(s) that the sectors can be discarded */
1902         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1903                               bytenr, &num_bytes, &bbio, 0);
1904         /* Error condition is -ENOMEM */
1905         if (!ret) {
1906                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1907                 int i;
1908
1909
1910                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1911                         if (!stripe->dev->can_discard)
1912                                 continue;
1913
1914                         ret = btrfs_issue_discard(stripe->dev->bdev,
1915                                                   stripe->physical,
1916                                                   stripe->length);
1917                         if (!ret)
1918                                 discarded_bytes += stripe->length;
1919                         else if (ret != -EOPNOTSUPP)
1920                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1921
1922                         /*
1923                          * Just in case we get back EOPNOTSUPP for some reason,
1924                          * just ignore the return value so we don't screw up
1925                          * people calling discard_extent.
1926                          */
1927                         ret = 0;
1928                 }
1929                 kfree(bbio);
1930         }
1931
1932         if (actual_bytes)
1933                 *actual_bytes = discarded_bytes;
1934
1935
1936         if (ret == -EOPNOTSUPP)
1937                 ret = 0;
1938         return ret;
1939 }
1940
1941 /* Can return -ENOMEM */
1942 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1943                          struct btrfs_root *root,
1944                          u64 bytenr, u64 num_bytes, u64 parent,
1945                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
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, for_cow);
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, for_cow);
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                                   struct btrfs_delayed_extent_op *extent_op)
1973 {
1974         struct btrfs_path *path;
1975         struct extent_buffer *leaf;
1976         struct btrfs_extent_item *item;
1977         u64 refs;
1978         int ret;
1979
1980         path = btrfs_alloc_path();
1981         if (!path)
1982                 return -ENOMEM;
1983
1984         path->reada = 1;
1985         path->leave_spinning = 1;
1986         /* this will setup the path even if it fails to insert the back ref */
1987         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1988                                            path, bytenr, num_bytes, parent,
1989                                            root_objectid, owner, offset,
1990                                            refs_to_add, extent_op);
1991         if (ret != -EAGAIN)
1992                 goto out;
1993
1994         leaf = path->nodes[0];
1995         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1996         refs = btrfs_extent_refs(leaf, item);
1997         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1998         if (extent_op)
1999                 __run_delayed_extent_op(extent_op, leaf, item);
2000
2001         btrfs_mark_buffer_dirty(leaf);
2002         btrfs_release_path(path);
2003
2004         path->reada = 1;
2005         path->leave_spinning = 1;
2006
2007         /* now insert the actual backref */
2008         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2009                                     path, bytenr, parent, root_objectid,
2010                                     owner, offset, refs_to_add);
2011         if (ret)
2012                 btrfs_abort_transaction(trans, root, ret);
2013 out:
2014         btrfs_free_path(path);
2015         return ret;
2016 }
2017
2018 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2019                                 struct btrfs_root *root,
2020                                 struct btrfs_delayed_ref_node *node,
2021                                 struct btrfs_delayed_extent_op *extent_op,
2022                                 int insert_reserved)
2023 {
2024         int ret = 0;
2025         struct btrfs_delayed_data_ref *ref;
2026         struct btrfs_key ins;
2027         u64 parent = 0;
2028         u64 ref_root = 0;
2029         u64 flags = 0;
2030
2031         ins.objectid = node->bytenr;
2032         ins.offset = node->num_bytes;
2033         ins.type = BTRFS_EXTENT_ITEM_KEY;
2034
2035         ref = btrfs_delayed_node_to_data_ref(node);
2036         trace_run_delayed_data_ref(node, ref, node->action);
2037
2038         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2039                 parent = ref->parent;
2040         else
2041                 ref_root = ref->root;
2042
2043         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2044                 if (extent_op)
2045                         flags |= extent_op->flags_to_set;
2046                 ret = alloc_reserved_file_extent(trans, root,
2047                                                  parent, ref_root, flags,
2048                                                  ref->objectid, ref->offset,
2049                                                  &ins, node->ref_mod);
2050         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2051                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2052                                              node->num_bytes, parent,
2053                                              ref_root, ref->objectid,
2054                                              ref->offset, node->ref_mod,
2055                                              extent_op);
2056         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2057                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2058                                           node->num_bytes, parent,
2059                                           ref_root, ref->objectid,
2060                                           ref->offset, node->ref_mod,
2061                                           extent_op);
2062         } else {
2063                 BUG();
2064         }
2065         return ret;
2066 }
2067
2068 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2069                                     struct extent_buffer *leaf,
2070                                     struct btrfs_extent_item *ei)
2071 {
2072         u64 flags = btrfs_extent_flags(leaf, ei);
2073         if (extent_op->update_flags) {
2074                 flags |= extent_op->flags_to_set;
2075                 btrfs_set_extent_flags(leaf, ei, flags);
2076         }
2077
2078         if (extent_op->update_key) {
2079                 struct btrfs_tree_block_info *bi;
2080                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2081                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2082                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2083         }
2084 }
2085
2086 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2087                                  struct btrfs_root *root,
2088                                  struct btrfs_delayed_ref_node *node,
2089                                  struct btrfs_delayed_extent_op *extent_op)
2090 {
2091         struct btrfs_key key;
2092         struct btrfs_path *path;
2093         struct btrfs_extent_item *ei;
2094         struct extent_buffer *leaf;
2095         u32 item_size;
2096         int ret;
2097         int err = 0;
2098         int metadata = !extent_op->is_data;
2099
2100         if (trans->aborted)
2101                 return 0;
2102
2103         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2104                 metadata = 0;
2105
2106         path = btrfs_alloc_path();
2107         if (!path)
2108                 return -ENOMEM;
2109
2110         key.objectid = node->bytenr;
2111
2112         if (metadata) {
2113                 key.type = BTRFS_METADATA_ITEM_KEY;
2114                 key.offset = extent_op->level;
2115         } else {
2116                 key.type = BTRFS_EXTENT_ITEM_KEY;
2117                 key.offset = node->num_bytes;
2118         }
2119
2120 again:
2121         path->reada = 1;
2122         path->leave_spinning = 1;
2123         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2124                                 path, 0, 1);
2125         if (ret < 0) {
2126                 err = ret;
2127                 goto out;
2128         }
2129         if (ret > 0) {
2130                 if (metadata) {
2131                         if (path->slots[0] > 0) {
2132                                 path->slots[0]--;
2133                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2134                                                       path->slots[0]);
2135                                 if (key.objectid == node->bytenr &&
2136                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2137                                     key.offset == node->num_bytes)
2138                                         ret = 0;
2139                         }
2140                         if (ret > 0) {
2141                                 btrfs_release_path(path);
2142                                 metadata = 0;
2143
2144                                 key.objectid = node->bytenr;
2145                                 key.offset = node->num_bytes;
2146                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2147                                 goto again;
2148                         }
2149                 } else {
2150                         err = -EIO;
2151                         goto out;
2152                 }
2153         }
2154
2155         leaf = path->nodes[0];
2156         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2157 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2158         if (item_size < sizeof(*ei)) {
2159                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2160                                              path, (u64)-1, 0);
2161                 if (ret < 0) {
2162                         err = ret;
2163                         goto out;
2164                 }
2165                 leaf = path->nodes[0];
2166                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2167         }
2168 #endif
2169         BUG_ON(item_size < sizeof(*ei));
2170         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2171         __run_delayed_extent_op(extent_op, leaf, ei);
2172
2173         btrfs_mark_buffer_dirty(leaf);
2174 out:
2175         btrfs_free_path(path);
2176         return err;
2177 }
2178
2179 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2180                                 struct btrfs_root *root,
2181                                 struct btrfs_delayed_ref_node *node,
2182                                 struct btrfs_delayed_extent_op *extent_op,
2183                                 int insert_reserved)
2184 {
2185         int ret = 0;
2186         struct btrfs_delayed_tree_ref *ref;
2187         struct btrfs_key ins;
2188         u64 parent = 0;
2189         u64 ref_root = 0;
2190         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2191                                                  SKINNY_METADATA);
2192
2193         ref = btrfs_delayed_node_to_tree_ref(node);
2194         trace_run_delayed_tree_ref(node, ref, node->action);
2195
2196         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2197                 parent = ref->parent;
2198         else
2199                 ref_root = ref->root;
2200
2201         ins.objectid = node->bytenr;
2202         if (skinny_metadata) {
2203                 ins.offset = ref->level;
2204                 ins.type = BTRFS_METADATA_ITEM_KEY;
2205         } else {
2206                 ins.offset = node->num_bytes;
2207                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2208         }
2209
2210         BUG_ON(node->ref_mod != 1);
2211         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2212                 BUG_ON(!extent_op || !extent_op->update_flags);
2213                 ret = alloc_reserved_tree_block(trans, root,
2214                                                 parent, ref_root,
2215                                                 extent_op->flags_to_set,
2216                                                 &extent_op->key,
2217                                                 ref->level, &ins);
2218         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2219                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2220                                              node->num_bytes, parent, ref_root,
2221                                              ref->level, 0, 1, extent_op);
2222         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2223                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2224                                           node->num_bytes, parent, ref_root,
2225                                           ref->level, 0, 1, extent_op);
2226         } else {
2227                 BUG();
2228         }
2229         return ret;
2230 }
2231
2232 /* helper function to actually process a single delayed ref entry */
2233 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2234                                struct btrfs_root *root,
2235                                struct btrfs_delayed_ref_node *node,
2236                                struct btrfs_delayed_extent_op *extent_op,
2237                                int insert_reserved)
2238 {
2239         int ret = 0;
2240
2241         if (trans->aborted) {
2242                 if (insert_reserved)
2243                         btrfs_pin_extent(root, node->bytenr,
2244                                          node->num_bytes, 1);
2245                 return 0;
2246         }
2247
2248         if (btrfs_delayed_ref_is_head(node)) {
2249                 struct btrfs_delayed_ref_head *head;
2250                 /*
2251                  * we've hit the end of the chain and we were supposed
2252                  * to insert this extent into the tree.  But, it got
2253                  * deleted before we ever needed to insert it, so all
2254                  * we have to do is clean up the accounting
2255                  */
2256                 BUG_ON(extent_op);
2257                 head = btrfs_delayed_node_to_head(node);
2258                 trace_run_delayed_ref_head(node, head, node->action);
2259
2260                 if (insert_reserved) {
2261                         btrfs_pin_extent(root, node->bytenr,
2262                                          node->num_bytes, 1);
2263                         if (head->is_data) {
2264                                 ret = btrfs_del_csums(trans, root,
2265                                                       node->bytenr,
2266                                                       node->num_bytes);
2267                         }
2268                 }
2269                 return ret;
2270         }
2271
2272         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2273             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2274                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2275                                            insert_reserved);
2276         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2277                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2278                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2279                                            insert_reserved);
2280         else
2281                 BUG();
2282         return ret;
2283 }
2284
2285 static noinline struct btrfs_delayed_ref_node *
2286 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2287 {
2288         struct rb_node *node;
2289         struct btrfs_delayed_ref_node *ref;
2290         int action = BTRFS_ADD_DELAYED_REF;
2291 again:
2292         /*
2293          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2294          * this prevents ref count from going down to zero when
2295          * there still are pending delayed ref.
2296          */
2297         node = rb_prev(&head->node.rb_node);
2298         while (1) {
2299                 if (!node)
2300                         break;
2301                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2302                                 rb_node);
2303                 if (ref->bytenr != head->node.bytenr)
2304                         break;
2305                 if (ref->action == action)
2306                         return ref;
2307                 node = rb_prev(node);
2308         }
2309         if (action == BTRFS_ADD_DELAYED_REF) {
2310                 action = BTRFS_DROP_DELAYED_REF;
2311                 goto again;
2312         }
2313         return NULL;
2314 }
2315
2316 /*
2317  * Returns 0 on success or if called with an already aborted transaction.
2318  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2319  */
2320 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2321                                        struct btrfs_root *root,
2322                                        struct list_head *cluster)
2323 {
2324         struct btrfs_delayed_ref_root *delayed_refs;
2325         struct btrfs_delayed_ref_node *ref;
2326         struct btrfs_delayed_ref_head *locked_ref = NULL;
2327         struct btrfs_delayed_extent_op *extent_op;
2328         struct btrfs_fs_info *fs_info = root->fs_info;
2329         int ret;
2330         int count = 0;
2331         int must_insert_reserved = 0;
2332
2333         delayed_refs = &trans->transaction->delayed_refs;
2334         while (1) {
2335                 if (!locked_ref) {
2336                         /* pick a new head ref from the cluster list */
2337                         if (list_empty(cluster))
2338                                 break;
2339
2340                         locked_ref = list_entry(cluster->next,
2341                                      struct btrfs_delayed_ref_head, cluster);
2342
2343                         /* grab the lock that says we are going to process
2344                          * all the refs for this head */
2345                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2346
2347                         /*
2348                          * we may have dropped the spin lock to get the head
2349                          * mutex lock, and that might have given someone else
2350                          * time to free the head.  If that's true, it has been
2351                          * removed from our list and we can move on.
2352                          */
2353                         if (ret == -EAGAIN) {
2354                                 locked_ref = NULL;
2355                                 count++;
2356                                 continue;
2357                         }
2358                 }
2359
2360                 /*
2361                  * We need to try and merge add/drops of the same ref since we
2362                  * can run into issues with relocate dropping the implicit ref
2363                  * and then it being added back again before the drop can
2364                  * finish.  If we merged anything we need to re-loop so we can
2365                  * get a good ref.
2366                  */
2367                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2368                                          locked_ref);
2369
2370                 /*
2371                  * locked_ref is the head node, so we have to go one
2372                  * node back for any delayed ref updates
2373                  */
2374                 ref = select_delayed_ref(locked_ref);
2375
2376                 if (ref && ref->seq &&
2377                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2378                         /*
2379                          * there are still refs with lower seq numbers in the
2380                          * process of being added. Don't run this ref yet.
2381                          */
2382                         list_del_init(&locked_ref->cluster);
2383                         btrfs_delayed_ref_unlock(locked_ref);
2384                         locked_ref = NULL;
2385                         delayed_refs->num_heads_ready++;
2386                         spin_unlock(&delayed_refs->lock);
2387                         cond_resched();
2388                         spin_lock(&delayed_refs->lock);
2389                         continue;
2390                 }
2391
2392                 /*
2393                  * record the must insert reserved flag before we
2394                  * drop the spin lock.
2395                  */
2396                 must_insert_reserved = locked_ref->must_insert_reserved;
2397                 locked_ref->must_insert_reserved = 0;
2398
2399                 extent_op = locked_ref->extent_op;
2400                 locked_ref->extent_op = NULL;
2401
2402                 if (!ref) {
2403                         /* All delayed refs have been processed, Go ahead
2404                          * and send the head node to run_one_delayed_ref,
2405                          * so that any accounting fixes can happen
2406                          */
2407                         ref = &locked_ref->node;
2408
2409                         if (extent_op && must_insert_reserved) {
2410                                 btrfs_free_delayed_extent_op(extent_op);
2411                                 extent_op = NULL;
2412                         }
2413
2414                         if (extent_op) {
2415                                 spin_unlock(&delayed_refs->lock);
2416
2417                                 ret = run_delayed_extent_op(trans, root,
2418                                                             ref, extent_op);
2419                                 btrfs_free_delayed_extent_op(extent_op);
2420
2421                                 if (ret) {
2422                                         /*
2423                                          * Need to reset must_insert_reserved if
2424                                          * there was an error so the abort stuff
2425                                          * can cleanup the reserved space
2426                                          * properly.
2427                                          */
2428                                         if (must_insert_reserved)
2429                                                 locked_ref->must_insert_reserved = 1;
2430                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2431                                         spin_lock(&delayed_refs->lock);
2432                                         btrfs_delayed_ref_unlock(locked_ref);
2433                                         return ret;
2434                                 }
2435
2436                                 goto next;
2437                         }
2438                 }
2439
2440                 ref->in_tree = 0;
2441                 rb_erase(&ref->rb_node, &delayed_refs->root);
2442                 if (btrfs_delayed_ref_is_head(ref)) {
2443                         rb_erase(&locked_ref->href_node,
2444                                  &delayed_refs->href_root);
2445                 }
2446                 delayed_refs->num_entries--;
2447                 if (!btrfs_delayed_ref_is_head(ref)) {
2448                         /*
2449                          * when we play the delayed ref, also correct the
2450                          * ref_mod on head
2451                          */
2452                         switch (ref->action) {
2453                         case BTRFS_ADD_DELAYED_REF:
2454                         case BTRFS_ADD_DELAYED_EXTENT:
2455                                 locked_ref->node.ref_mod -= ref->ref_mod;
2456                                 break;
2457                         case BTRFS_DROP_DELAYED_REF:
2458                                 locked_ref->node.ref_mod += ref->ref_mod;
2459                                 break;
2460                         default:
2461                                 WARN_ON(1);
2462                         }
2463                 } else {
2464                         list_del_init(&locked_ref->cluster);
2465                 }
2466                 spin_unlock(&delayed_refs->lock);
2467
2468                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2469                                           must_insert_reserved);
2470
2471                 btrfs_free_delayed_extent_op(extent_op);
2472                 if (ret) {
2473                         btrfs_delayed_ref_unlock(locked_ref);
2474                         btrfs_put_delayed_ref(ref);
2475                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2476                         spin_lock(&delayed_refs->lock);
2477                         return ret;
2478                 }
2479
2480                 /*
2481                  * If this node is a head, that means all the refs in this head
2482                  * have been dealt with, and we will pick the next head to deal
2483                  * with, so we must unlock the head and drop it from the cluster
2484                  * list before we release it.
2485                  */
2486                 if (btrfs_delayed_ref_is_head(ref)) {
2487                         btrfs_delayed_ref_unlock(locked_ref);
2488                         locked_ref = NULL;
2489                 }
2490                 btrfs_put_delayed_ref(ref);
2491                 count++;
2492 next:
2493                 cond_resched();
2494                 spin_lock(&delayed_refs->lock);
2495         }
2496         return count;
2497 }
2498
2499 #ifdef SCRAMBLE_DELAYED_REFS
2500 /*
2501  * Normally delayed refs get processed in ascending bytenr order. This
2502  * correlates in most cases to the order added. To expose dependencies on this
2503  * order, we start to process the tree in the middle instead of the beginning
2504  */
2505 static u64 find_middle(struct rb_root *root)
2506 {
2507         struct rb_node *n = root->rb_node;
2508         struct btrfs_delayed_ref_node *entry;
2509         int alt = 1;
2510         u64 middle;
2511         u64 first = 0, last = 0;
2512
2513         n = rb_first(root);
2514         if (n) {
2515                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2516                 first = entry->bytenr;
2517         }
2518         n = rb_last(root);
2519         if (n) {
2520                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2521                 last = entry->bytenr;
2522         }
2523         n = root->rb_node;
2524
2525         while (n) {
2526                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2527                 WARN_ON(!entry->in_tree);
2528
2529                 middle = entry->bytenr;
2530
2531                 if (alt)
2532                         n = n->rb_left;
2533                 else
2534                         n = n->rb_right;
2535
2536                 alt = 1 - alt;
2537         }
2538         return middle;
2539 }
2540 #endif
2541
2542 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2543                                          struct btrfs_fs_info *fs_info)
2544 {
2545         struct qgroup_update *qgroup_update;
2546         int ret = 0;
2547
2548         if (list_empty(&trans->qgroup_ref_list) !=
2549             !trans->delayed_ref_elem.seq) {
2550                 /* list without seq or seq without list */
2551                 btrfs_err(fs_info,
2552                         "qgroup accounting update error, list is%s empty, seq is %#x.%x",
2553                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2554                         (u32)(trans->delayed_ref_elem.seq >> 32),
2555                         (u32)trans->delayed_ref_elem.seq);
2556                 BUG();
2557         }
2558
2559         if (!trans->delayed_ref_elem.seq)
2560                 return 0;
2561
2562         while (!list_empty(&trans->qgroup_ref_list)) {
2563                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2564                                                  struct qgroup_update, list);
2565                 list_del(&qgroup_update->list);
2566                 if (!ret)
2567                         ret = btrfs_qgroup_account_ref(
2568                                         trans, fs_info, qgroup_update->node,
2569                                         qgroup_update->extent_op);
2570                 kfree(qgroup_update);
2571         }
2572
2573         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2574
2575         return ret;
2576 }
2577
2578 static int refs_newer(struct btrfs_delayed_ref_root *delayed_refs, int seq,
2579                       int count)
2580 {
2581         int val = atomic_read(&delayed_refs->ref_seq);
2582
2583         if (val < seq || val >= seq + count)
2584                 return 1;
2585         return 0;
2586 }
2587
2588 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2589 {
2590         u64 num_bytes;
2591
2592         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2593                              sizeof(struct btrfs_extent_inline_ref));
2594         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2595                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2596
2597         /*
2598          * We don't ever fill up leaves all the way so multiply by 2 just to be
2599          * closer to what we're really going to want to ouse.
2600          */
2601         return div64_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2602 }
2603
2604 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2605                                        struct btrfs_root *root)
2606 {
2607         struct btrfs_block_rsv *global_rsv;
2608         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2609         u64 num_bytes;
2610         int ret = 0;
2611
2612         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2613         num_heads = heads_to_leaves(root, num_heads);
2614         if (num_heads > 1)
2615                 num_bytes += (num_heads - 1) * root->leafsize;
2616         num_bytes <<= 1;
2617         global_rsv = &root->fs_info->global_block_rsv;
2618
2619         /*
2620          * If we can't allocate any more chunks lets make sure we have _lots_ of
2621          * wiggle room since running delayed refs can create more delayed refs.
2622          */
2623         if (global_rsv->space_info->full)
2624                 num_bytes <<= 1;
2625
2626         spin_lock(&global_rsv->lock);
2627         if (global_rsv->reserved <= num_bytes)
2628                 ret = 1;
2629         spin_unlock(&global_rsv->lock);
2630         return ret;
2631 }
2632
2633 /*
2634  * this starts processing the delayed reference count updates and
2635  * extent insertions we have queued up so far.  count can be
2636  * 0, which means to process everything in the tree at the start
2637  * of the run (but not newly added entries), or it can be some target
2638  * number you'd like to process.
2639  *
2640  * Returns 0 on success or if called with an aborted transaction
2641  * Returns <0 on error and aborts the transaction
2642  */
2643 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2644                            struct btrfs_root *root, unsigned long count)
2645 {
2646         struct rb_node *node;
2647         struct btrfs_delayed_ref_root *delayed_refs;
2648         struct btrfs_delayed_ref_head *head;
2649         struct list_head cluster;
2650         int ret;
2651         u64 delayed_start;
2652         int run_all = count == (unsigned long)-1;
2653         int run_most = 0;
2654         int loops;
2655
2656         /* We'll clean this up in btrfs_cleanup_transaction */
2657         if (trans->aborted)
2658                 return 0;
2659
2660         if (root == root->fs_info->extent_root)
2661                 root = root->fs_info->tree_root;
2662
2663         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2664
2665         delayed_refs = &trans->transaction->delayed_refs;
2666         INIT_LIST_HEAD(&cluster);
2667         if (count == 0) {
2668                 count = delayed_refs->num_entries * 2;
2669                 run_most = 1;
2670         }
2671
2672         if (!run_all && !run_most) {
2673                 int old;
2674                 int seq = atomic_read(&delayed_refs->ref_seq);
2675
2676 progress:
2677                 old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2678                 if (old) {
2679                         DEFINE_WAIT(__wait);
2680                         if (delayed_refs->flushing ||
2681                             !btrfs_should_throttle_delayed_refs(trans, root))
2682                                 return 0;
2683
2684                         prepare_to_wait(&delayed_refs->wait, &__wait,
2685                                         TASK_UNINTERRUPTIBLE);
2686
2687                         old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2688                         if (old) {
2689                                 schedule();
2690                                 finish_wait(&delayed_refs->wait, &__wait);
2691
2692                                 if (!refs_newer(delayed_refs, seq, 256))
2693                                         goto progress;
2694                                 else
2695                                         return 0;
2696                         } else {
2697                                 finish_wait(&delayed_refs->wait, &__wait);
2698                                 goto again;
2699                         }
2700                 }
2701
2702         } else {
2703                 atomic_inc(&delayed_refs->procs_running_refs);
2704         }
2705
2706 again:
2707         loops = 0;
2708         spin_lock(&delayed_refs->lock);
2709
2710 #ifdef SCRAMBLE_DELAYED_REFS
2711         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2712 #endif
2713
2714         while (1) {
2715                 if (!(run_all || run_most) &&
2716                     !btrfs_should_throttle_delayed_refs(trans, root))
2717                         break;
2718
2719                 /*
2720                  * go find something we can process in the rbtree.  We start at
2721                  * the beginning of the tree, and then build a cluster
2722                  * of refs to process starting at the first one we are able to
2723                  * lock
2724                  */
2725                 delayed_start = delayed_refs->run_delayed_start;
2726                 ret = btrfs_find_ref_cluster(trans, &cluster,
2727                                              delayed_refs->run_delayed_start);
2728                 if (ret)
2729                         break;
2730
2731                 ret = run_clustered_refs(trans, root, &cluster);
2732                 if (ret < 0) {
2733                         btrfs_release_ref_cluster(&cluster);
2734                         spin_unlock(&delayed_refs->lock);
2735                         btrfs_abort_transaction(trans, root, ret);
2736                         atomic_dec(&delayed_refs->procs_running_refs);
2737                         wake_up(&delayed_refs->wait);
2738                         return ret;
2739                 }
2740
2741                 atomic_add(ret, &delayed_refs->ref_seq);
2742
2743                 count -= min_t(unsigned long, ret, count);
2744
2745                 if (count == 0)
2746                         break;
2747
2748                 if (delayed_start >= delayed_refs->run_delayed_start) {
2749                         if (loops == 0) {
2750                                 /*
2751                                  * btrfs_find_ref_cluster looped. let's do one
2752                                  * more cycle. if we don't run any delayed ref
2753                                  * during that cycle (because we can't because
2754                                  * all of them are blocked), bail out.
2755                                  */
2756                                 loops = 1;
2757                         } else {
2758                                 /*
2759                                  * no runnable refs left, stop trying
2760                                  */
2761                                 BUG_ON(run_all);
2762                                 break;
2763                         }
2764                 }
2765                 if (ret) {
2766                         /* refs were run, let's reset staleness detection */
2767                         loops = 0;
2768                 }
2769         }
2770
2771         if (run_all) {
2772                 if (!list_empty(&trans->new_bgs)) {
2773                         spin_unlock(&delayed_refs->lock);
2774                         btrfs_create_pending_block_groups(trans, root);
2775                         spin_lock(&delayed_refs->lock);
2776                 }
2777
2778                 node = rb_first(&delayed_refs->href_root);
2779                 if (!node)
2780                         goto out;
2781                 count = (unsigned long)-1;
2782
2783                 while (node) {
2784                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2785                                         href_node);
2786                         if (btrfs_delayed_ref_is_head(&head->node)) {
2787                                 struct btrfs_delayed_ref_node *ref;
2788
2789                                 ref = &head->node;
2790                                 atomic_inc(&ref->refs);
2791
2792                                 spin_unlock(&delayed_refs->lock);
2793                                 /*
2794                                  * Mutex was contended, block until it's
2795                                  * released and try again
2796                                  */
2797                                 mutex_lock(&head->mutex);
2798                                 mutex_unlock(&head->mutex);
2799
2800                                 btrfs_put_delayed_ref(ref);
2801                                 cond_resched();
2802                                 goto again;
2803                         } else {
2804                                 WARN_ON(1);
2805                         }
2806                         node = rb_next(node);
2807                 }
2808                 spin_unlock(&delayed_refs->lock);
2809                 schedule_timeout(1);
2810                 goto again;
2811         }
2812 out:
2813         atomic_dec(&delayed_refs->procs_running_refs);
2814         smp_mb();
2815         if (waitqueue_active(&delayed_refs->wait))
2816                 wake_up(&delayed_refs->wait);
2817
2818         spin_unlock(&delayed_refs->lock);
2819         assert_qgroups_uptodate(trans);
2820         return 0;
2821 }
2822
2823 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2824                                 struct btrfs_root *root,
2825                                 u64 bytenr, u64 num_bytes, u64 flags,
2826                                 int level, int is_data)
2827 {
2828         struct btrfs_delayed_extent_op *extent_op;
2829         int ret;
2830
2831         extent_op = btrfs_alloc_delayed_extent_op();
2832         if (!extent_op)
2833                 return -ENOMEM;
2834
2835         extent_op->flags_to_set = flags;
2836         extent_op->update_flags = 1;
2837         extent_op->update_key = 0;
2838         extent_op->is_data = is_data ? 1 : 0;
2839         extent_op->level = level;
2840
2841         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2842                                           num_bytes, extent_op);
2843         if (ret)
2844                 btrfs_free_delayed_extent_op(extent_op);
2845         return ret;
2846 }
2847
2848 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2849                                       struct btrfs_root *root,
2850                                       struct btrfs_path *path,
2851                                       u64 objectid, u64 offset, u64 bytenr)
2852 {
2853         struct btrfs_delayed_ref_head *head;
2854         struct btrfs_delayed_ref_node *ref;
2855         struct btrfs_delayed_data_ref *data_ref;
2856         struct btrfs_delayed_ref_root *delayed_refs;
2857         struct rb_node *node;
2858         int ret = 0;
2859
2860         ret = -ENOENT;
2861         delayed_refs = &trans->transaction->delayed_refs;
2862         spin_lock(&delayed_refs->lock);
2863         head = btrfs_find_delayed_ref_head(trans, bytenr);
2864         if (!head)
2865                 goto out;
2866
2867         if (!mutex_trylock(&head->mutex)) {
2868                 atomic_inc(&head->node.refs);
2869                 spin_unlock(&delayed_refs->lock);
2870
2871                 btrfs_release_path(path);
2872
2873                 /*
2874                  * Mutex was contended, block until it's released and let
2875                  * caller try again
2876                  */
2877                 mutex_lock(&head->mutex);
2878                 mutex_unlock(&head->mutex);
2879                 btrfs_put_delayed_ref(&head->node);
2880                 return -EAGAIN;
2881         }
2882
2883         node = rb_prev(&head->node.rb_node);
2884         if (!node)
2885                 goto out_unlock;
2886
2887         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2888
2889         if (ref->bytenr != bytenr)
2890                 goto out_unlock;
2891
2892         ret = 1;
2893         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2894                 goto out_unlock;
2895
2896         data_ref = btrfs_delayed_node_to_data_ref(ref);
2897
2898         node = rb_prev(node);
2899         if (node) {
2900                 int seq = ref->seq;
2901
2902                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2903                 if (ref->bytenr == bytenr && ref->seq == seq)
2904                         goto out_unlock;
2905         }
2906
2907         if (data_ref->root != root->root_key.objectid ||
2908             data_ref->objectid != objectid || data_ref->offset != offset)
2909                 goto out_unlock;
2910
2911         ret = 0;
2912 out_unlock:
2913         mutex_unlock(&head->mutex);
2914 out:
2915         spin_unlock(&delayed_refs->lock);
2916         return ret;
2917 }
2918
2919 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2920                                         struct btrfs_root *root,
2921                                         struct btrfs_path *path,
2922                                         u64 objectid, u64 offset, u64 bytenr)
2923 {
2924         struct btrfs_root *extent_root = root->fs_info->extent_root;
2925         struct extent_buffer *leaf;
2926         struct btrfs_extent_data_ref *ref;
2927         struct btrfs_extent_inline_ref *iref;
2928         struct btrfs_extent_item *ei;
2929         struct btrfs_key key;
2930         u32 item_size;
2931         int ret;
2932
2933         key.objectid = bytenr;
2934         key.offset = (u64)-1;
2935         key.type = BTRFS_EXTENT_ITEM_KEY;
2936
2937         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2938         if (ret < 0)
2939                 goto out;
2940         BUG_ON(ret == 0); /* Corruption */
2941
2942         ret = -ENOENT;
2943         if (path->slots[0] == 0)
2944                 goto out;
2945
2946         path->slots[0]--;
2947         leaf = path->nodes[0];
2948         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2949
2950         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2951                 goto out;
2952
2953         ret = 1;
2954         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2955 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2956         if (item_size < sizeof(*ei)) {
2957                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2958                 goto out;
2959         }
2960 #endif
2961         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2962
2963         if (item_size != sizeof(*ei) +
2964             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2965                 goto out;
2966
2967         if (btrfs_extent_generation(leaf, ei) <=
2968             btrfs_root_last_snapshot(&root->root_item))
2969                 goto out;
2970
2971         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2972         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2973             BTRFS_EXTENT_DATA_REF_KEY)
2974                 goto out;
2975
2976         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2977         if (btrfs_extent_refs(leaf, ei) !=
2978             btrfs_extent_data_ref_count(leaf, ref) ||
2979             btrfs_extent_data_ref_root(leaf, ref) !=
2980             root->root_key.objectid ||
2981             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2982             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2983                 goto out;
2984
2985         ret = 0;
2986 out:
2987         return ret;
2988 }
2989
2990 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2991                           struct btrfs_root *root,
2992                           u64 objectid, u64 offset, u64 bytenr)
2993 {
2994         struct btrfs_path *path;
2995         int ret;
2996         int ret2;
2997
2998         path = btrfs_alloc_path();
2999         if (!path)
3000                 return -ENOENT;
3001
3002         do {
3003                 ret = check_committed_ref(trans, root, path, objectid,
3004                                           offset, bytenr);
3005                 if (ret && ret != -ENOENT)
3006                         goto out;
3007
3008                 ret2 = check_delayed_ref(trans, root, path, objectid,
3009                                          offset, bytenr);
3010         } while (ret2 == -EAGAIN);
3011
3012         if (ret2 && ret2 != -ENOENT) {
3013                 ret = ret2;
3014                 goto out;
3015         }
3016
3017         if (ret != -ENOENT || ret2 != -ENOENT)
3018                 ret = 0;
3019 out:
3020         btrfs_free_path(path);
3021         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3022                 WARN_ON(ret > 0);
3023         return ret;
3024 }
3025
3026 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3027                            struct btrfs_root *root,
3028                            struct extent_buffer *buf,
3029                            int full_backref, int inc, int for_cow)
3030 {
3031         u64 bytenr;
3032         u64 num_bytes;
3033         u64 parent;
3034         u64 ref_root;
3035         u32 nritems;
3036         struct btrfs_key key;
3037         struct btrfs_file_extent_item *fi;
3038         int i;
3039         int level;
3040         int ret = 0;
3041         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3042                             u64, u64, u64, u64, u64, u64, int);
3043
3044         ref_root = btrfs_header_owner(buf);
3045         nritems = btrfs_header_nritems(buf);
3046         level = btrfs_header_level(buf);
3047
3048         if (!root->ref_cows && level == 0)
3049                 return 0;
3050
3051         if (inc)
3052                 process_func = btrfs_inc_extent_ref;
3053         else
3054                 process_func = btrfs_free_extent;
3055
3056         if (full_backref)
3057                 parent = buf->start;
3058         else
3059                 parent = 0;
3060
3061         for (i = 0; i < nritems; i++) {
3062                 if (level == 0) {
3063                         btrfs_item_key_to_cpu(buf, &key, i);
3064                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
3065                                 continue;
3066                         fi = btrfs_item_ptr(buf, i,
3067                                             struct btrfs_file_extent_item);
3068                         if (btrfs_file_extent_type(buf, fi) ==
3069                             BTRFS_FILE_EXTENT_INLINE)
3070                                 continue;
3071                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3072                         if (bytenr == 0)
3073                                 continue;
3074
3075                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3076                         key.offset -= btrfs_file_extent_offset(buf, fi);
3077                         ret = process_func(trans, root, bytenr, num_bytes,
3078                                            parent, ref_root, key.objectid,
3079                                            key.offset, for_cow);
3080                         if (ret)
3081                                 goto fail;
3082                 } else {
3083                         bytenr = btrfs_node_blockptr(buf, i);
3084                         num_bytes = btrfs_level_size(root, level - 1);
3085                         ret = process_func(trans, root, bytenr, num_bytes,
3086                                            parent, ref_root, level - 1, 0,
3087                                            for_cow);
3088                         if (ret)
3089                                 goto fail;
3090                 }
3091         }
3092         return 0;
3093 fail:
3094         return ret;
3095 }
3096
3097 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3098                   struct extent_buffer *buf, int full_backref, int for_cow)
3099 {
3100         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
3101 }
3102
3103 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3104                   struct extent_buffer *buf, int full_backref, int for_cow)
3105 {
3106         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
3107 }
3108
3109 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3110                                  struct btrfs_root *root,
3111                                  struct btrfs_path *path,
3112                                  struct btrfs_block_group_cache *cache)
3113 {
3114         int ret;
3115         struct btrfs_root *extent_root = root->fs_info->extent_root;
3116         unsigned long bi;
3117         struct extent_buffer *leaf;
3118
3119         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3120         if (ret < 0)
3121                 goto fail;
3122         BUG_ON(ret); /* Corruption */
3123
3124         leaf = path->nodes[0];
3125         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3126         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3127         btrfs_mark_buffer_dirty(leaf);
3128         btrfs_release_path(path);
3129 fail:
3130         if (ret) {
3131                 btrfs_abort_transaction(trans, root, ret);
3132                 return ret;
3133         }
3134         return 0;
3135
3136 }
3137
3138 static struct btrfs_block_group_cache *
3139 next_block_group(struct btrfs_root *root,
3140                  struct btrfs_block_group_cache *cache)
3141 {
3142         struct rb_node *node;
3143         spin_lock(&root->fs_info->block_group_cache_lock);
3144         node = rb_next(&cache->cache_node);
3145         btrfs_put_block_group(cache);
3146         if (node) {
3147                 cache = rb_entry(node, struct btrfs_block_group_cache,
3148                                  cache_node);
3149                 btrfs_get_block_group(cache);
3150         } else
3151                 cache = NULL;
3152         spin_unlock(&root->fs_info->block_group_cache_lock);
3153         return cache;
3154 }
3155
3156 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3157                             struct btrfs_trans_handle *trans,
3158                             struct btrfs_path *path)
3159 {
3160         struct btrfs_root *root = block_group->fs_info->tree_root;
3161         struct inode *inode = NULL;
3162         u64 alloc_hint = 0;
3163         int dcs = BTRFS_DC_ERROR;
3164         int num_pages = 0;
3165         int retries = 0;
3166         int ret = 0;
3167
3168         /*
3169          * If this block group is smaller than 100 megs don't bother caching the
3170          * block group.
3171          */
3172         if (block_group->key.offset < (100 * 1024 * 1024)) {
3173                 spin_lock(&block_group->lock);
3174                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3175                 spin_unlock(&block_group->lock);
3176                 return 0;
3177         }
3178
3179 again:
3180         inode = lookup_free_space_inode(root, block_group, path);
3181         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3182                 ret = PTR_ERR(inode);
3183                 btrfs_release_path(path);
3184                 goto out;
3185         }
3186
3187         if (IS_ERR(inode)) {
3188                 BUG_ON(retries);
3189                 retries++;
3190
3191                 if (block_group->ro)
3192                         goto out_free;
3193
3194                 ret = create_free_space_inode(root, trans, block_group, path);
3195                 if (ret)
3196                         goto out_free;
3197                 goto again;
3198         }
3199
3200         /* We've already setup this transaction, go ahead and exit */
3201         if (block_group->cache_generation == trans->transid &&
3202             i_size_read(inode)) {
3203                 dcs = BTRFS_DC_SETUP;
3204                 goto out_put;
3205         }
3206
3207         /*
3208          * We want to set the generation to 0, that way if anything goes wrong
3209          * from here on out we know not to trust this cache when we load up next
3210          * time.
3211          */
3212         BTRFS_I(inode)->generation = 0;
3213         ret = btrfs_update_inode(trans, root, inode);
3214         WARN_ON(ret);
3215
3216         if (i_size_read(inode) > 0) {
3217                 ret = btrfs_check_trunc_cache_free_space(root,
3218                                         &root->fs_info->global_block_rsv);
3219                 if (ret)
3220                         goto out_put;
3221
3222                 ret = btrfs_truncate_free_space_cache(root, trans, inode);
3223                 if (ret)
3224                         goto out_put;
3225         }
3226
3227         spin_lock(&block_group->lock);
3228         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3229             !btrfs_test_opt(root, SPACE_CACHE)) {
3230                 /*
3231                  * don't bother trying to write stuff out _if_
3232                  * a) we're not cached,
3233                  * b) we're with nospace_cache mount option.
3234                  */
3235                 dcs = BTRFS_DC_WRITTEN;
3236                 spin_unlock(&block_group->lock);
3237                 goto out_put;
3238         }
3239         spin_unlock(&block_group->lock);
3240
3241         /*
3242          * Try to preallocate enough space based on how big the block group is.
3243          * Keep in mind this has to include any pinned space which could end up
3244          * taking up quite a bit since it's not folded into the other space
3245          * cache.
3246          */
3247         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3248         if (!num_pages)
3249                 num_pages = 1;
3250
3251         num_pages *= 16;
3252         num_pages *= PAGE_CACHE_SIZE;
3253
3254         ret = btrfs_check_data_free_space(inode, num_pages);
3255         if (ret)
3256                 goto out_put;
3257
3258         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3259                                               num_pages, num_pages,
3260                                               &alloc_hint);
3261         if (!ret)
3262                 dcs = BTRFS_DC_SETUP;
3263         btrfs_free_reserved_data_space(inode, num_pages);
3264
3265 out_put:
3266         iput(inode);
3267 out_free:
3268         btrfs_release_path(path);
3269 out:
3270         spin_lock(&block_group->lock);
3271         if (!ret && dcs == BTRFS_DC_SETUP)
3272                 block_group->cache_generation = trans->transid;
3273         block_group->disk_cache_state = dcs;
3274         spin_unlock(&block_group->lock);
3275
3276         return ret;
3277 }
3278
3279 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3280                                    struct btrfs_root *root)
3281 {
3282         struct btrfs_block_group_cache *cache;
3283         int err = 0;
3284         struct btrfs_path *path;
3285         u64 last = 0;
3286
3287         path = btrfs_alloc_path();
3288         if (!path)
3289                 return -ENOMEM;
3290
3291 again:
3292         while (1) {
3293                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3294                 while (cache) {
3295                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3296                                 break;
3297                         cache = next_block_group(root, cache);
3298                 }
3299                 if (!cache) {
3300                         if (last == 0)
3301                                 break;
3302                         last = 0;
3303                         continue;
3304                 }
3305                 err = cache_save_setup(cache, trans, path);
3306                 last = cache->key.objectid + cache->key.offset;
3307                 btrfs_put_block_group(cache);
3308         }
3309
3310         while (1) {
3311                 if (last == 0) {
3312                         err = btrfs_run_delayed_refs(trans, root,
3313                                                      (unsigned long)-1);
3314                         if (err) /* File system offline */
3315                                 goto out;
3316                 }
3317
3318                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3319                 while (cache) {
3320                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3321                                 btrfs_put_block_group(cache);
3322                                 goto again;
3323                         }
3324
3325                         if (cache->dirty)
3326                                 break;
3327                         cache = next_block_group(root, cache);
3328                 }
3329                 if (!cache) {
3330                         if (last == 0)
3331                                 break;
3332                         last = 0;
3333                         continue;
3334                 }
3335
3336                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3337                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3338                 cache->dirty = 0;
3339                 last = cache->key.objectid + cache->key.offset;
3340
3341                 err = write_one_cache_group(trans, root, path, cache);
3342                 btrfs_put_block_group(cache);
3343                 if (err) /* File system offline */
3344                         goto out;
3345         }
3346
3347         while (1) {
3348                 /*
3349                  * I don't think this is needed since we're just marking our
3350                  * preallocated extent as written, but just in case it can't
3351                  * hurt.
3352                  */
3353                 if (last == 0) {
3354                         err = btrfs_run_delayed_refs(trans, root,
3355                                                      (unsigned long)-1);
3356                         if (err) /* File system offline */
3357                                 goto out;
3358                 }
3359
3360                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3361                 while (cache) {
3362                         /*
3363                          * Really this shouldn't happen, but it could if we
3364                          * couldn't write the entire preallocated extent and
3365                          * splitting the extent resulted in a new block.
3366                          */
3367                         if (cache->dirty) {
3368                                 btrfs_put_block_group(cache);
3369                                 goto again;
3370                         }
3371                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3372                                 break;
3373                         cache = next_block_group(root, cache);
3374                 }
3375                 if (!cache) {
3376                         if (last == 0)
3377                                 break;
3378                         last = 0;
3379                         continue;
3380                 }
3381
3382                 err = btrfs_write_out_cache(root, trans, cache, path);
3383
3384                 /*
3385                  * If we didn't have an error then the cache state is still
3386                  * NEED_WRITE, so we can set it to WRITTEN.
3387                  */
3388                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3389                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3390                 last = cache->key.objectid + cache->key.offset;
3391                 btrfs_put_block_group(cache);
3392         }
3393 out:
3394
3395         btrfs_free_path(path);
3396         return err;
3397 }
3398
3399 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3400 {
3401         struct btrfs_block_group_cache *block_group;
3402         int readonly = 0;
3403
3404         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3405         if (!block_group || block_group->ro)
3406                 readonly = 1;
3407         if (block_group)
3408                 btrfs_put_block_group(block_group);
3409         return readonly;
3410 }
3411
3412 static const char *alloc_name(u64 flags)
3413 {
3414         switch (flags) {
3415         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3416                 return "mixed";
3417         case BTRFS_BLOCK_GROUP_METADATA:
3418                 return "metadata";
3419         case BTRFS_BLOCK_GROUP_DATA:
3420                 return "data";
3421         case BTRFS_BLOCK_GROUP_SYSTEM:
3422                 return "system";
3423         default:
3424                 WARN_ON(1);
3425                 return "invalid-combination";
3426         };
3427 }
3428
3429 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3430                              u64 total_bytes, u64 bytes_used,
3431                              struct btrfs_space_info **space_info)
3432 {
3433         struct btrfs_space_info *found;
3434         int i;
3435         int factor;
3436         int ret;
3437
3438         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3439                      BTRFS_BLOCK_GROUP_RAID10))
3440                 factor = 2;
3441         else
3442                 factor = 1;
3443
3444         found = __find_space_info(info, flags);
3445         if (found) {
3446                 spin_lock(&found->lock);
3447                 found->total_bytes += total_bytes;
3448                 found->disk_total += total_bytes * factor;
3449                 found->bytes_used += bytes_used;
3450                 found->disk_used += bytes_used * factor;
3451                 found->full = 0;
3452                 spin_unlock(&found->lock);
3453                 *space_info = found;
3454                 return 0;
3455         }
3456         found = kzalloc(sizeof(*found), GFP_NOFS);
3457         if (!found)
3458                 return -ENOMEM;
3459
3460         ret = percpu_counter_init(&found->total_bytes_pinned, 0);
3461         if (ret) {
3462                 kfree(found);
3463                 return ret;
3464         }
3465
3466         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
3467                 INIT_LIST_HEAD(&found->block_groups[i]);
3468                 kobject_init(&found->block_group_kobjs[i], &btrfs_raid_ktype);
3469         }
3470         init_rwsem(&found->groups_sem);
3471         spin_lock_init(&found->lock);
3472         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3473         found->total_bytes = total_bytes;
3474         found->disk_total = total_bytes * factor;
3475         found->bytes_used = bytes_used;
3476         found->disk_used = bytes_used * factor;
3477         found->bytes_pinned = 0;
3478         found->bytes_reserved = 0;
3479         found->bytes_readonly = 0;
3480         found->bytes_may_use = 0;
3481         found->full = 0;
3482         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3483         found->chunk_alloc = 0;
3484         found->flush = 0;
3485         init_waitqueue_head(&found->wait);
3486
3487         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3488                                     info->space_info_kobj, "%s",
3489                                     alloc_name(found->flags));
3490         if (ret) {
3491                 kfree(found);
3492                 return ret;
3493         }
3494
3495         *space_info = found;
3496         list_add_rcu(&found->list, &info->space_info);
3497         if (flags & BTRFS_BLOCK_GROUP_DATA)
3498                 info->data_sinfo = found;
3499
3500         return ret;
3501 }
3502
3503 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3504 {
3505         u64 extra_flags = chunk_to_extended(flags) &
3506                                 BTRFS_EXTENDED_PROFILE_MASK;
3507
3508         write_seqlock(&fs_info->profiles_lock);
3509         if (flags & BTRFS_BLOCK_GROUP_DATA)
3510                 fs_info->avail_data_alloc_bits |= extra_flags;
3511         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3512                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3513         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3514                 fs_info->avail_system_alloc_bits |= extra_flags;
3515         write_sequnlock(&fs_info->profiles_lock);
3516 }
3517
3518 /*
3519  * returns target flags in extended format or 0 if restripe for this
3520  * chunk_type is not in progress
3521  *
3522  * should be called with either volume_mutex or balance_lock held
3523  */
3524 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3525 {
3526         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3527         u64 target = 0;
3528
3529         if (!bctl)
3530                 return 0;
3531
3532         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3533             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3534                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3535         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3536                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3537                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3538         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3539                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3540                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3541         }
3542
3543         return target;
3544 }
3545
3546 /*
3547  * @flags: available profiles in extended format (see ctree.h)
3548  *
3549  * Returns reduced profile in chunk format.  If profile changing is in
3550  * progress (either running or paused) picks the target profile (if it's
3551  * already available), otherwise falls back to plain reducing.
3552  */
3553 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3554 {
3555         /*
3556          * we add in the count of missing devices because we want
3557          * to make sure that any RAID levels on a degraded FS
3558          * continue to be honored.
3559          */
3560         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3561                 root->fs_info->fs_devices->missing_devices;
3562         u64 target;
3563         u64 tmp;
3564
3565         /*
3566          * see if restripe for this chunk_type is in progress, if so
3567          * try to reduce to the target profile
3568          */
3569         spin_lock(&root->fs_info->balance_lock);
3570         target = get_restripe_target(root->fs_info, flags);
3571         if (target) {
3572                 /* pick target profile only if it's already available */
3573                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3574                         spin_unlock(&root->fs_info->balance_lock);
3575                         return extended_to_chunk(target);
3576                 }
3577         }
3578         spin_unlock(&root->fs_info->balance_lock);
3579
3580         /* First, mask out the RAID levels which aren't possible */
3581         if (num_devices == 1)
3582                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3583                            BTRFS_BLOCK_GROUP_RAID5);
3584         if (num_devices < 3)
3585                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3586         if (num_devices < 4)
3587                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3588
3589         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3590                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3591                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3592         flags &= ~tmp;
3593
3594         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3595                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3596         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3597                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3598         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3599                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3600         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3601                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3602         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3603                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3604
3605         return extended_to_chunk(flags | tmp);
3606 }
3607
3608 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3609 {
3610         unsigned seq;
3611
3612         do {
3613                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3614
3615                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3616                         flags |= root->fs_info->avail_data_alloc_bits;
3617                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3618                         flags |= root->fs_info->avail_system_alloc_bits;
3619                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3620                         flags |= root->fs_info->avail_metadata_alloc_bits;
3621         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3622
3623         return btrfs_reduce_alloc_profile(root, flags);
3624 }
3625
3626 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3627 {
3628         u64 flags;
3629         u64 ret;
3630
3631         if (data)
3632                 flags = BTRFS_BLOCK_GROUP_DATA;
3633         else if (root == root->fs_info->chunk_root)
3634                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3635         else
3636                 flags = BTRFS_BLOCK_GROUP_METADATA;
3637
3638         ret = get_alloc_profile(root, flags);
3639         return ret;
3640 }
3641
3642 /*
3643  * This will check the space that the inode allocates from to make sure we have
3644  * enough space for bytes.
3645  */
3646 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3647 {
3648         struct btrfs_space_info *data_sinfo;
3649         struct btrfs_root *root = BTRFS_I(inode)->root;
3650         struct btrfs_fs_info *fs_info = root->fs_info;
3651         u64 used;
3652         int ret = 0, committed = 0, alloc_chunk = 1;
3653
3654         /* make sure bytes are sectorsize aligned */
3655         bytes = ALIGN(bytes, root->sectorsize);
3656
3657         if (btrfs_is_free_space_inode(inode)) {
3658                 committed = 1;
3659                 ASSERT(current->journal_info);
3660         }
3661
3662         data_sinfo = fs_info->data_sinfo;
3663         if (!data_sinfo)
3664                 goto alloc;
3665
3666 again:
3667         /* make sure we have enough space to handle the data first */
3668         spin_lock(&data_sinfo->lock);
3669         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3670                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3671                 data_sinfo->bytes_may_use;
3672
3673         if (used + bytes > data_sinfo->total_bytes) {
3674                 struct btrfs_trans_handle *trans;
3675
3676                 /*
3677                  * if we don't have enough free bytes in this space then we need
3678                  * to alloc a new chunk.
3679                  */
3680                 if (!data_sinfo->full && alloc_chunk) {
3681                         u64 alloc_target;
3682
3683                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3684                         spin_unlock(&data_sinfo->lock);
3685 alloc:
3686                         alloc_target = btrfs_get_alloc_profile(root, 1);
3687                         /*
3688                          * It is ugly that we don't call nolock join
3689                          * transaction for the free space inode case here.
3690                          * But it is safe because we only do the data space
3691                          * reservation for the free space cache in the
3692                          * transaction context, the common join transaction
3693                          * just increase the counter of the current transaction
3694                          * handler, doesn't try to acquire the trans_lock of
3695                          * the fs.
3696                          */
3697                         trans = btrfs_join_transaction(root);
3698                         if (IS_ERR(trans))
3699                                 return PTR_ERR(trans);
3700
3701                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3702                                              alloc_target,
3703                                              CHUNK_ALLOC_NO_FORCE);
3704                         btrfs_end_transaction(trans, root);
3705                         if (ret < 0) {
3706                                 if (ret != -ENOSPC)
3707                                         return ret;
3708                                 else
3709                                         goto commit_trans;
3710                         }
3711
3712                         if (!data_sinfo)
3713                                 data_sinfo = fs_info->data_sinfo;
3714
3715                         goto again;
3716                 }
3717
3718                 /*
3719                  * If we don't have enough pinned space to deal with this
3720                  * allocation don't bother committing the transaction.
3721                  */
3722                 if (percpu_counter_compare(&data_sinfo->total_bytes_pinned,
3723                                            bytes) < 0)
3724                         committed = 1;
3725                 spin_unlock(&data_sinfo->lock);
3726
3727                 /* commit the current transaction and try again */
3728 commit_trans:
3729                 if (!committed &&
3730                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3731                         committed = 1;
3732
3733                         trans = btrfs_join_transaction(root);
3734                         if (IS_ERR(trans))
3735                                 return PTR_ERR(trans);
3736                         ret = btrfs_commit_transaction(trans, root);
3737                         if (ret)
3738                                 return ret;
3739                         goto again;
3740                 }
3741
3742                 trace_btrfs_space_reservation(root->fs_info,
3743                                               "space_info:enospc",
3744                                               data_sinfo->flags, bytes, 1);
3745                 return -ENOSPC;
3746         }
3747         data_sinfo->bytes_may_use += bytes;
3748         trace_btrfs_space_reservation(root->fs_info, "space_info",
3749                                       data_sinfo->flags, bytes, 1);
3750         spin_unlock(&data_sinfo->lock);
3751
3752         return 0;
3753 }
3754
3755 /*
3756  * Called if we need to clear a data reservation for this inode.
3757  */
3758 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3759 {
3760         struct btrfs_root *root = BTRFS_I(inode)->root;
3761         struct btrfs_space_info *data_sinfo;
3762
3763         /* make sure bytes are sectorsize aligned */
3764         bytes = ALIGN(bytes, root->sectorsize);
3765
3766         data_sinfo = root->fs_info->data_sinfo;
3767         spin_lock(&data_sinfo->lock);
3768         WARN_ON(data_sinfo->bytes_may_use < bytes);
3769         data_sinfo->bytes_may_use -= bytes;
3770         trace_btrfs_space_reservation(root->fs_info, "space_info",
3771                                       data_sinfo->flags, bytes, 0);
3772         spin_unlock(&data_sinfo->lock);
3773 }
3774
3775 static void force_metadata_allocation(struct btrfs_fs_info *info)
3776 {
3777         struct list_head *head = &info->space_info;
3778         struct btrfs_space_info *found;
3779
3780         rcu_read_lock();
3781         list_for_each_entry_rcu(found, head, list) {
3782                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3783                         found->force_alloc = CHUNK_ALLOC_FORCE;
3784         }
3785         rcu_read_unlock();
3786 }
3787
3788 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3789 {
3790         return (global->size << 1);
3791 }
3792
3793 static int should_alloc_chunk(struct btrfs_root *root,
3794                               struct btrfs_space_info *sinfo, int force)
3795 {
3796         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3797         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3798         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3799         u64 thresh;
3800
3801         if (force == CHUNK_ALLOC_FORCE)
3802                 return 1;
3803
3804         /*
3805          * We need to take into account the global rsv because for all intents
3806          * and purposes it's used space.  Don't worry about locking the
3807          * global_rsv, it doesn't change except when the transaction commits.
3808          */
3809         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3810                 num_allocated += calc_global_rsv_need_space(global_rsv);
3811
3812         /*
3813          * in limited mode, we want to have some free space up to
3814          * about 1% of the FS size.
3815          */
3816         if (force == CHUNK_ALLOC_LIMITED) {
3817                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3818                 thresh = max_t(u64, 64 * 1024 * 1024,
3819                                div_factor_fine(thresh, 1));
3820
3821                 if (num_bytes - num_allocated < thresh)
3822                         return 1;
3823         }
3824
3825         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3826                 return 0;
3827         return 1;
3828 }
3829
3830 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3831 {
3832         u64 num_dev;
3833
3834         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3835                     BTRFS_BLOCK_GROUP_RAID0 |
3836                     BTRFS_BLOCK_GROUP_RAID5 |
3837                     BTRFS_BLOCK_GROUP_RAID6))
3838                 num_dev = root->fs_info->fs_devices->rw_devices;
3839         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3840                 num_dev = 2;
3841         else
3842                 num_dev = 1;    /* DUP or single */
3843
3844         /* metadata for updaing devices and chunk tree */
3845         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3846 }
3847
3848 static void check_system_chunk(struct btrfs_trans_handle *trans,
3849                                struct btrfs_root *root, u64 type)
3850 {
3851         struct btrfs_space_info *info;
3852         u64 left;
3853         u64 thresh;
3854
3855         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3856         spin_lock(&info->lock);
3857         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3858                 info->bytes_reserved - info->bytes_readonly;
3859         spin_unlock(&info->lock);
3860
3861         thresh = get_system_chunk_thresh(root, type);
3862         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3863                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3864                         left, thresh, type);
3865                 dump_space_info(info, 0, 0);
3866         }
3867
3868         if (left < thresh) {
3869                 u64 flags;
3870
3871                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3872                 btrfs_alloc_chunk(trans, root, flags);
3873         }
3874 }
3875
3876 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3877                           struct btrfs_root *extent_root, u64 flags, int force)
3878 {
3879         struct btrfs_space_info *space_info;
3880         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3881         int wait_for_alloc = 0;
3882         int ret = 0;
3883
3884         /* Don't re-enter if we're already allocating a chunk */
3885         if (trans->allocating_chunk)
3886                 return -ENOSPC;
3887
3888         space_info = __find_space_info(extent_root->fs_info, flags);
3889         if (!space_info) {
3890                 ret = update_space_info(extent_root->fs_info, flags,
3891                                         0, 0, &space_info);
3892                 BUG_ON(ret); /* -ENOMEM */
3893         }
3894         BUG_ON(!space_info); /* Logic error */
3895
3896 again:
3897         spin_lock(&space_info->lock);
3898         if (force < space_info->force_alloc)
3899                 force = space_info->force_alloc;
3900         if (space_info->full) {
3901                 if (should_alloc_chunk(extent_root, space_info, force))
3902                         ret = -ENOSPC;
3903                 else
3904                         ret = 0;
3905                 spin_unlock(&space_info->lock);
3906                 return ret;
3907         }
3908
3909         if (!should_alloc_chunk(extent_root, space_info, force)) {
3910                 spin_unlock(&space_info->lock);
3911                 return 0;
3912         } else if (space_info->chunk_alloc) {
3913                 wait_for_alloc = 1;
3914         } else {
3915                 space_info->chunk_alloc = 1;
3916         }
3917
3918         spin_unlock(&space_info->lock);
3919
3920         mutex_lock(&fs_info->chunk_mutex);
3921
3922         /*
3923          * The chunk_mutex is held throughout the entirety of a chunk
3924          * allocation, so once we've acquired the chunk_mutex we know that the
3925          * other guy is done and we need to recheck and see if we should
3926          * allocate.
3927          */
3928         if (wait_for_alloc) {
3929                 mutex_unlock(&fs_info->chunk_mutex);
3930                 wait_for_alloc = 0;
3931                 goto again;
3932         }
3933
3934         trans->allocating_chunk = true;
3935
3936         /*
3937          * If we have mixed data/metadata chunks we want to make sure we keep
3938          * allocating mixed chunks instead of individual chunks.
3939          */
3940         if (btrfs_mixed_space_info(space_info))
3941                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3942
3943         /*
3944          * if we're doing a data chunk, go ahead and make sure that
3945          * we keep a reasonable number of metadata chunks allocated in the
3946          * FS as well.
3947          */
3948         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3949                 fs_info->data_chunk_allocations++;
3950                 if (!(fs_info->data_chunk_allocations %
3951                       fs_info->metadata_ratio))
3952                         force_metadata_allocation(fs_info);
3953         }
3954
3955         /*
3956          * Check if we have enough space in SYSTEM chunk because we may need
3957          * to update devices.
3958          */
3959         check_system_chunk(trans, extent_root, flags);
3960
3961         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3962         trans->allocating_chunk = false;
3963
3964         spin_lock(&space_info->lock);
3965         if (ret < 0 && ret != -ENOSPC)
3966                 goto out;
3967         if (ret)
3968                 space_info->full = 1;
3969         else
3970                 ret = 1;
3971
3972         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3973 out:
3974         space_info->chunk_alloc = 0;
3975         spin_unlock(&space_info->lock);
3976         mutex_unlock(&fs_info->chunk_mutex);
3977         return ret;
3978 }
3979
3980 static int can_overcommit(struct btrfs_root *root,
3981                           struct btrfs_space_info *space_info, u64 bytes,
3982                           enum btrfs_reserve_flush_enum flush)
3983 {
3984         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3985         u64 profile = btrfs_get_alloc_profile(root, 0);
3986         u64 space_size;
3987         u64 avail;
3988         u64 used;
3989
3990         used = space_info->bytes_used + space_info->bytes_reserved +
3991                 space_info->bytes_pinned + space_info->bytes_readonly;
3992
3993         /*
3994          * We only want to allow over committing if we have lots of actual space
3995          * free, but if we don't have enough space to handle the global reserve
3996          * space then we could end up having a real enospc problem when trying
3997          * to allocate a chunk or some other such important allocation.
3998          */
3999         spin_lock(&global_rsv->lock);
4000         space_size = calc_global_rsv_need_space(global_rsv);
4001         spin_unlock(&global_rsv->lock);
4002         if (used + space_size >= space_info->total_bytes)
4003                 return 0;
4004
4005         used += space_info->bytes_may_use;
4006
4007         spin_lock(&root->fs_info->free_chunk_lock);
4008         avail = root->fs_info->free_chunk_space;
4009         spin_unlock(&root->fs_info->free_chunk_lock);
4010
4011         /*
4012          * If we have dup, raid1 or raid10 then only half of the free
4013          * space is actually useable.  For raid56, the space info used
4014          * doesn't include the parity drive, so we don't have to
4015          * change the math
4016          */
4017         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4018                        BTRFS_BLOCK_GROUP_RAID1 |
4019                        BTRFS_BLOCK_GROUP_RAID10))
4020                 avail >>= 1;
4021
4022         /*
4023          * If we aren't flushing all things, let us overcommit up to
4024          * 1/2th of the space. If we can flush, don't let us overcommit
4025          * too much, let it overcommit up to 1/8 of the space.
4026          */
4027         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4028                 avail >>= 3;
4029         else
4030                 avail >>= 1;
4031
4032         if (used + bytes < space_info->total_bytes + avail)
4033                 return 1;
4034         return 0;
4035 }
4036
4037 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4038                                          unsigned long nr_pages)
4039 {
4040         struct super_block *sb = root->fs_info->sb;
4041
4042         if (down_read_trylock(&sb->s_umount)) {
4043                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4044                 up_read(&sb->s_umount);
4045         } else {
4046                 /*
4047                  * We needn't worry the filesystem going from r/w to r/o though
4048                  * we don't acquire ->s_umount mutex, because the filesystem
4049                  * should guarantee the delalloc inodes list be empty after
4050                  * the filesystem is readonly(all dirty pages are written to
4051                  * the disk).
4052                  */
4053                 btrfs_start_delalloc_roots(root->fs_info, 0);
4054                 if (!current->journal_info)
4055                         btrfs_wait_ordered_roots(root->fs_info, -1);
4056         }
4057 }
4058
4059 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4060 {
4061         u64 bytes;
4062         int nr;
4063
4064         bytes = btrfs_calc_trans_metadata_size(root, 1);
4065         nr = (int)div64_u64(to_reclaim, bytes);
4066         if (!nr)
4067                 nr = 1;
4068         return nr;
4069 }
4070
4071 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4072
4073 /*
4074  * shrink metadata reservation for delalloc
4075  */
4076 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4077                             bool wait_ordered)
4078 {
4079         struct btrfs_block_rsv *block_rsv;
4080         struct btrfs_space_info *space_info;
4081         struct btrfs_trans_handle *trans;
4082         u64 delalloc_bytes;
4083         u64 max_reclaim;
4084         long time_left;
4085         unsigned long nr_pages;
4086         int loops;
4087         int items;
4088         enum btrfs_reserve_flush_enum flush;
4089
4090         /* Calc the number of the pages we need flush for space reservation */
4091         items = calc_reclaim_items_nr(root, to_reclaim);
4092         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4093
4094         trans = (struct btrfs_trans_handle *)current->journal_info;
4095         block_rsv = &root->fs_info->delalloc_block_rsv;
4096         space_info = block_rsv->space_info;
4097
4098         delalloc_bytes = percpu_counter_sum_positive(
4099                                                 &root->fs_info->delalloc_bytes);
4100         if (delalloc_bytes == 0) {
4101                 if (trans)
4102                         return;
4103                 if (wait_ordered)
4104                         btrfs_wait_ordered_roots(root->fs_info, items);
4105                 return;
4106         }
4107
4108         loops = 0;
4109         while (delalloc_bytes && loops < 3) {
4110                 max_reclaim = min(delalloc_bytes, to_reclaim);
4111                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4112                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
4113                 /*
4114                  * We need to wait for the async pages to actually start before
4115                  * we do anything.
4116                  */
4117                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4118                 if (!max_reclaim)
4119                         goto skip_async;
4120
4121                 if (max_reclaim <= nr_pages)
4122                         max_reclaim = 0;
4123                 else
4124                         max_reclaim -= nr_pages;
4125
4126                 wait_event(root->fs_info->async_submit_wait,
4127                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4128                            (int)max_reclaim);
4129 skip_async:
4130                 if (!trans)
4131                         flush = BTRFS_RESERVE_FLUSH_ALL;
4132                 else
4133                         flush = BTRFS_RESERVE_NO_FLUSH;
4134                 spin_lock(&space_info->lock);
4135                 if (can_overcommit(root, space_info, orig, flush)) {
4136                         spin_unlock(&space_info->lock);
4137                         break;
4138                 }
4139                 spin_unlock(&space_info->lock);
4140
4141                 loops++;
4142                 if (wait_ordered && !trans) {
4143                         btrfs_wait_ordered_roots(root->fs_info, items);
4144                 } else {
4145                         time_left = schedule_timeout_killable(1);
4146                         if (time_left)
4147                                 break;
4148                 }
4149                 delalloc_bytes = percpu_counter_sum_positive(
4150                                                 &root->fs_info->delalloc_bytes);
4151         }
4152 }
4153
4154 /**
4155  * maybe_commit_transaction - possibly commit the transaction if its ok to
4156  * @root - the root we're allocating for
4157  * @bytes - the number of bytes we want to reserve
4158  * @force - force the commit
4159  *
4160  * This will check to make sure that committing the transaction will actually
4161  * get us somewhere and then commit the transaction if it does.  Otherwise it
4162  * will return -ENOSPC.
4163  */
4164 static int may_commit_transaction(struct btrfs_root *root,
4165                                   struct btrfs_space_info *space_info,
4166                                   u64 bytes, int force)
4167 {
4168         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4169         struct btrfs_trans_handle *trans;
4170
4171         trans = (struct btrfs_trans_handle *)current->journal_info;
4172         if (trans)
4173                 return -EAGAIN;
4174
4175         if (force)
4176                 goto commit;
4177
4178         /* See if there is enough pinned space to make this reservation */
4179         spin_lock(&space_info->lock);
4180         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4181                                    bytes) >= 0) {
4182                 spin_unlock(&space_info->lock);
4183                 goto commit;
4184         }
4185         spin_unlock(&space_info->lock);
4186
4187         /*
4188          * See if there is some space in the delayed insertion reservation for
4189          * this reservation.
4190          */
4191         if (space_info != delayed_rsv->space_info)
4192                 return -ENOSPC;
4193
4194         spin_lock(&space_info->lock);
4195         spin_lock(&delayed_rsv->lock);
4196         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4197                                    bytes - delayed_rsv->size) >= 0) {
4198                 spin_unlock(&delayed_rsv->lock);
4199                 spin_unlock(&space_info->lock);
4200                 return -ENOSPC;
4201         }
4202         spin_unlock(&delayed_rsv->lock);
4203         spin_unlock(&space_info->lock);
4204
4205 commit:
4206         trans = btrfs_join_transaction(root);
4207         if (IS_ERR(trans))
4208                 return -ENOSPC;
4209
4210         return btrfs_commit_transaction(trans, root);
4211 }
4212
4213 enum flush_state {
4214         FLUSH_DELAYED_ITEMS_NR  =       1,
4215         FLUSH_DELAYED_ITEMS     =       2,
4216         FLUSH_DELALLOC          =       3,
4217         FLUSH_DELALLOC_WAIT     =       4,
4218         ALLOC_CHUNK             =       5,
4219         COMMIT_TRANS            =       6,
4220 };
4221
4222 static int flush_space(struct btrfs_root *root,
4223                        struct btrfs_space_info *space_info, u64 num_bytes,
4224                        u64 orig_bytes, int state)
4225 {
4226         struct btrfs_trans_handle *trans;
4227         int nr;
4228         int ret = 0;
4229
4230         switch (state) {
4231         case FLUSH_DELAYED_ITEMS_NR:
4232         case FLUSH_DELAYED_ITEMS:
4233                 if (state == FLUSH_DELAYED_ITEMS_NR)
4234                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4235                 else
4236                         nr = -1;
4237
4238                 trans = btrfs_join_transaction(root);
4239                 if (IS_ERR(trans)) {
4240                         ret = PTR_ERR(trans);
4241                         break;
4242                 }
4243                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4244                 btrfs_end_transaction(trans, root);
4245                 break;
4246         case FLUSH_DELALLOC:
4247         case FLUSH_DELALLOC_WAIT:
4248                 shrink_delalloc(root, num_bytes, orig_bytes,
4249                                 state == FLUSH_DELALLOC_WAIT);
4250                 break;
4251         case ALLOC_CHUNK:
4252                 trans = btrfs_join_transaction(root);
4253                 if (IS_ERR(trans)) {
4254                         ret = PTR_ERR(trans);
4255                         break;
4256                 }
4257                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4258                                      btrfs_get_alloc_profile(root, 0),
4259                                      CHUNK_ALLOC_NO_FORCE);
4260                 btrfs_end_transaction(trans, root);
4261                 if (ret == -ENOSPC)
4262                         ret = 0;
4263                 break;
4264         case COMMIT_TRANS:
4265                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4266                 break;
4267         default:
4268                 ret = -ENOSPC;
4269                 break;
4270         }
4271
4272         return ret;
4273 }
4274 /**
4275  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4276  * @root - the root we're allocating for
4277  * @block_rsv - the block_rsv we're allocating for
4278  * @orig_bytes - the number of bytes we want
4279  * @flush - whether or not we can flush to make our reservation
4280  *
4281  * This will reserve orgi_bytes number of bytes from the space info associated
4282  * with the block_rsv.  If there is not enough space it will make an attempt to
4283  * flush out space to make room.  It will do this by flushing delalloc if
4284  * possible or committing the transaction.  If flush is 0 then no attempts to
4285  * regain reservations will be made and this will fail if there is not enough
4286  * space already.
4287  */
4288 static int reserve_metadata_bytes(struct btrfs_root *root,
4289                                   struct btrfs_block_rsv *block_rsv,
4290                                   u64 orig_bytes,
4291                                   enum btrfs_reserve_flush_enum flush)
4292 {
4293         struct btrfs_space_info *space_info = block_rsv->space_info;
4294         u64 used;
4295         u64 num_bytes = orig_bytes;
4296         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4297         int ret = 0;
4298         bool flushing = false;
4299
4300 again:
4301         ret = 0;
4302         spin_lock(&space_info->lock);
4303         /*
4304          * We only want to wait if somebody other than us is flushing and we
4305          * are actually allowed to flush all things.
4306          */
4307         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4308                space_info->flush) {
4309                 spin_unlock(&space_info->lock);
4310                 /*
4311                  * If we have a trans handle we can't wait because the flusher
4312                  * may have to commit the transaction, which would mean we would
4313                  * deadlock since we are waiting for the flusher to finish, but
4314                  * hold the current transaction open.
4315                  */
4316                 if (current->journal_info)
4317                         return -EAGAIN;
4318                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4319                 /* Must have been killed, return */
4320                 if (ret)
4321                         return -EINTR;
4322
4323                 spin_lock(&space_info->lock);
4324         }
4325
4326         ret = -ENOSPC;
4327         used = space_info->bytes_used + space_info->bytes_reserved +
4328                 space_info->bytes_pinned + space_info->bytes_readonly +
4329                 space_info->bytes_may_use;
4330
4331         /*
4332          * The idea here is that we've not already over-reserved the block group
4333          * then we can go ahead and save our reservation first and then start
4334          * flushing if we need to.  Otherwise if we've already overcommitted
4335          * lets start flushing stuff first and then come back and try to make
4336          * our reservation.
4337          */
4338         if (used <= space_info->total_bytes) {
4339                 if (used + orig_bytes <= space_info->total_bytes) {
4340                         space_info->bytes_may_use += orig_bytes;
4341                         trace_btrfs_space_reservation(root->fs_info,
4342                                 "space_info", space_info->flags, orig_bytes, 1);
4343                         ret = 0;
4344                 } else {
4345                         /*
4346                          * Ok set num_bytes to orig_bytes since we aren't
4347                          * overocmmitted, this way we only try and reclaim what
4348                          * we need.
4349                          */
4350                         num_bytes = orig_bytes;
4351                 }
4352         } else {
4353                 /*
4354                  * Ok we're over committed, set num_bytes to the overcommitted
4355                  * amount plus the amount of bytes that we need for this
4356                  * reservation.
4357                  */
4358                 num_bytes = used - space_info->total_bytes +
4359                         (orig_bytes * 2);
4360         }
4361
4362         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4363                 space_info->bytes_may_use += orig_bytes;
4364                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4365                                               space_info->flags, orig_bytes,
4366                                               1);
4367                 ret = 0;
4368         }
4369
4370         /*
4371          * Couldn't make our reservation, save our place so while we're trying
4372          * to reclaim space we can actually use it instead of somebody else
4373          * stealing it from us.
4374          *
4375          * We make the other tasks wait for the flush only when we can flush
4376          * all things.
4377          */
4378         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4379                 flushing = true;
4380                 space_info->flush = 1;
4381         }
4382
4383         spin_unlock(&space_info->lock);
4384
4385         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4386                 goto out;
4387
4388         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4389                           flush_state);
4390         flush_state++;
4391
4392         /*
4393          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4394          * would happen. So skip delalloc flush.
4395          */
4396         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4397             (flush_state == FLUSH_DELALLOC ||
4398              flush_state == FLUSH_DELALLOC_WAIT))
4399                 flush_state = ALLOC_CHUNK;
4400
4401         if (!ret)
4402                 goto again;
4403         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4404                  flush_state < COMMIT_TRANS)
4405                 goto again;
4406         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4407                  flush_state <= COMMIT_TRANS)
4408                 goto again;
4409
4410 out:
4411         if (ret == -ENOSPC &&
4412             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4413                 struct btrfs_block_rsv *global_rsv =
4414                         &root->fs_info->global_block_rsv;
4415
4416                 if (block_rsv != global_rsv &&
4417                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4418                         ret = 0;
4419         }
4420         if (ret == -ENOSPC)
4421                 trace_btrfs_space_reservation(root->fs_info,
4422                                               "space_info:enospc",
4423                                               space_info->flags, orig_bytes, 1);
4424         if (flushing) {
4425                 spin_lock(&space_info->lock);
4426                 space_info->flush = 0;
4427                 wake_up_all(&space_info->wait);
4428                 spin_unlock(&space_info->lock);
4429         }
4430         return ret;
4431 }
4432
4433 static struct btrfs_block_rsv *get_block_rsv(
4434                                         const struct btrfs_trans_handle *trans,
4435                                         const struct btrfs_root *root)
4436 {
4437         struct btrfs_block_rsv *block_rsv = NULL;
4438
4439         if (root->ref_cows)
4440                 block_rsv = trans->block_rsv;
4441
4442         if (root == root->fs_info->csum_root && trans->adding_csums)
4443                 block_rsv = trans->block_rsv;
4444
4445         if (root == root->fs_info->uuid_root)
4446                 block_rsv = trans->block_rsv;
4447
4448         if (!block_rsv)
4449                 block_rsv = root->block_rsv;
4450
4451         if (!block_rsv)
4452                 block_rsv = &root->fs_info->empty_block_rsv;
4453
4454         return block_rsv;
4455 }
4456
4457 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4458                                u64 num_bytes)
4459 {
4460         int ret = -ENOSPC;
4461         spin_lock(&block_rsv->lock);
4462         if (block_rsv->reserved >= num_bytes) {
4463                 block_rsv->reserved -= num_bytes;
4464                 if (block_rsv->reserved < block_rsv->size)
4465                         block_rsv->full = 0;
4466                 ret = 0;
4467         }
4468         spin_unlock(&block_rsv->lock);
4469         return ret;
4470 }
4471
4472 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4473                                 u64 num_bytes, int update_size)
4474 {
4475         spin_lock(&block_rsv->lock);
4476         block_rsv->reserved += num_bytes;
4477         if (update_size)
4478                 block_rsv->size += num_bytes;
4479         else if (block_rsv->reserved >= block_rsv->size)
4480                 block_rsv->full = 1;
4481         spin_unlock(&block_rsv->lock);
4482 }
4483
4484 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4485                              struct btrfs_block_rsv *dest, u64 num_bytes,
4486                              int min_factor)
4487 {
4488         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4489         u64 min_bytes;
4490
4491         if (global_rsv->space_info != dest->space_info)
4492                 return -ENOSPC;
4493
4494         spin_lock(&global_rsv->lock);
4495         min_bytes = div_factor(global_rsv->size, min_factor);
4496         if (global_rsv->reserved < min_bytes + num_bytes) {
4497                 spin_unlock(&global_rsv->lock);
4498                 return -ENOSPC;
4499         }
4500         global_rsv->reserved -= num_bytes;
4501         if (global_rsv->reserved < global_rsv->size)
4502                 global_rsv->full = 0;
4503         spin_unlock(&global_rsv->lock);
4504
4505         block_rsv_add_bytes(dest, num_bytes, 1);
4506         return 0;
4507 }
4508
4509 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4510                                     struct btrfs_block_rsv *block_rsv,
4511                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4512 {
4513         struct btrfs_space_info *space_info = block_rsv->space_info;
4514
4515         spin_lock(&block_rsv->lock);
4516         if (num_bytes == (u64)-1)
4517                 num_bytes = block_rsv->size;
4518         block_rsv->size -= num_bytes;
4519         if (block_rsv->reserved >= block_rsv->size) {
4520                 num_bytes = block_rsv->reserved - block_rsv->size;
4521                 block_rsv->reserved = block_rsv->size;
4522                 block_rsv->full = 1;
4523         } else {
4524                 num_bytes = 0;
4525         }
4526         spin_unlock(&block_rsv->lock);
4527
4528         if (num_bytes > 0) {
4529                 if (dest) {
4530                         spin_lock(&dest->lock);
4531                         if (!dest->full) {
4532                                 u64 bytes_to_add;
4533
4534                                 bytes_to_add = dest->size - dest->reserved;
4535                                 bytes_to_add = min(num_bytes, bytes_to_add);
4536                                 dest->reserved += bytes_to_add;
4537                                 if (dest->reserved >= dest->size)
4538                                         dest->full = 1;
4539                                 num_bytes -= bytes_to_add;
4540                         }
4541                         spin_unlock(&dest->lock);
4542                 }
4543                 if (num_bytes) {
4544                         spin_lock(&space_info->lock);
4545                         space_info->bytes_may_use -= num_bytes;
4546                         trace_btrfs_space_reservation(fs_info, "space_info",
4547                                         space_info->flags, num_bytes, 0);
4548                         spin_unlock(&space_info->lock);
4549                 }
4550         }
4551 }
4552
4553 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4554                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4555 {
4556         int ret;
4557
4558         ret = block_rsv_use_bytes(src, num_bytes);
4559         if (ret)
4560                 return ret;
4561
4562         block_rsv_add_bytes(dst, num_bytes, 1);
4563         return 0;
4564 }
4565
4566 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4567 {
4568         memset(rsv, 0, sizeof(*rsv));
4569         spin_lock_init(&rsv->lock);
4570         rsv->type = type;
4571 }
4572
4573 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4574                                               unsigned short type)
4575 {
4576         struct btrfs_block_rsv *block_rsv;
4577         struct btrfs_fs_info *fs_info = root->fs_info;
4578
4579         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4580         if (!block_rsv)
4581                 return NULL;
4582
4583         btrfs_init_block_rsv(block_rsv, type);
4584         block_rsv->space_info = __find_space_info(fs_info,
4585                                                   BTRFS_BLOCK_GROUP_METADATA);
4586         return block_rsv;
4587 }
4588
4589 void btrfs_free_block_rsv(struct btrfs_root *root,
4590                           struct btrfs_block_rsv *rsv)
4591 {
4592         if (!rsv)
4593                 return;
4594         btrfs_block_rsv_release(root, rsv, (u64)-1);
4595         kfree(rsv);
4596 }
4597
4598 int btrfs_block_rsv_add(struct btrfs_root *root,
4599                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4600                         enum btrfs_reserve_flush_enum flush)
4601 {
4602         int ret;
4603
4604         if (num_bytes == 0)
4605                 return 0;
4606
4607         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4608         if (!ret) {
4609                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4610                 return 0;
4611         }
4612
4613         return ret;
4614 }
4615
4616 int btrfs_block_rsv_check(struct btrfs_root *root,
4617                           struct btrfs_block_rsv *block_rsv, int min_factor)
4618 {
4619         u64 num_bytes = 0;
4620         int ret = -ENOSPC;
4621
4622         if (!block_rsv)
4623                 return 0;
4624
4625         spin_lock(&block_rsv->lock);
4626         num_bytes = div_factor(block_rsv->size, min_factor);
4627         if (block_rsv->reserved >= num_bytes)
4628                 ret = 0;
4629         spin_unlock(&block_rsv->lock);
4630
4631         return ret;
4632 }
4633
4634 int btrfs_block_rsv_refill(struct btrfs_root *root,
4635                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4636                            enum btrfs_reserve_flush_enum flush)
4637 {
4638         u64 num_bytes = 0;
4639         int ret = -ENOSPC;
4640
4641         if (!block_rsv)
4642                 return 0;
4643
4644         spin_lock(&block_rsv->lock);
4645         num_bytes = min_reserved;
4646         if (block_rsv->reserved >= num_bytes)
4647                 ret = 0;
4648         else
4649                 num_bytes -= block_rsv->reserved;
4650         spin_unlock(&block_rsv->lock);
4651
4652         if (!ret)
4653                 return 0;
4654
4655         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4656         if (!ret) {
4657                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4658                 return 0;
4659         }
4660
4661         return ret;
4662 }
4663
4664 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4665                             struct btrfs_block_rsv *dst_rsv,
4666                             u64 num_bytes)
4667 {
4668         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4669 }
4670
4671 void btrfs_block_rsv_release(struct btrfs_root *root,
4672                              struct btrfs_block_rsv *block_rsv,
4673                              u64 num_bytes)
4674 {
4675         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4676         if (global_rsv->full || global_rsv == block_rsv ||
4677             block_rsv->space_info != global_rsv->space_info)
4678                 global_rsv = NULL;
4679         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4680                                 num_bytes);
4681 }
4682
4683 /*
4684  * helper to calculate size of global block reservation.
4685  * the desired value is sum of space used by extent tree,
4686  * checksum tree and root tree
4687  */
4688 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4689 {
4690         struct btrfs_space_info *sinfo;
4691         u64 num_bytes;
4692         u64 meta_used;
4693         u64 data_used;
4694         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4695
4696         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4697         spin_lock(&sinfo->lock);
4698         data_used = sinfo->bytes_used;
4699         spin_unlock(&sinfo->lock);
4700
4701         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4702         spin_lock(&sinfo->lock);
4703         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4704                 data_used = 0;
4705         meta_used = sinfo->bytes_used;
4706         spin_unlock(&sinfo->lock);
4707
4708         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4709                     csum_size * 2;
4710         num_bytes += div64_u64(data_used + meta_used, 50);
4711
4712         if (num_bytes * 3 > meta_used)
4713                 num_bytes = div64_u64(meta_used, 3);
4714
4715         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4716 }
4717
4718 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4719 {
4720         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4721         struct btrfs_space_info *sinfo = block_rsv->space_info;
4722         u64 num_bytes;
4723
4724         num_bytes = calc_global_metadata_size(fs_info);
4725
4726         spin_lock(&sinfo->lock);
4727         spin_lock(&block_rsv->lock);
4728
4729         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4730
4731         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4732                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4733                     sinfo->bytes_may_use;
4734
4735         if (sinfo->total_bytes > num_bytes) {
4736                 num_bytes = sinfo->total_bytes - num_bytes;
4737                 block_rsv->reserved += num_bytes;
4738                 sinfo->bytes_may_use += num_bytes;
4739                 trace_btrfs_space_reservation(fs_info, "space_info",
4740                                       sinfo->flags, num_bytes, 1);
4741         }
4742
4743         if (block_rsv->reserved >= block_rsv->size) {
4744                 num_bytes = block_rsv->reserved - block_rsv->size;
4745                 sinfo->bytes_may_use -= num_bytes;
4746                 trace_btrfs_space_reservation(fs_info, "space_info",
4747                                       sinfo->flags, num_bytes, 0);
4748                 block_rsv->reserved = block_rsv->size;
4749                 block_rsv->full = 1;
4750         }
4751
4752         spin_unlock(&block_rsv->lock);
4753         spin_unlock(&sinfo->lock);
4754 }
4755
4756 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4757 {
4758         struct btrfs_space_info *space_info;
4759
4760         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4761         fs_info->chunk_block_rsv.space_info = space_info;
4762
4763         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4764         fs_info->global_block_rsv.space_info = space_info;
4765         fs_info->delalloc_block_rsv.space_info = space_info;
4766         fs_info->trans_block_rsv.space_info = space_info;
4767         fs_info->empty_block_rsv.space_info = space_info;
4768         fs_info->delayed_block_rsv.space_info = space_info;
4769
4770         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4771         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4772         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4773         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4774         if (fs_info->quota_root)
4775                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
4776         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4777
4778         update_global_block_rsv(fs_info);
4779 }
4780
4781 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4782 {
4783         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4784                                 (u64)-1);
4785         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4786         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4787         WARN_ON(fs_info->trans_block_rsv.size > 0);
4788         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4789         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4790         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4791         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4792         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4793 }
4794
4795 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4796                                   struct btrfs_root *root)
4797 {
4798         if (!trans->block_rsv)
4799                 return;
4800
4801         if (!trans->bytes_reserved)
4802                 return;
4803
4804         trace_btrfs_space_reservation(root->fs_info, "transaction",
4805                                       trans->transid, trans->bytes_reserved, 0);
4806         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4807         trans->bytes_reserved = 0;
4808 }
4809
4810 /* Can only return 0 or -ENOSPC */
4811 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4812                                   struct inode *inode)
4813 {
4814         struct btrfs_root *root = BTRFS_I(inode)->root;
4815         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4816         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4817
4818         /*
4819          * We need to hold space in order to delete our orphan item once we've
4820          * added it, so this takes the reservation so we can release it later
4821          * when we are truly done with the orphan item.
4822          */
4823         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4824         trace_btrfs_space_reservation(root->fs_info, "orphan",
4825                                       btrfs_ino(inode), num_bytes, 1);
4826         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4827 }
4828
4829 void btrfs_orphan_release_metadata(struct inode *inode)
4830 {
4831         struct btrfs_root *root = BTRFS_I(inode)->root;
4832         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4833         trace_btrfs_space_reservation(root->fs_info, "orphan",
4834                                       btrfs_ino(inode), num_bytes, 0);
4835         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4836 }
4837
4838 /*
4839  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4840  * root: the root of the parent directory
4841  * rsv: block reservation
4842  * items: the number of items that we need do reservation
4843  * qgroup_reserved: used to return the reserved size in qgroup
4844  *
4845  * This function is used to reserve the space for snapshot/subvolume
4846  * creation and deletion. Those operations are different with the
4847  * common file/directory operations, they change two fs/file trees
4848  * and root tree, the number of items that the qgroup reserves is
4849  * different with the free space reservation. So we can not use
4850  * the space reseravtion mechanism in start_transaction().
4851  */
4852 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4853                                      struct btrfs_block_rsv *rsv,
4854                                      int items,
4855                                      u64 *qgroup_reserved,
4856                                      bool use_global_rsv)
4857 {
4858         u64 num_bytes;
4859         int ret;
4860         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4861
4862         if (root->fs_info->quota_enabled) {
4863                 /* One for parent inode, two for dir entries */
4864                 num_bytes = 3 * root->leafsize;
4865                 ret = btrfs_qgroup_reserve(root, num_bytes);
4866                 if (ret)
4867                         return ret;
4868         } else {
4869                 num_bytes = 0;
4870         }
4871
4872         *qgroup_reserved = num_bytes;
4873
4874         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4875         rsv->space_info = __find_space_info(root->fs_info,
4876                                             BTRFS_BLOCK_GROUP_METADATA);
4877         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4878                                   BTRFS_RESERVE_FLUSH_ALL);
4879
4880         if (ret == -ENOSPC && use_global_rsv)
4881                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
4882
4883         if (ret) {
4884                 if (*qgroup_reserved)
4885                         btrfs_qgroup_free(root, *qgroup_reserved);
4886         }
4887
4888         return ret;
4889 }
4890
4891 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4892                                       struct btrfs_block_rsv *rsv,
4893                                       u64 qgroup_reserved)
4894 {
4895         btrfs_block_rsv_release(root, rsv, (u64)-1);
4896         if (qgroup_reserved)
4897                 btrfs_qgroup_free(root, qgroup_reserved);
4898 }
4899
4900 /**
4901  * drop_outstanding_extent - drop an outstanding extent
4902  * @inode: the inode we're dropping the extent for
4903  *
4904  * This is called when we are freeing up an outstanding extent, either called
4905  * after an error or after an extent is written.  This will return the number of
4906  * reserved extents that need to be freed.  This must be called with
4907  * BTRFS_I(inode)->lock held.
4908  */
4909 static unsigned drop_outstanding_extent(struct inode *inode)
4910 {
4911         unsigned drop_inode_space = 0;
4912         unsigned dropped_extents = 0;
4913
4914         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4915         BTRFS_I(inode)->outstanding_extents--;
4916
4917         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4918             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4919                                &BTRFS_I(inode)->runtime_flags))
4920                 drop_inode_space = 1;
4921
4922         /*
4923          * If we have more or the same amount of outsanding extents than we have
4924          * reserved then we need to leave the reserved extents count alone.
4925          */
4926         if (BTRFS_I(inode)->outstanding_extents >=
4927             BTRFS_I(inode)->reserved_extents)
4928                 return drop_inode_space;
4929
4930         dropped_extents = BTRFS_I(inode)->reserved_extents -
4931                 BTRFS_I(inode)->outstanding_extents;
4932         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4933         return dropped_extents + drop_inode_space;
4934 }
4935
4936 /**
4937  * calc_csum_metadata_size - return the amount of metada space that must be
4938  *      reserved/free'd for the given bytes.
4939  * @inode: the inode we're manipulating
4940  * @num_bytes: the number of bytes in question
4941  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4942  *
4943  * This adjusts the number of csum_bytes in the inode and then returns the
4944  * correct amount of metadata that must either be reserved or freed.  We
4945  * calculate how many checksums we can fit into one leaf and then divide the
4946  * number of bytes that will need to be checksumed by this value to figure out
4947  * how many checksums will be required.  If we are adding bytes then the number
4948  * may go up and we will return the number of additional bytes that must be
4949  * reserved.  If it is going down we will return the number of bytes that must
4950  * be freed.
4951  *
4952  * This must be called with BTRFS_I(inode)->lock held.
4953  */
4954 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4955                                    int reserve)
4956 {
4957         struct btrfs_root *root = BTRFS_I(inode)->root;
4958         u64 csum_size;
4959         int num_csums_per_leaf;
4960         int num_csums;
4961         int old_csums;
4962
4963         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4964             BTRFS_I(inode)->csum_bytes == 0)
4965                 return 0;
4966
4967         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4968         if (reserve)
4969                 BTRFS_I(inode)->csum_bytes += num_bytes;
4970         else
4971                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4972         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4973         num_csums_per_leaf = (int)div64_u64(csum_size,
4974                                             sizeof(struct btrfs_csum_item) +
4975                                             sizeof(struct btrfs_disk_key));
4976         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4977         num_csums = num_csums + num_csums_per_leaf - 1;
4978         num_csums = num_csums / num_csums_per_leaf;
4979
4980         old_csums = old_csums + num_csums_per_leaf - 1;
4981         old_csums = old_csums / num_csums_per_leaf;
4982
4983         /* No change, no need to reserve more */
4984         if (old_csums == num_csums)
4985                 return 0;
4986
4987         if (reserve)
4988                 return btrfs_calc_trans_metadata_size(root,
4989                                                       num_csums - old_csums);
4990
4991         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4992 }
4993
4994 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4995 {
4996         struct btrfs_root *root = BTRFS_I(inode)->root;
4997         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4998         u64 to_reserve = 0;
4999         u64 csum_bytes;
5000         unsigned nr_extents = 0;
5001         int extra_reserve = 0;
5002         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5003         int ret = 0;
5004         bool delalloc_lock = true;
5005         u64 to_free = 0;
5006         unsigned dropped;
5007
5008         /* If we are a free space inode we need to not flush since we will be in
5009          * the middle of a transaction commit.  We also don't need the delalloc
5010          * mutex since we won't race with anybody.  We need this mostly to make
5011          * lockdep shut its filthy mouth.
5012          */
5013         if (btrfs_is_free_space_inode(inode)) {
5014                 flush = BTRFS_RESERVE_NO_FLUSH;
5015                 delalloc_lock = false;
5016         }
5017
5018         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5019             btrfs_transaction_in_commit(root->fs_info))
5020                 schedule_timeout(1);
5021
5022         if (delalloc_lock)
5023                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5024
5025         num_bytes = ALIGN(num_bytes, root->sectorsize);
5026
5027         spin_lock(&BTRFS_I(inode)->lock);
5028         BTRFS_I(inode)->outstanding_extents++;
5029
5030         if (BTRFS_I(inode)->outstanding_extents >
5031             BTRFS_I(inode)->reserved_extents)
5032                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5033                         BTRFS_I(inode)->reserved_extents;
5034
5035         /*
5036          * Add an item to reserve for updating the inode when we complete the
5037          * delalloc io.
5038          */
5039         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5040                       &BTRFS_I(inode)->runtime_flags)) {
5041                 nr_extents++;
5042                 extra_reserve = 1;
5043         }
5044
5045         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5046         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5047         csum_bytes = BTRFS_I(inode)->csum_bytes;
5048         spin_unlock(&BTRFS_I(inode)->lock);
5049
5050         if (root->fs_info->quota_enabled) {
5051                 ret = btrfs_qgroup_reserve(root, num_bytes +
5052                                            nr_extents * root->leafsize);
5053                 if (ret)
5054                         goto out_fail;
5055         }
5056
5057         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5058         if (unlikely(ret)) {
5059                 if (root->fs_info->quota_enabled)
5060                         btrfs_qgroup_free(root, num_bytes +
5061                                                 nr_extents * root->leafsize);
5062                 goto out_fail;
5063         }
5064
5065         spin_lock(&BTRFS_I(inode)->lock);
5066         if (extra_reserve) {
5067                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5068                         &BTRFS_I(inode)->runtime_flags);
5069                 nr_extents--;
5070         }
5071         BTRFS_I(inode)->reserved_extents += nr_extents;
5072         spin_unlock(&BTRFS_I(inode)->lock);
5073
5074         if (delalloc_lock)
5075                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5076
5077         if (to_reserve)
5078                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5079                                               btrfs_ino(inode), to_reserve, 1);
5080         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5081
5082         return 0;
5083
5084 out_fail:
5085         spin_lock(&BTRFS_I(inode)->lock);
5086         dropped = drop_outstanding_extent(inode);
5087         /*
5088          * If the inodes csum_bytes is the same as the original
5089          * csum_bytes then we know we haven't raced with any free()ers
5090          * so we can just reduce our inodes csum bytes and carry on.
5091          */
5092         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5093                 calc_csum_metadata_size(inode, num_bytes, 0);
5094         } else {
5095                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5096                 u64 bytes;
5097
5098                 /*
5099                  * This is tricky, but first we need to figure out how much we
5100                  * free'd from any free-ers that occured during this
5101                  * reservation, so we reset ->csum_bytes to the csum_bytes
5102                  * before we dropped our lock, and then call the free for the
5103                  * number of bytes that were freed while we were trying our
5104                  * reservation.
5105                  */
5106                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5107                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5108                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5109
5110
5111                 /*
5112                  * Now we need to see how much we would have freed had we not
5113                  * been making this reservation and our ->csum_bytes were not
5114                  * artificially inflated.
5115                  */
5116                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5117                 bytes = csum_bytes - orig_csum_bytes;
5118                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5119
5120                 /*
5121                  * Now reset ->csum_bytes to what it should be.  If bytes is
5122                  * more than to_free then we would have free'd more space had we
5123                  * not had an artificially high ->csum_bytes, so we need to free
5124                  * the remainder.  If bytes is the same or less then we don't
5125                  * need to do anything, the other free-ers did the correct
5126                  * thing.
5127                  */
5128                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5129                 if (bytes > to_free)
5130                         to_free = bytes - to_free;
5131                 else
5132                         to_free = 0;
5133         }
5134         spin_unlock(&BTRFS_I(inode)->lock);
5135         if (dropped)
5136                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5137
5138         if (to_free) {
5139                 btrfs_block_rsv_release(root, block_rsv, to_free);
5140                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5141                                               btrfs_ino(inode), to_free, 0);
5142         }
5143         if (delalloc_lock)
5144                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5145         return ret;
5146 }
5147
5148 /**
5149  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5150  * @inode: the inode to release the reservation for
5151  * @num_bytes: the number of bytes we're releasing
5152  *
5153  * This will release the metadata reservation for an inode.  This can be called
5154  * once we complete IO for a given set of bytes to release their metadata
5155  * reservations.
5156  */
5157 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5158 {
5159         struct btrfs_root *root = BTRFS_I(inode)->root;
5160         u64 to_free = 0;
5161         unsigned dropped;
5162
5163         num_bytes = ALIGN(num_bytes, root->sectorsize);
5164         spin_lock(&BTRFS_I(inode)->lock);
5165         dropped = drop_outstanding_extent(inode);
5166
5167         if (num_bytes)
5168                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5169         spin_unlock(&BTRFS_I(inode)->lock);
5170         if (dropped > 0)
5171                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5172
5173         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5174                                       btrfs_ino(inode), to_free, 0);
5175         if (root->fs_info->quota_enabled) {
5176                 btrfs_qgroup_free(root, num_bytes +
5177                                         dropped * root->leafsize);
5178         }
5179
5180         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5181                                 to_free);
5182 }
5183
5184 /**
5185  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5186  * @inode: inode we're writing to
5187  * @num_bytes: the number of bytes we want to allocate
5188  *
5189  * This will do the following things
5190  *
5191  * o reserve space in the data space info for num_bytes
5192  * o reserve space in the metadata space info based on number of outstanding
5193  *   extents and how much csums will be needed
5194  * o add to the inodes ->delalloc_bytes
5195  * o add it to the fs_info's delalloc inodes list.
5196  *
5197  * This will return 0 for success and -ENOSPC if there is no space left.
5198  */
5199 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5200 {
5201         int ret;
5202
5203         ret = btrfs_check_data_free_space(inode, num_bytes);
5204         if (ret)
5205                 return ret;
5206
5207         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5208         if (ret) {
5209                 btrfs_free_reserved_data_space(inode, num_bytes);
5210                 return ret;
5211         }
5212
5213         return 0;
5214 }
5215
5216 /**
5217  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5218  * @inode: inode we're releasing space for
5219  * @num_bytes: the number of bytes we want to free up
5220  *
5221  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5222  * called in the case that we don't need the metadata AND data reservations
5223  * anymore.  So if there is an error or we insert an inline extent.
5224  *
5225  * This function will release the metadata space that was not used and will
5226  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5227  * list if there are no delalloc bytes left.
5228  */
5229 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5230 {
5231         btrfs_delalloc_release_metadata(inode, num_bytes);
5232         btrfs_free_reserved_data_space(inode, num_bytes);
5233 }
5234
5235 static int update_block_group(struct btrfs_root *root,
5236                               u64 bytenr, u64 num_bytes, int alloc)
5237 {
5238         struct btrfs_block_group_cache *cache = NULL;
5239         struct btrfs_fs_info *info = root->fs_info;
5240         u64 total = num_bytes;
5241         u64 old_val;
5242         u64 byte_in_group;
5243         int factor;
5244
5245         /* block accounting for super block */
5246         spin_lock(&info->delalloc_root_lock);
5247         old_val = btrfs_super_bytes_used(info->super_copy);
5248         if (alloc)
5249                 old_val += num_bytes;
5250         else
5251                 old_val -= num_bytes;
5252         btrfs_set_super_bytes_used(info->super_copy, old_val);
5253         spin_unlock(&info->delalloc_root_lock);
5254
5255         while (total) {
5256                 cache = btrfs_lookup_block_group(info, bytenr);
5257                 if (!cache)
5258                         return -ENOENT;
5259                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5260                                     BTRFS_BLOCK_GROUP_RAID1 |
5261                                     BTRFS_BLOCK_GROUP_RAID10))
5262                         factor = 2;
5263                 else
5264                         factor = 1;
5265                 /*
5266                  * If this block group has free space cache written out, we
5267                  * need to make sure to load it if we are removing space.  This
5268                  * is because we need the unpinning stage to actually add the
5269                  * space back to the block group, otherwise we will leak space.
5270                  */
5271                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5272                         cache_block_group(cache, 1);
5273
5274                 byte_in_group = bytenr - cache->key.objectid;
5275                 WARN_ON(byte_in_group > cache->key.offset);
5276
5277                 spin_lock(&cache->space_info->lock);
5278                 spin_lock(&cache->lock);
5279
5280                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5281                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5282                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5283
5284                 cache->dirty = 1;
5285                 old_val = btrfs_block_group_used(&cache->item);
5286                 num_bytes = min(total, cache->key.offset - byte_in_group);
5287                 if (alloc) {
5288                         old_val += num_bytes;
5289                         btrfs_set_block_group_used(&cache->item, old_val);
5290                         cache->reserved -= num_bytes;
5291                         cache->space_info->bytes_reserved -= num_bytes;
5292                         cache->space_info->bytes_used += num_bytes;
5293                         cache->space_info->disk_used += num_bytes * factor;
5294                         spin_unlock(&cache->lock);
5295                         spin_unlock(&cache->space_info->lock);
5296                 } else {
5297                         old_val -= num_bytes;
5298                         btrfs_set_block_group_used(&cache->item, old_val);
5299                         cache->pinned += num_bytes;
5300                         cache->space_info->bytes_pinned += num_bytes;
5301                         cache->space_info->bytes_used -= num_bytes;
5302                         cache->space_info->disk_used -= num_bytes * factor;
5303                         spin_unlock(&cache->lock);
5304                         spin_unlock(&cache->space_info->lock);
5305
5306                         set_extent_dirty(info->pinned_extents,
5307                                          bytenr, bytenr + num_bytes - 1,
5308                                          GFP_NOFS | __GFP_NOFAIL);
5309                 }
5310                 btrfs_put_block_group(cache);
5311                 total -= num_bytes;
5312                 bytenr += num_bytes;
5313         }
5314         return 0;
5315 }
5316
5317 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5318 {
5319         struct btrfs_block_group_cache *cache;
5320         u64 bytenr;
5321
5322         spin_lock(&root->fs_info->block_group_cache_lock);
5323         bytenr = root->fs_info->first_logical_byte;
5324         spin_unlock(&root->fs_info->block_group_cache_lock);
5325
5326         if (bytenr < (u64)-1)
5327                 return bytenr;
5328
5329         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5330         if (!cache)
5331                 return 0;
5332
5333         bytenr = cache->key.objectid;
5334         btrfs_put_block_group(cache);
5335
5336         return bytenr;
5337 }
5338
5339 static int pin_down_extent(struct btrfs_root *root,
5340                            struct btrfs_block_group_cache *cache,
5341                            u64 bytenr, u64 num_bytes, int reserved)
5342 {
5343         spin_lock(&cache->space_info->lock);
5344         spin_lock(&cache->lock);
5345         cache->pinned += num_bytes;
5346         cache->space_info->bytes_pinned += num_bytes;
5347         if (reserved) {
5348                 cache->reserved -= num_bytes;
5349                 cache->space_info->bytes_reserved -= num_bytes;
5350         }
5351         spin_unlock(&cache->lock);
5352         spin_unlock(&cache->space_info->lock);
5353
5354         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5355                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5356         if (reserved)
5357                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5358         return 0;
5359 }
5360
5361 /*
5362  * this function must be called within transaction
5363  */
5364 int btrfs_pin_extent(struct btrfs_root *root,
5365                      u64 bytenr, u64 num_bytes, int reserved)
5366 {
5367         struct btrfs_block_group_cache *cache;
5368
5369         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5370         BUG_ON(!cache); /* Logic error */
5371
5372         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5373
5374         btrfs_put_block_group(cache);
5375         return 0;
5376 }
5377
5378 /*
5379  * this function must be called within transaction
5380  */
5381 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5382                                     u64 bytenr, u64 num_bytes)
5383 {
5384         struct btrfs_block_group_cache *cache;
5385         int ret;
5386
5387         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5388         if (!cache)
5389                 return -EINVAL;
5390
5391         /*
5392          * pull in the free space cache (if any) so that our pin
5393          * removes the free space from the cache.  We have load_only set
5394          * to one because the slow code to read in the free extents does check
5395          * the pinned extents.
5396          */
5397         cache_block_group(cache, 1);
5398
5399         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5400
5401         /* remove us from the free space cache (if we're there at all) */
5402         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5403         btrfs_put_block_group(cache);
5404         return ret;
5405 }
5406
5407 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5408 {
5409         int ret;
5410         struct btrfs_block_group_cache *block_group;
5411         struct btrfs_caching_control *caching_ctl;
5412
5413         block_group = btrfs_lookup_block_group(root->fs_info, start);
5414         if (!block_group)
5415                 return -EINVAL;
5416
5417         cache_block_group(block_group, 0);
5418         caching_ctl = get_caching_control(block_group);
5419
5420         if (!caching_ctl) {
5421                 /* Logic error */
5422                 BUG_ON(!block_group_cache_done(block_group));
5423                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5424         } else {
5425                 mutex_lock(&caching_ctl->mutex);
5426
5427                 if (start >= caching_ctl->progress) {
5428                         ret = add_excluded_extent(root, start, num_bytes);
5429                 } else if (start + num_bytes <= caching_ctl->progress) {
5430                         ret = btrfs_remove_free_space(block_group,
5431                                                       start, num_bytes);
5432                 } else {
5433                         num_bytes = caching_ctl->progress - start;
5434                         ret = btrfs_remove_free_space(block_group,
5435                                                       start, num_bytes);
5436                         if (ret)
5437                                 goto out_lock;
5438
5439                         num_bytes = (start + num_bytes) -
5440                                 caching_ctl->progress;
5441                         start = caching_ctl->progress;
5442                         ret = add_excluded_extent(root, start, num_bytes);
5443                 }
5444 out_lock:
5445                 mutex_unlock(&caching_ctl->mutex);
5446                 put_caching_control(caching_ctl);
5447         }
5448         btrfs_put_block_group(block_group);
5449         return ret;
5450 }
5451
5452 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5453                                  struct extent_buffer *eb)
5454 {
5455         struct btrfs_file_extent_item *item;
5456         struct btrfs_key key;
5457         int found_type;
5458         int i;
5459
5460         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5461                 return 0;
5462
5463         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5464                 btrfs_item_key_to_cpu(eb, &key, i);
5465                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5466                         continue;
5467                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5468                 found_type = btrfs_file_extent_type(eb, item);
5469                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5470                         continue;
5471                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5472                         continue;
5473                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5474                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5475                 __exclude_logged_extent(log, key.objectid, key.offset);
5476         }
5477
5478         return 0;
5479 }
5480
5481 /**
5482  * btrfs_update_reserved_bytes - update the block_group and space info counters
5483  * @cache:      The cache we are manipulating
5484  * @num_bytes:  The number of bytes in question
5485  * @reserve:    One of the reservation enums
5486  *
5487  * This is called by the allocator when it reserves space, or by somebody who is
5488  * freeing space that was never actually used on disk.  For example if you
5489  * reserve some space for a new leaf in transaction A and before transaction A
5490  * commits you free that leaf, you call this with reserve set to 0 in order to
5491  * clear the reservation.
5492  *
5493  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5494  * ENOSPC accounting.  For data we handle the reservation through clearing the
5495  * delalloc bits in the io_tree.  We have to do this since we could end up
5496  * allocating less disk space for the amount of data we have reserved in the
5497  * case of compression.
5498  *
5499  * If this is a reservation and the block group has become read only we cannot
5500  * make the reservation and return -EAGAIN, otherwise this function always
5501  * succeeds.
5502  */
5503 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5504                                        u64 num_bytes, int reserve)
5505 {
5506         struct btrfs_space_info *space_info = cache->space_info;
5507         int ret = 0;
5508
5509         spin_lock(&space_info->lock);
5510         spin_lock(&cache->lock);
5511         if (reserve != RESERVE_FREE) {
5512                 if (cache->ro) {
5513                         ret = -EAGAIN;
5514                 } else {
5515                         cache->reserved += num_bytes;
5516                         space_info->bytes_reserved += num_bytes;
5517                         if (reserve == RESERVE_ALLOC) {
5518                                 trace_btrfs_space_reservation(cache->fs_info,
5519                                                 "space_info", space_info->flags,
5520                                                 num_bytes, 0);
5521                                 space_info->bytes_may_use -= num_bytes;
5522                         }
5523                 }
5524         } else {
5525                 if (cache->ro)
5526                         space_info->bytes_readonly += num_bytes;
5527                 cache->reserved -= num_bytes;
5528                 space_info->bytes_reserved -= num_bytes;
5529         }
5530         spin_unlock(&cache->lock);
5531         spin_unlock(&space_info->lock);
5532         return ret;
5533 }
5534
5535 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5536                                 struct btrfs_root *root)
5537 {
5538         struct btrfs_fs_info *fs_info = root->fs_info;
5539         struct btrfs_caching_control *next;
5540         struct btrfs_caching_control *caching_ctl;
5541         struct btrfs_block_group_cache *cache;
5542         struct btrfs_space_info *space_info;
5543
5544         down_write(&fs_info->extent_commit_sem);
5545
5546         list_for_each_entry_safe(caching_ctl, next,
5547                                  &fs_info->caching_block_groups, list) {
5548                 cache = caching_ctl->block_group;
5549                 if (block_group_cache_done(cache)) {
5550                         cache->last_byte_to_unpin = (u64)-1;
5551                         list_del_init(&caching_ctl->list);
5552                         put_caching_control(caching_ctl);
5553                 } else {
5554                         cache->last_byte_to_unpin = caching_ctl->progress;
5555                 }
5556         }
5557
5558         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5559                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5560         else
5561                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5562
5563         up_write(&fs_info->extent_commit_sem);
5564
5565         list_for_each_entry_rcu(space_info, &fs_info->space_info, list)
5566                 percpu_counter_set(&space_info->total_bytes_pinned, 0);
5567
5568         update_global_block_rsv(fs_info);
5569 }
5570
5571 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5572 {
5573         struct btrfs_fs_info *fs_info = root->fs_info;
5574         struct btrfs_block_group_cache *cache = NULL;
5575         struct btrfs_space_info *space_info;
5576         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5577         u64 len;
5578         bool readonly;
5579
5580         while (start <= end) {
5581                 readonly = false;
5582                 if (!cache ||
5583                     start >= cache->key.objectid + cache->key.offset) {
5584                         if (cache)
5585                                 btrfs_put_block_group(cache);
5586                         cache = btrfs_lookup_block_group(fs_info, start);
5587                         BUG_ON(!cache); /* Logic error */
5588                 }
5589
5590                 len = cache->key.objectid + cache->key.offset - start;
5591                 len = min(len, end + 1 - start);
5592
5593                 if (start < cache->last_byte_to_unpin) {
5594                         len = min(len, cache->last_byte_to_unpin - start);
5595                         btrfs_add_free_space(cache, start, len);
5596                 }
5597
5598                 start += len;
5599                 space_info = cache->space_info;
5600
5601                 spin_lock(&space_info->lock);
5602                 spin_lock(&cache->lock);
5603                 cache->pinned -= len;
5604                 space_info->bytes_pinned -= len;
5605                 if (cache->ro) {
5606                         space_info->bytes_readonly += len;
5607                         readonly = true;
5608                 }
5609                 spin_unlock(&cache->lock);
5610                 if (!readonly && global_rsv->space_info == space_info) {
5611                         spin_lock(&global_rsv->lock);
5612                         if (!global_rsv->full) {
5613                                 len = min(len, global_rsv->size -
5614                                           global_rsv->reserved);
5615                                 global_rsv->reserved += len;
5616                                 space_info->bytes_may_use += len;
5617                                 if (global_rsv->reserved >= global_rsv->size)
5618                                         global_rsv->full = 1;
5619                         }
5620                         spin_unlock(&global_rsv->lock);
5621                 }
5622                 spin_unlock(&space_info->lock);
5623         }
5624
5625         if (cache)
5626                 btrfs_put_block_group(cache);
5627         return 0;
5628 }
5629
5630 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5631                                struct btrfs_root *root)
5632 {
5633         struct btrfs_fs_info *fs_info = root->fs_info;
5634         struct extent_io_tree *unpin;
5635         u64 start;
5636         u64 end;
5637         int ret;
5638
5639         if (trans->aborted)
5640                 return 0;
5641
5642         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5643                 unpin = &fs_info->freed_extents[1];
5644         else
5645                 unpin = &fs_info->freed_extents[0];
5646
5647         while (1) {
5648                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5649                                             EXTENT_DIRTY, NULL);
5650                 if (ret)
5651                         break;
5652
5653                 if (btrfs_test_opt(root, DISCARD))
5654                         ret = btrfs_discard_extent(root, start,
5655                                                    end + 1 - start, NULL);
5656
5657                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5658                 unpin_extent_range(root, start, end);
5659                 cond_resched();
5660         }
5661
5662         return 0;
5663 }
5664
5665 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
5666                              u64 owner, u64 root_objectid)
5667 {
5668         struct btrfs_space_info *space_info;
5669         u64 flags;
5670
5671         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5672                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
5673                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
5674                 else
5675                         flags = BTRFS_BLOCK_GROUP_METADATA;
5676         } else {
5677                 flags = BTRFS_BLOCK_GROUP_DATA;
5678         }
5679
5680         space_info = __find_space_info(fs_info, flags);
5681         BUG_ON(!space_info); /* Logic bug */
5682         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
5683 }
5684
5685
5686 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5687                                 struct btrfs_root *root,
5688                                 u64 bytenr, u64 num_bytes, u64 parent,
5689                                 u64 root_objectid, u64 owner_objectid,
5690                                 u64 owner_offset, int refs_to_drop,
5691                                 struct btrfs_delayed_extent_op *extent_op)
5692 {
5693         struct btrfs_key key;
5694         struct btrfs_path *path;
5695         struct btrfs_fs_info *info = root->fs_info;
5696         struct btrfs_root *extent_root = info->extent_root;
5697         struct extent_buffer *leaf;
5698         struct btrfs_extent_item *ei;
5699         struct btrfs_extent_inline_ref *iref;
5700         int ret;
5701         int is_data;
5702         int extent_slot = 0;
5703         int found_extent = 0;
5704         int num_to_del = 1;
5705         u32 item_size;
5706         u64 refs;
5707         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5708                                                  SKINNY_METADATA);
5709
5710         path = btrfs_alloc_path();
5711         if (!path)
5712                 return -ENOMEM;
5713
5714         path->reada = 1;
5715         path->leave_spinning = 1;
5716
5717         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5718         BUG_ON(!is_data && refs_to_drop != 1);
5719
5720         if (is_data)
5721                 skinny_metadata = 0;
5722
5723         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5724                                     bytenr, num_bytes, parent,
5725                                     root_objectid, owner_objectid,
5726                                     owner_offset);
5727         if (ret == 0) {
5728                 extent_slot = path->slots[0];
5729                 while (extent_slot >= 0) {
5730                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5731                                               extent_slot);
5732                         if (key.objectid != bytenr)
5733                                 break;
5734                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5735                             key.offset == num_bytes) {
5736                                 found_extent = 1;
5737                                 break;
5738                         }
5739                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5740                             key.offset == owner_objectid) {
5741                                 found_extent = 1;
5742                                 break;
5743                         }
5744                         if (path->slots[0] - extent_slot > 5)
5745                                 break;
5746                         extent_slot--;
5747                 }
5748 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5749                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5750                 if (found_extent && item_size < sizeof(*ei))
5751                         found_extent = 0;
5752 #endif
5753                 if (!found_extent) {
5754                         BUG_ON(iref);
5755                         ret = remove_extent_backref(trans, extent_root, path,
5756                                                     NULL, refs_to_drop,
5757                                                     is_data);
5758                         if (ret) {
5759                                 btrfs_abort_transaction(trans, extent_root, ret);
5760                                 goto out;
5761                         }
5762                         btrfs_release_path(path);
5763                         path->leave_spinning = 1;
5764
5765                         key.objectid = bytenr;
5766                         key.type = BTRFS_EXTENT_ITEM_KEY;
5767                         key.offset = num_bytes;
5768
5769                         if (!is_data && skinny_metadata) {
5770                                 key.type = BTRFS_METADATA_ITEM_KEY;
5771                                 key.offset = owner_objectid;
5772                         }
5773
5774                         ret = btrfs_search_slot(trans, extent_root,
5775                                                 &key, path, -1, 1);
5776                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5777                                 /*
5778                                  * Couldn't find our skinny metadata item,
5779                                  * see if we have ye olde extent item.
5780                                  */
5781                                 path->slots[0]--;
5782                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5783                                                       path->slots[0]);
5784                                 if (key.objectid == bytenr &&
5785                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5786                                     key.offset == num_bytes)
5787                                         ret = 0;
5788                         }
5789
5790                         if (ret > 0 && skinny_metadata) {
5791                                 skinny_metadata = false;
5792                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5793                                 key.offset = num_bytes;
5794                                 btrfs_release_path(path);
5795                                 ret = btrfs_search_slot(trans, extent_root,
5796                                                         &key, path, -1, 1);
5797                         }
5798
5799                         if (ret) {
5800                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5801                                         ret, bytenr);
5802                                 if (ret > 0)
5803                                         btrfs_print_leaf(extent_root,
5804                                                          path->nodes[0]);
5805                         }
5806                         if (ret < 0) {
5807                                 btrfs_abort_transaction(trans, extent_root, ret);
5808                                 goto out;
5809                         }
5810                         extent_slot = path->slots[0];
5811                 }
5812         } else if (WARN_ON(ret == -ENOENT)) {
5813                 btrfs_print_leaf(extent_root, path->nodes[0]);
5814                 btrfs_err(info,
5815                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5816                         bytenr, parent, root_objectid, owner_objectid,
5817                         owner_offset);
5818         } else {
5819                 btrfs_abort_transaction(trans, extent_root, ret);
5820                 goto out;
5821         }
5822
5823         leaf = path->nodes[0];
5824         item_size = btrfs_item_size_nr(leaf, extent_slot);
5825 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5826         if (item_size < sizeof(*ei)) {
5827                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5828                 ret = convert_extent_item_v0(trans, extent_root, path,
5829                                              owner_objectid, 0);
5830                 if (ret < 0) {
5831                         btrfs_abort_transaction(trans, extent_root, ret);
5832                         goto out;
5833                 }
5834
5835                 btrfs_release_path(path);
5836                 path->leave_spinning = 1;
5837
5838                 key.objectid = bytenr;
5839                 key.type = BTRFS_EXTENT_ITEM_KEY;
5840                 key.offset = num_bytes;
5841
5842                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5843                                         -1, 1);
5844                 if (ret) {
5845                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5846                                 ret, bytenr);
5847                         btrfs_print_leaf(extent_root, path->nodes[0]);
5848                 }
5849                 if (ret < 0) {
5850                         btrfs_abort_transaction(trans, extent_root, ret);
5851                         goto out;
5852                 }
5853
5854                 extent_slot = path->slots[0];
5855                 leaf = path->nodes[0];
5856                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5857         }
5858 #endif
5859         BUG_ON(item_size < sizeof(*ei));
5860         ei = btrfs_item_ptr(leaf, extent_slot,
5861                             struct btrfs_extent_item);
5862         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
5863             key.type == BTRFS_EXTENT_ITEM_KEY) {
5864                 struct btrfs_tree_block_info *bi;
5865                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5866                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5867                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5868         }
5869
5870         refs = btrfs_extent_refs(leaf, ei);
5871         if (refs < refs_to_drop) {
5872                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
5873                           "for bytenr %Lu\n", refs_to_drop, refs, bytenr);
5874                 ret = -EINVAL;
5875                 btrfs_abort_transaction(trans, extent_root, ret);
5876                 goto out;
5877         }
5878         refs -= refs_to_drop;
5879
5880         if (refs > 0) {
5881                 if (extent_op)
5882                         __run_delayed_extent_op(extent_op, leaf, ei);
5883                 /*
5884                  * In the case of inline back ref, reference count will
5885                  * be updated by remove_extent_backref
5886                  */
5887                 if (iref) {
5888                         BUG_ON(!found_extent);
5889                 } else {
5890                         btrfs_set_extent_refs(leaf, ei, refs);
5891                         btrfs_mark_buffer_dirty(leaf);
5892                 }
5893                 if (found_extent) {
5894                         ret = remove_extent_backref(trans, extent_root, path,
5895                                                     iref, refs_to_drop,
5896                                                     is_data);
5897                         if (ret) {
5898                                 btrfs_abort_transaction(trans, extent_root, ret);
5899                                 goto out;
5900                         }
5901                 }
5902                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
5903                                  root_objectid);
5904         } else {
5905                 if (found_extent) {
5906                         BUG_ON(is_data && refs_to_drop !=
5907                                extent_data_ref_count(root, path, iref));
5908                         if (iref) {
5909                                 BUG_ON(path->slots[0] != extent_slot);
5910                         } else {
5911                                 BUG_ON(path->slots[0] != extent_slot + 1);
5912                                 path->slots[0] = extent_slot;
5913                                 num_to_del = 2;
5914                         }
5915                 }
5916
5917                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5918                                       num_to_del);
5919                 if (ret) {
5920                         btrfs_abort_transaction(trans, extent_root, ret);
5921                         goto out;
5922                 }
5923                 btrfs_release_path(path);
5924
5925                 if (is_data) {
5926                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5927                         if (ret) {
5928                                 btrfs_abort_transaction(trans, extent_root, ret);
5929                                 goto out;
5930                         }
5931                 }
5932
5933                 ret = update_block_group(root, bytenr, num_bytes, 0);
5934                 if (ret) {
5935                         btrfs_abort_transaction(trans, extent_root, ret);
5936                         goto out;
5937                 }
5938         }
5939 out:
5940         btrfs_free_path(path);
5941         return ret;
5942 }
5943
5944 /*
5945  * when we free an block, it is possible (and likely) that we free the last
5946  * delayed ref for that extent as well.  This searches the delayed ref tree for
5947  * a given extent, and if there are no other delayed refs to be processed, it
5948  * removes it from the tree.
5949  */
5950 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5951                                       struct btrfs_root *root, u64 bytenr)
5952 {
5953         struct btrfs_delayed_ref_head *head;
5954         struct btrfs_delayed_ref_root *delayed_refs;
5955         struct btrfs_delayed_ref_node *ref;
5956         struct rb_node *node;
5957         int ret = 0;
5958
5959         delayed_refs = &trans->transaction->delayed_refs;
5960         spin_lock(&delayed_refs->lock);
5961         head = btrfs_find_delayed_ref_head(trans, bytenr);
5962         if (!head)
5963                 goto out;
5964
5965         node = rb_prev(&head->node.rb_node);
5966         if (!node)
5967                 goto out;
5968
5969         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5970
5971         /* there are still entries for this ref, we can't drop it */
5972         if (ref->bytenr == bytenr)
5973                 goto out;
5974
5975         if (head->extent_op) {
5976                 if (!head->must_insert_reserved)
5977                         goto out;
5978                 btrfs_free_delayed_extent_op(head->extent_op);
5979                 head->extent_op = NULL;
5980         }
5981
5982         /*
5983          * waiting for the lock here would deadlock.  If someone else has it
5984          * locked they are already in the process of dropping it anyway
5985          */
5986         if (!mutex_trylock(&head->mutex))
5987                 goto out;
5988
5989         /*
5990          * at this point we have a head with no other entries.  Go
5991          * ahead and process it.
5992          */
5993         head->node.in_tree = 0;
5994         rb_erase(&head->node.rb_node, &delayed_refs->root);
5995         rb_erase(&head->href_node, &delayed_refs->href_root);
5996
5997         delayed_refs->num_entries--;
5998
5999         /*
6000          * we don't take a ref on the node because we're removing it from the
6001          * tree, so we just steal the ref the tree was holding.
6002          */
6003         delayed_refs->num_heads--;
6004         if (list_empty(&head->cluster))
6005                 delayed_refs->num_heads_ready--;
6006
6007         list_del_init(&head->cluster);
6008         spin_unlock(&delayed_refs->lock);
6009
6010         BUG_ON(head->extent_op);
6011         if (head->must_insert_reserved)
6012                 ret = 1;
6013
6014         mutex_unlock(&head->mutex);
6015         btrfs_put_delayed_ref(&head->node);
6016         return ret;
6017 out:
6018         spin_unlock(&delayed_refs->lock);
6019         return 0;
6020 }
6021
6022 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6023                            struct btrfs_root *root,
6024                            struct extent_buffer *buf,
6025                            u64 parent, int last_ref)
6026 {
6027         struct btrfs_block_group_cache *cache = NULL;
6028         int pin = 1;
6029         int ret;
6030
6031         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6032                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6033                                         buf->start, buf->len,
6034                                         parent, root->root_key.objectid,
6035                                         btrfs_header_level(buf),
6036                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6037                 BUG_ON(ret); /* -ENOMEM */
6038         }
6039
6040         if (!last_ref)
6041                 return;
6042
6043         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6044
6045         if (btrfs_header_generation(buf) == trans->transid) {
6046                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6047                         ret = check_ref_cleanup(trans, root, buf->start);
6048                         if (!ret)
6049                                 goto out;
6050                 }
6051
6052                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6053                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6054                         goto out;
6055                 }
6056
6057                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6058
6059                 btrfs_add_free_space(cache, buf->start, buf->len);
6060                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
6061                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6062                 pin = 0;
6063         }
6064 out:
6065         if (pin)
6066                 add_pinned_bytes(root->fs_info, buf->len,
6067                                  btrfs_header_level(buf),
6068                                  root->root_key.objectid);
6069
6070         /*
6071          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6072          * anymore.
6073          */
6074         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6075         btrfs_put_block_group(cache);
6076 }
6077
6078 /* Can return -ENOMEM */
6079 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6080                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6081                       u64 owner, u64 offset, int for_cow)
6082 {
6083         int ret;
6084         struct btrfs_fs_info *fs_info = root->fs_info;
6085
6086         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6087
6088         /*
6089          * tree log blocks never actually go into the extent allocation
6090          * tree, just update pinning info and exit early.
6091          */
6092         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6093                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6094                 /* unlocks the pinned mutex */
6095                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6096                 ret = 0;
6097         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6098                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6099                                         num_bytes,
6100                                         parent, root_objectid, (int)owner,
6101                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
6102         } else {
6103                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6104                                                 num_bytes,
6105                                                 parent, root_objectid, owner,
6106                                                 offset, BTRFS_DROP_DELAYED_REF,
6107                                                 NULL, for_cow);
6108         }
6109         return ret;
6110 }
6111
6112 static u64 stripe_align(struct btrfs_root *root,
6113                         struct btrfs_block_group_cache *cache,
6114                         u64 val, u64 num_bytes)
6115 {
6116         u64 ret = ALIGN(val, root->stripesize);
6117         return ret;
6118 }
6119
6120 /*
6121  * when we wait for progress in the block group caching, its because
6122  * our allocation attempt failed at least once.  So, we must sleep
6123  * and let some progress happen before we try again.
6124  *
6125  * This function will sleep at least once waiting for new free space to
6126  * show up, and then it will check the block group free space numbers
6127  * for our min num_bytes.  Another option is to have it go ahead
6128  * and look in the rbtree for a free extent of a given size, but this
6129  * is a good start.
6130  *
6131  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6132  * any of the information in this block group.
6133  */
6134 static noinline void
6135 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6136                                 u64 num_bytes)
6137 {
6138         struct btrfs_caching_control *caching_ctl;
6139
6140         caching_ctl = get_caching_control(cache);
6141         if (!caching_ctl)
6142                 return;
6143
6144         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6145                    (cache->free_space_ctl->free_space >= num_bytes));
6146
6147         put_caching_control(caching_ctl);
6148 }
6149
6150 static noinline int
6151 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6152 {
6153         struct btrfs_caching_control *caching_ctl;
6154         int ret = 0;
6155
6156         caching_ctl = get_caching_control(cache);
6157         if (!caching_ctl)
6158                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6159
6160         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6161         if (cache->cached == BTRFS_CACHE_ERROR)
6162                 ret = -EIO;
6163         put_caching_control(caching_ctl);
6164         return ret;
6165 }
6166
6167 int __get_raid_index(u64 flags)
6168 {
6169         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6170                 return BTRFS_RAID_RAID10;
6171         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6172                 return BTRFS_RAID_RAID1;
6173         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6174                 return BTRFS_RAID_DUP;
6175         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6176                 return BTRFS_RAID_RAID0;
6177         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6178                 return BTRFS_RAID_RAID5;
6179         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6180                 return BTRFS_RAID_RAID6;
6181
6182         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6183 }
6184
6185 int get_block_group_index(struct btrfs_block_group_cache *cache)
6186 {
6187         return __get_raid_index(cache->flags);
6188 }
6189
6190 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6191         [BTRFS_RAID_RAID10]     = "raid10",
6192         [BTRFS_RAID_RAID1]      = "raid1",
6193         [BTRFS_RAID_DUP]        = "dup",
6194         [BTRFS_RAID_RAID0]      = "raid0",
6195         [BTRFS_RAID_SINGLE]     = "single",
6196         [BTRFS_RAID_RAID5]      = "raid5",
6197         [BTRFS_RAID_RAID6]      = "raid6",
6198 };
6199
6200 static const char *get_raid_name(enum btrfs_raid_types type)
6201 {
6202         if (type >= BTRFS_NR_RAID_TYPES)
6203                 return NULL;
6204
6205         return btrfs_raid_type_names[type];
6206 }
6207
6208 enum btrfs_loop_type {
6209         LOOP_CACHING_NOWAIT = 0,
6210         LOOP_CACHING_WAIT = 1,
6211         LOOP_ALLOC_CHUNK = 2,
6212         LOOP_NO_EMPTY_SIZE = 3,
6213 };
6214
6215 /*
6216  * walks the btree of allocated extents and find a hole of a given size.
6217  * The key ins is changed to record the hole:
6218  * ins->objectid == start position
6219  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6220  * ins->offset == the size of the hole.
6221  * Any available blocks before search_start are skipped.
6222  *
6223  * If there is no suitable free space, we will record the max size of
6224  * the free space extent currently.
6225  */
6226 static noinline int find_free_extent(struct btrfs_root *orig_root,
6227                                      u64 num_bytes, u64 empty_size,
6228                                      u64 hint_byte, struct btrfs_key *ins,
6229                                      u64 flags)
6230 {
6231         int ret = 0;
6232         struct btrfs_root *root = orig_root->fs_info->extent_root;
6233         struct btrfs_free_cluster *last_ptr = NULL;
6234         struct btrfs_block_group_cache *block_group = NULL;
6235         struct btrfs_block_group_cache *used_block_group;
6236         u64 search_start = 0;
6237         u64 max_extent_size = 0;
6238         int empty_cluster = 2 * 1024 * 1024;
6239         struct btrfs_space_info *space_info;
6240         int loop = 0;
6241         int index = __get_raid_index(flags);
6242         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6243                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6244         bool failed_cluster_refill = false;
6245         bool failed_alloc = false;
6246         bool use_cluster = true;
6247         bool have_caching_bg = false;
6248
6249         WARN_ON(num_bytes < root->sectorsize);
6250         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
6251         ins->objectid = 0;
6252         ins->offset = 0;
6253
6254         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6255
6256         space_info = __find_space_info(root->fs_info, flags);
6257         if (!space_info) {
6258                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6259                 return -ENOSPC;
6260         }
6261
6262         /*
6263          * If the space info is for both data and metadata it means we have a
6264          * small filesystem and we can't use the clustering stuff.
6265          */
6266         if (btrfs_mixed_space_info(space_info))
6267                 use_cluster = false;
6268
6269         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6270                 last_ptr = &root->fs_info->meta_alloc_cluster;
6271                 if (!btrfs_test_opt(root, SSD))
6272                         empty_cluster = 64 * 1024;
6273         }
6274
6275         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6276             btrfs_test_opt(root, SSD)) {
6277                 last_ptr = &root->fs_info->data_alloc_cluster;
6278         }
6279
6280         if (last_ptr) {
6281                 spin_lock(&last_ptr->lock);
6282                 if (last_ptr->block_group)
6283                         hint_byte = last_ptr->window_start;
6284                 spin_unlock(&last_ptr->lock);
6285         }
6286
6287         search_start = max(search_start, first_logical_byte(root, 0));
6288         search_start = max(search_start, hint_byte);
6289
6290         if (!last_ptr)
6291                 empty_cluster = 0;
6292
6293         if (search_start == hint_byte) {
6294                 block_group = btrfs_lookup_block_group(root->fs_info,
6295                                                        search_start);
6296                 used_block_group = block_group;
6297                 /*
6298                  * we don't want to use the block group if it doesn't match our
6299                  * allocation bits, or if its not cached.
6300                  *
6301                  * However if we are re-searching with an ideal block group
6302                  * picked out then we don't care that the block group is cached.
6303                  */
6304                 if (block_group && block_group_bits(block_group, flags) &&
6305                     block_group->cached != BTRFS_CACHE_NO) {
6306                         down_read(&space_info->groups_sem);
6307                         if (list_empty(&block_group->list) ||
6308                             block_group->ro) {
6309                                 /*
6310                                  * someone is removing this block group,
6311                                  * we can't jump into the have_block_group
6312                                  * target because our list pointers are not
6313                                  * valid
6314                                  */
6315                                 btrfs_put_block_group(block_group);
6316                                 up_read(&space_info->groups_sem);
6317                         } else {
6318                                 index = get_block_group_index(block_group);
6319                                 goto have_block_group;
6320                         }
6321                 } else if (block_group) {
6322                         btrfs_put_block_group(block_group);
6323                 }
6324         }
6325 search:
6326         have_caching_bg = false;
6327         down_read(&space_info->groups_sem);
6328         list_for_each_entry(block_group, &space_info->block_groups[index],
6329                             list) {
6330                 u64 offset;
6331                 int cached;
6332
6333                 used_block_group = block_group;
6334                 btrfs_get_block_group(block_group);
6335                 search_start = block_group->key.objectid;
6336
6337                 /*
6338                  * this can happen if we end up cycling through all the
6339                  * raid types, but we want to make sure we only allocate
6340                  * for the proper type.
6341                  */
6342                 if (!block_group_bits(block_group, flags)) {
6343                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6344                                 BTRFS_BLOCK_GROUP_RAID1 |
6345                                 BTRFS_BLOCK_GROUP_RAID5 |
6346                                 BTRFS_BLOCK_GROUP_RAID6 |
6347                                 BTRFS_BLOCK_GROUP_RAID10;
6348
6349                         /*
6350                          * if they asked for extra copies and this block group
6351                          * doesn't provide them, bail.  This does allow us to
6352                          * fill raid0 from raid1.
6353                          */
6354                         if ((flags & extra) && !(block_group->flags & extra))
6355                                 goto loop;
6356                 }
6357
6358 have_block_group:
6359                 cached = block_group_cache_done(block_group);
6360                 if (unlikely(!cached)) {
6361                         ret = cache_block_group(block_group, 0);
6362                         BUG_ON(ret < 0);
6363                         ret = 0;
6364                 }
6365
6366                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6367                         goto loop;
6368                 if (unlikely(block_group->ro))
6369                         goto loop;
6370
6371                 /*
6372                  * Ok we want to try and use the cluster allocator, so
6373                  * lets look there
6374                  */
6375                 if (last_ptr) {
6376                         unsigned long aligned_cluster;
6377                         /*
6378                          * the refill lock keeps out other
6379                          * people trying to start a new cluster
6380                          */
6381                         spin_lock(&last_ptr->refill_lock);
6382                         used_block_group = last_ptr->block_group;
6383                         if (used_block_group != block_group &&
6384                             (!used_block_group ||
6385                              used_block_group->ro ||
6386                              !block_group_bits(used_block_group, flags))) {
6387                                 used_block_group = block_group;
6388                                 goto refill_cluster;
6389                         }
6390
6391                         if (used_block_group != block_group)
6392                                 btrfs_get_block_group(used_block_group);
6393
6394                         offset = btrfs_alloc_from_cluster(used_block_group,
6395                                                 last_ptr,
6396                                                 num_bytes,
6397                                                 used_block_group->key.objectid,
6398                                                 &max_extent_size);
6399                         if (offset) {
6400                                 /* we have a block, we're done */
6401                                 spin_unlock(&last_ptr->refill_lock);
6402                                 trace_btrfs_reserve_extent_cluster(root,
6403                                         block_group, search_start, num_bytes);
6404                                 goto checks;
6405                         }
6406
6407                         WARN_ON(last_ptr->block_group != used_block_group);
6408                         if (used_block_group != block_group) {
6409                                 btrfs_put_block_group(used_block_group);
6410                                 used_block_group = block_group;
6411                         }
6412 refill_cluster:
6413                         BUG_ON(used_block_group != block_group);
6414                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6415                          * set up a new clusters, so lets just skip it
6416                          * and let the allocator find whatever block
6417                          * it can find.  If we reach this point, we
6418                          * will have tried the cluster allocator
6419                          * plenty of times and not have found
6420                          * anything, so we are likely way too
6421                          * fragmented for the clustering stuff to find
6422                          * anything.
6423                          *
6424                          * However, if the cluster is taken from the
6425                          * current block group, release the cluster
6426                          * first, so that we stand a better chance of
6427                          * succeeding in the unclustered
6428                          * allocation.  */
6429                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6430                             last_ptr->block_group != block_group) {
6431                                 spin_unlock(&last_ptr->refill_lock);
6432                                 goto unclustered_alloc;
6433                         }
6434
6435                         /*
6436                          * this cluster didn't work out, free it and
6437                          * start over
6438                          */
6439                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6440
6441                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6442                                 spin_unlock(&last_ptr->refill_lock);
6443                                 goto unclustered_alloc;
6444                         }
6445
6446                         aligned_cluster = max_t(unsigned long,
6447                                                 empty_cluster + empty_size,
6448                                               block_group->full_stripe_len);
6449
6450                         /* allocate a cluster in this block group */
6451                         ret = btrfs_find_space_cluster(root, block_group,
6452                                                        last_ptr, search_start,
6453                                                        num_bytes,
6454                                                        aligned_cluster);
6455                         if (ret == 0) {
6456                                 /*
6457                                  * now pull our allocation out of this
6458                                  * cluster
6459                                  */
6460                                 offset = btrfs_alloc_from_cluster(block_group,
6461                                                         last_ptr,
6462                                                         num_bytes,
6463                                                         search_start,
6464                                                         &max_extent_size);
6465                                 if (offset) {
6466                                         /* we found one, proceed */
6467                                         spin_unlock(&last_ptr->refill_lock);
6468                                         trace_btrfs_reserve_extent_cluster(root,
6469                                                 block_group, search_start,
6470                                                 num_bytes);
6471                                         goto checks;
6472                                 }
6473                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6474                                    && !failed_cluster_refill) {
6475                                 spin_unlock(&last_ptr->refill_lock);
6476
6477                                 failed_cluster_refill = true;
6478                                 wait_block_group_cache_progress(block_group,
6479                                        num_bytes + empty_cluster + empty_size);
6480                                 goto have_block_group;
6481                         }
6482
6483                         /*
6484                          * at this point we either didn't find a cluster
6485                          * or we weren't able to allocate a block from our
6486                          * cluster.  Free the cluster we've been trying
6487                          * to use, and go to the next block group
6488                          */
6489                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6490                         spin_unlock(&last_ptr->refill_lock);
6491                         goto loop;
6492                 }
6493
6494 unclustered_alloc:
6495                 spin_lock(&block_group->free_space_ctl->tree_lock);
6496                 if (cached &&
6497                     block_group->free_space_ctl->free_space <
6498                     num_bytes + empty_cluster + empty_size) {
6499                         if (block_group->free_space_ctl->free_space >
6500                             max_extent_size)
6501                                 max_extent_size =
6502                                         block_group->free_space_ctl->free_space;
6503                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6504                         goto loop;
6505                 }
6506                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6507
6508                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6509                                                     num_bytes, empty_size,
6510                                                     &max_extent_size);
6511                 /*
6512                  * If we didn't find a chunk, and we haven't failed on this
6513                  * block group before, and this block group is in the middle of
6514                  * caching and we are ok with waiting, then go ahead and wait
6515                  * for progress to be made, and set failed_alloc to true.
6516                  *
6517                  * If failed_alloc is true then we've already waited on this
6518                  * block group once and should move on to the next block group.
6519                  */
6520                 if (!offset && !failed_alloc && !cached &&
6521                     loop > LOOP_CACHING_NOWAIT) {
6522                         wait_block_group_cache_progress(block_group,
6523                                                 num_bytes + empty_size);
6524                         failed_alloc = true;
6525                         goto have_block_group;
6526                 } else if (!offset) {
6527                         if (!cached)
6528                                 have_caching_bg = true;
6529                         goto loop;
6530                 }
6531 checks:
6532                 search_start = stripe_align(root, used_block_group,
6533                                             offset, num_bytes);
6534
6535                 /* move on to the next group */
6536                 if (search_start + num_bytes >
6537                     used_block_group->key.objectid + used_block_group->key.offset) {
6538                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6539                         goto loop;
6540                 }
6541
6542                 if (offset < search_start)
6543                         btrfs_add_free_space(used_block_group, offset,
6544                                              search_start - offset);
6545                 BUG_ON(offset > search_start);
6546
6547                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
6548                                                   alloc_type);
6549                 if (ret == -EAGAIN) {
6550                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6551                         goto loop;
6552                 }
6553
6554                 /* we are all good, lets return */
6555                 ins->objectid = search_start;
6556                 ins->offset = num_bytes;
6557
6558                 trace_btrfs_reserve_extent(orig_root, block_group,
6559                                            search_start, num_bytes);
6560                 if (used_block_group != block_group)
6561                         btrfs_put_block_group(used_block_group);
6562                 btrfs_put_block_group(block_group);
6563                 break;
6564 loop:
6565                 failed_cluster_refill = false;
6566                 failed_alloc = false;
6567                 BUG_ON(index != get_block_group_index(block_group));
6568                 if (used_block_group != block_group)
6569                         btrfs_put_block_group(used_block_group);
6570                 btrfs_put_block_group(block_group);
6571         }
6572         up_read(&space_info->groups_sem);
6573
6574         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6575                 goto search;
6576
6577         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6578                 goto search;
6579
6580         /*
6581          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6582          *                      caching kthreads as we move along
6583          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6584          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6585          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6586          *                      again
6587          */
6588         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6589                 index = 0;
6590                 loop++;
6591                 if (loop == LOOP_ALLOC_CHUNK) {
6592                         struct btrfs_trans_handle *trans;
6593
6594                         trans = btrfs_join_transaction(root);
6595                         if (IS_ERR(trans)) {
6596                                 ret = PTR_ERR(trans);
6597                                 goto out;
6598                         }
6599
6600                         ret = do_chunk_alloc(trans, root, flags,
6601                                              CHUNK_ALLOC_FORCE);
6602                         /*
6603                          * Do not bail out on ENOSPC since we
6604                          * can do more things.
6605                          */
6606                         if (ret < 0 && ret != -ENOSPC)
6607                                 btrfs_abort_transaction(trans,
6608                                                         root, ret);
6609                         else
6610                                 ret = 0;
6611                         btrfs_end_transaction(trans, root);
6612                         if (ret)
6613                                 goto out;
6614                 }
6615
6616                 if (loop == LOOP_NO_EMPTY_SIZE) {
6617                         empty_size = 0;
6618                         empty_cluster = 0;
6619                 }
6620
6621                 goto search;
6622         } else if (!ins->objectid) {
6623                 ret = -ENOSPC;
6624         } else if (ins->objectid) {
6625                 ret = 0;
6626         }
6627 out:
6628         if (ret == -ENOSPC)
6629                 ins->offset = max_extent_size;
6630         return ret;
6631 }
6632
6633 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6634                             int dump_block_groups)
6635 {
6636         struct btrfs_block_group_cache *cache;
6637         int index = 0;
6638
6639         spin_lock(&info->lock);
6640         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
6641                info->flags,
6642                info->total_bytes - info->bytes_used - info->bytes_pinned -
6643                info->bytes_reserved - info->bytes_readonly,
6644                (info->full) ? "" : "not ");
6645         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
6646                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6647                info->total_bytes, info->bytes_used, info->bytes_pinned,
6648                info->bytes_reserved, info->bytes_may_use,
6649                info->bytes_readonly);
6650         spin_unlock(&info->lock);
6651
6652         if (!dump_block_groups)
6653                 return;
6654
6655         down_read(&info->groups_sem);
6656 again:
6657         list_for_each_entry(cache, &info->block_groups[index], list) {
6658                 spin_lock(&cache->lock);
6659                 printk(KERN_INFO "BTRFS: "
6660                            "block group %llu has %llu bytes, "
6661                            "%llu used %llu pinned %llu reserved %s\n",
6662                        cache->key.objectid, cache->key.offset,
6663                        btrfs_block_group_used(&cache->item), cache->pinned,
6664                        cache->reserved, cache->ro ? "[readonly]" : "");
6665                 btrfs_dump_free_space(cache, bytes);
6666                 spin_unlock(&cache->lock);
6667         }
6668         if (++index < BTRFS_NR_RAID_TYPES)
6669                 goto again;
6670         up_read(&info->groups_sem);
6671 }
6672
6673 int btrfs_reserve_extent(struct btrfs_root *root,
6674                          u64 num_bytes, u64 min_alloc_size,
6675                          u64 empty_size, u64 hint_byte,
6676                          struct btrfs_key *ins, int is_data)
6677 {
6678         bool final_tried = false;
6679         u64 flags;
6680         int ret;
6681
6682         flags = btrfs_get_alloc_profile(root, is_data);
6683 again:
6684         WARN_ON(num_bytes < root->sectorsize);
6685         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
6686                                flags);
6687
6688         if (ret == -ENOSPC) {
6689                 if (!final_tried && ins->offset) {
6690                         num_bytes = min(num_bytes >> 1, ins->offset);
6691                         num_bytes = round_down(num_bytes, root->sectorsize);
6692                         num_bytes = max(num_bytes, min_alloc_size);
6693                         if (num_bytes == min_alloc_size)
6694                                 final_tried = true;
6695                         goto again;
6696                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6697                         struct btrfs_space_info *sinfo;
6698
6699                         sinfo = __find_space_info(root->fs_info, flags);
6700                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6701                                 flags, num_bytes);
6702                         if (sinfo)
6703                                 dump_space_info(sinfo, num_bytes, 1);
6704                 }
6705         }
6706
6707         return ret;
6708 }
6709
6710 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6711                                         u64 start, u64 len, int pin)
6712 {
6713         struct btrfs_block_group_cache *cache;
6714         int ret = 0;
6715
6716         cache = btrfs_lookup_block_group(root->fs_info, start);
6717         if (!cache) {
6718                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6719                         start);
6720                 return -ENOSPC;
6721         }
6722
6723         if (btrfs_test_opt(root, DISCARD))
6724                 ret = btrfs_discard_extent(root, start, len, NULL);
6725
6726         if (pin)
6727                 pin_down_extent(root, cache, start, len, 1);
6728         else {
6729                 btrfs_add_free_space(cache, start, len);
6730                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6731         }
6732         btrfs_put_block_group(cache);
6733
6734         trace_btrfs_reserved_extent_free(root, start, len);
6735
6736         return ret;
6737 }
6738
6739 int btrfs_free_reserved_extent(struct btrfs_root *root,
6740                                         u64 start, u64 len)
6741 {
6742         return __btrfs_free_reserved_extent(root, start, len, 0);
6743 }
6744
6745 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6746                                        u64 start, u64 len)
6747 {
6748         return __btrfs_free_reserved_extent(root, start, len, 1);
6749 }
6750
6751 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6752                                       struct btrfs_root *root,
6753                                       u64 parent, u64 root_objectid,
6754                                       u64 flags, u64 owner, u64 offset,
6755                                       struct btrfs_key *ins, int ref_mod)
6756 {
6757         int ret;
6758         struct btrfs_fs_info *fs_info = root->fs_info;
6759         struct btrfs_extent_item *extent_item;
6760         struct btrfs_extent_inline_ref *iref;
6761         struct btrfs_path *path;
6762         struct extent_buffer *leaf;
6763         int type;
6764         u32 size;
6765
6766         if (parent > 0)
6767                 type = BTRFS_SHARED_DATA_REF_KEY;
6768         else
6769                 type = BTRFS_EXTENT_DATA_REF_KEY;
6770
6771         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6772
6773         path = btrfs_alloc_path();
6774         if (!path)
6775                 return -ENOMEM;
6776
6777         path->leave_spinning = 1;
6778         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6779                                       ins, size);
6780         if (ret) {
6781                 btrfs_free_path(path);
6782                 return ret;
6783         }
6784
6785         leaf = path->nodes[0];
6786         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6787                                      struct btrfs_extent_item);
6788         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6789         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6790         btrfs_set_extent_flags(leaf, extent_item,
6791                                flags | BTRFS_EXTENT_FLAG_DATA);
6792
6793         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6794         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6795         if (parent > 0) {
6796                 struct btrfs_shared_data_ref *ref;
6797                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6798                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6799                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6800         } else {
6801                 struct btrfs_extent_data_ref *ref;
6802                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6803                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6804                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6805                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6806                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6807         }
6808
6809         btrfs_mark_buffer_dirty(path->nodes[0]);
6810         btrfs_free_path(path);
6811
6812         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6813         if (ret) { /* -ENOENT, logic error */
6814                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6815                         ins->objectid, ins->offset);
6816                 BUG();
6817         }
6818         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6819         return ret;
6820 }
6821
6822 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6823                                      struct btrfs_root *root,
6824                                      u64 parent, u64 root_objectid,
6825                                      u64 flags, struct btrfs_disk_key *key,
6826                                      int level, struct btrfs_key *ins)
6827 {
6828         int ret;
6829         struct btrfs_fs_info *fs_info = root->fs_info;
6830         struct btrfs_extent_item *extent_item;
6831         struct btrfs_tree_block_info *block_info;
6832         struct btrfs_extent_inline_ref *iref;
6833         struct btrfs_path *path;
6834         struct extent_buffer *leaf;
6835         u32 size = sizeof(*extent_item) + sizeof(*iref);
6836         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6837                                                  SKINNY_METADATA);
6838
6839         if (!skinny_metadata)
6840                 size += sizeof(*block_info);
6841
6842         path = btrfs_alloc_path();
6843         if (!path) {
6844                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
6845                                                    root->leafsize);
6846                 return -ENOMEM;
6847         }
6848
6849         path->leave_spinning = 1;
6850         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6851                                       ins, size);
6852         if (ret) {
6853                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
6854                                                    root->leafsize);
6855                 btrfs_free_path(path);
6856                 return ret;
6857         }
6858
6859         leaf = path->nodes[0];
6860         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6861                                      struct btrfs_extent_item);
6862         btrfs_set_extent_refs(leaf, extent_item, 1);
6863         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6864         btrfs_set_extent_flags(leaf, extent_item,
6865                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6866
6867         if (skinny_metadata) {
6868                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6869         } else {
6870                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6871                 btrfs_set_tree_block_key(leaf, block_info, key);
6872                 btrfs_set_tree_block_level(leaf, block_info, level);
6873                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6874         }
6875
6876         if (parent > 0) {
6877                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6878                 btrfs_set_extent_inline_ref_type(leaf, iref,
6879                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6880                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6881         } else {
6882                 btrfs_set_extent_inline_ref_type(leaf, iref,
6883                                                  BTRFS_TREE_BLOCK_REF_KEY);
6884                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6885         }
6886
6887         btrfs_mark_buffer_dirty(leaf);
6888         btrfs_free_path(path);
6889
6890         ret = update_block_group(root, ins->objectid, root->leafsize, 1);
6891         if (ret) { /* -ENOENT, logic error */
6892                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6893                         ins->objectid, ins->offset);
6894                 BUG();
6895         }
6896
6897         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->leafsize);
6898         return ret;
6899 }
6900
6901 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6902                                      struct btrfs_root *root,
6903                                      u64 root_objectid, u64 owner,
6904                                      u64 offset, struct btrfs_key *ins)
6905 {
6906         int ret;
6907
6908         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6909
6910         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6911                                          ins->offset, 0,
6912                                          root_objectid, owner, offset,
6913                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6914         return ret;
6915 }
6916
6917 /*
6918  * this is used by the tree logging recovery code.  It records that
6919  * an extent has been allocated and makes sure to clear the free
6920  * space cache bits as well
6921  */
6922 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6923                                    struct btrfs_root *root,
6924                                    u64 root_objectid, u64 owner, u64 offset,
6925                                    struct btrfs_key *ins)
6926 {
6927         int ret;
6928         struct btrfs_block_group_cache *block_group;
6929
6930         /*
6931          * Mixed block groups will exclude before processing the log so we only
6932          * need to do the exlude dance if this fs isn't mixed.
6933          */
6934         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
6935                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
6936                 if (ret)
6937                         return ret;
6938         }
6939
6940         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6941         if (!block_group)
6942                 return -EINVAL;
6943
6944         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6945                                           RESERVE_ALLOC_NO_ACCOUNT);
6946         BUG_ON(ret); /* logic error */
6947         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6948                                          0, owner, offset, ins, 1);
6949         btrfs_put_block_group(block_group);
6950         return ret;
6951 }
6952
6953 static struct extent_buffer *
6954 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6955                       u64 bytenr, u32 blocksize, int level)
6956 {
6957         struct extent_buffer *buf;
6958
6959         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6960         if (!buf)
6961                 return ERR_PTR(-ENOMEM);
6962         btrfs_set_header_generation(buf, trans->transid);
6963         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6964         btrfs_tree_lock(buf);
6965         clean_tree_block(trans, root, buf);
6966         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6967
6968         btrfs_set_lock_blocking(buf);
6969         btrfs_set_buffer_uptodate(buf);
6970
6971         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6972                 /*
6973                  * we allow two log transactions at a time, use different
6974                  * EXENT bit to differentiate dirty pages.
6975                  */
6976                 if (root->log_transid % 2 == 0)
6977                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6978                                         buf->start + buf->len - 1, GFP_NOFS);
6979                 else
6980                         set_extent_new(&root->dirty_log_pages, buf->start,
6981                                         buf->start + buf->len - 1, GFP_NOFS);
6982         } else {
6983                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6984                          buf->start + buf->len - 1, GFP_NOFS);
6985         }
6986         trans->blocks_used++;
6987         /* this returns a buffer locked for blocking */
6988         return buf;
6989 }
6990
6991 static struct btrfs_block_rsv *
6992 use_block_rsv(struct btrfs_trans_handle *trans,
6993               struct btrfs_root *root, u32 blocksize)
6994 {
6995         struct btrfs_block_rsv *block_rsv;
6996         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6997         int ret;
6998         bool global_updated = false;
6999
7000         block_rsv = get_block_rsv(trans, root);
7001
7002         if (unlikely(block_rsv->size == 0))
7003                 goto try_reserve;
7004 again:
7005         ret = block_rsv_use_bytes(block_rsv, blocksize);
7006         if (!ret)
7007                 return block_rsv;
7008
7009         if (block_rsv->failfast)
7010                 return ERR_PTR(ret);
7011
7012         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7013                 global_updated = true;
7014                 update_global_block_rsv(root->fs_info);
7015                 goto again;
7016         }
7017
7018         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7019                 static DEFINE_RATELIMIT_STATE(_rs,
7020                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7021                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7022                 if (__ratelimit(&_rs))
7023                         WARN(1, KERN_DEBUG
7024                                 "BTRFS: block rsv returned %d\n", ret);
7025         }
7026 try_reserve:
7027         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7028                                      BTRFS_RESERVE_NO_FLUSH);
7029         if (!ret)
7030                 return block_rsv;
7031         /*
7032          * If we couldn't reserve metadata bytes try and use some from
7033          * the global reserve if its space type is the same as the global
7034          * reservation.
7035          */
7036         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7037             block_rsv->space_info == global_rsv->space_info) {
7038                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7039                 if (!ret)
7040                         return global_rsv;
7041         }
7042         return ERR_PTR(ret);
7043 }
7044
7045 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7046                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7047 {
7048         block_rsv_add_bytes(block_rsv, blocksize, 0);
7049         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7050 }
7051
7052 /*
7053  * finds a free extent and does all the dirty work required for allocation
7054  * returns the key for the extent through ins, and a tree buffer for
7055  * the first block of the extent through buf.
7056  *
7057  * returns the tree buffer or NULL.
7058  */
7059 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
7060                                         struct btrfs_root *root, u32 blocksize,
7061                                         u64 parent, u64 root_objectid,
7062                                         struct btrfs_disk_key *key, int level,
7063                                         u64 hint, u64 empty_size)
7064 {
7065         struct btrfs_key ins;
7066         struct btrfs_block_rsv *block_rsv;
7067         struct extent_buffer *buf;
7068         u64 flags = 0;
7069         int ret;
7070         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7071                                                  SKINNY_METADATA);
7072
7073         block_rsv = use_block_rsv(trans, root, blocksize);
7074         if (IS_ERR(block_rsv))
7075                 return ERR_CAST(block_rsv);
7076
7077         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7078                                    empty_size, hint, &ins, 0);
7079         if (ret) {
7080                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7081                 return ERR_PTR(ret);
7082         }
7083
7084         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
7085                                     blocksize, level);
7086         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
7087
7088         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7089                 if (parent == 0)
7090                         parent = ins.objectid;
7091                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7092         } else
7093                 BUG_ON(parent > 0);
7094
7095         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7096                 struct btrfs_delayed_extent_op *extent_op;
7097                 extent_op = btrfs_alloc_delayed_extent_op();
7098                 BUG_ON(!extent_op); /* -ENOMEM */
7099                 if (key)
7100                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7101                 else
7102                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7103                 extent_op->flags_to_set = flags;
7104                 if (skinny_metadata)
7105                         extent_op->update_key = 0;
7106                 else
7107                         extent_op->update_key = 1;
7108                 extent_op->update_flags = 1;
7109                 extent_op->is_data = 0;
7110                 extent_op->level = level;
7111
7112                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7113                                         ins.objectid,
7114                                         ins.offset, parent, root_objectid,
7115                                         level, BTRFS_ADD_DELAYED_EXTENT,
7116                                         extent_op, 0);
7117                 BUG_ON(ret); /* -ENOMEM */
7118         }
7119         return buf;
7120 }
7121
7122 struct walk_control {
7123         u64 refs[BTRFS_MAX_LEVEL];
7124         u64 flags[BTRFS_MAX_LEVEL];
7125         struct btrfs_key update_progress;
7126         int stage;
7127         int level;
7128         int shared_level;
7129         int update_ref;
7130         int keep_locks;
7131         int reada_slot;
7132         int reada_count;
7133         int for_reloc;
7134 };
7135
7136 #define DROP_REFERENCE  1
7137 #define UPDATE_BACKREF  2
7138
7139 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7140                                      struct btrfs_root *root,
7141                                      struct walk_control *wc,
7142                                      struct btrfs_path *path)
7143 {
7144         u64 bytenr;
7145         u64 generation;
7146         u64 refs;
7147         u64 flags;
7148         u32 nritems;
7149         u32 blocksize;
7150         struct btrfs_key key;
7151         struct extent_buffer *eb;
7152         int ret;
7153         int slot;
7154         int nread = 0;
7155
7156         if (path->slots[wc->level] < wc->reada_slot) {
7157                 wc->reada_count = wc->reada_count * 2 / 3;
7158                 wc->reada_count = max(wc->reada_count, 2);
7159         } else {
7160                 wc->reada_count = wc->reada_count * 3 / 2;
7161                 wc->reada_count = min_t(int, wc->reada_count,
7162                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7163         }
7164
7165         eb = path->nodes[wc->level];
7166         nritems = btrfs_header_nritems(eb);
7167         blocksize = btrfs_level_size(root, wc->level - 1);
7168
7169         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7170                 if (nread >= wc->reada_count)
7171                         break;
7172
7173                 cond_resched();
7174                 bytenr = btrfs_node_blockptr(eb, slot);
7175                 generation = btrfs_node_ptr_generation(eb, slot);
7176
7177                 if (slot == path->slots[wc->level])
7178                         goto reada;
7179
7180                 if (wc->stage == UPDATE_BACKREF &&
7181                     generation <= root->root_key.offset)
7182                         continue;
7183
7184                 /* We don't lock the tree block, it's OK to be racy here */
7185                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7186                                                wc->level - 1, 1, &refs,
7187                                                &flags);
7188                 /* We don't care about errors in readahead. */
7189                 if (ret < 0)
7190                         continue;
7191                 BUG_ON(refs == 0);
7192
7193                 if (wc->stage == DROP_REFERENCE) {
7194                         if (refs == 1)
7195                                 goto reada;
7196
7197                         if (wc->level == 1 &&
7198                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7199                                 continue;
7200                         if (!wc->update_ref ||
7201                             generation <= root->root_key.offset)
7202                                 continue;
7203                         btrfs_node_key_to_cpu(eb, &key, slot);
7204                         ret = btrfs_comp_cpu_keys(&key,
7205                                                   &wc->update_progress);
7206                         if (ret < 0)
7207                                 continue;
7208                 } else {
7209                         if (wc->level == 1 &&
7210                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7211                                 continue;
7212                 }
7213 reada:
7214                 ret = readahead_tree_block(root, bytenr, blocksize,
7215                                            generation);
7216                 if (ret)
7217                         break;
7218                 nread++;
7219         }
7220         wc->reada_slot = slot;
7221 }
7222
7223 /*
7224  * helper to process tree block while walking down the tree.
7225  *
7226  * when wc->stage == UPDATE_BACKREF, this function updates
7227  * back refs for pointers in the block.
7228  *
7229  * NOTE: return value 1 means we should stop walking down.
7230  */
7231 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7232                                    struct btrfs_root *root,
7233                                    struct btrfs_path *path,
7234                                    struct walk_control *wc, int lookup_info)
7235 {
7236         int level = wc->level;
7237         struct extent_buffer *eb = path->nodes[level];
7238         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7239         int ret;
7240
7241         if (wc->stage == UPDATE_BACKREF &&
7242             btrfs_header_owner(eb) != root->root_key.objectid)
7243                 return 1;
7244
7245         /*
7246          * when reference count of tree block is 1, it won't increase
7247          * again. once full backref flag is set, we never clear it.
7248          */
7249         if (lookup_info &&
7250             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7251              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7252                 BUG_ON(!path->locks[level]);
7253                 ret = btrfs_lookup_extent_info(trans, root,
7254                                                eb->start, level, 1,
7255                                                &wc->refs[level],
7256                                                &wc->flags[level]);
7257                 BUG_ON(ret == -ENOMEM);
7258                 if (ret)
7259                         return ret;
7260                 BUG_ON(wc->refs[level] == 0);
7261         }
7262
7263         if (wc->stage == DROP_REFERENCE) {
7264                 if (wc->refs[level] > 1)
7265                         return 1;
7266
7267                 if (path->locks[level] && !wc->keep_locks) {
7268                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7269                         path->locks[level] = 0;
7270                 }
7271                 return 0;
7272         }
7273
7274         /* wc->stage == UPDATE_BACKREF */
7275         if (!(wc->flags[level] & flag)) {
7276                 BUG_ON(!path->locks[level]);
7277                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
7278                 BUG_ON(ret); /* -ENOMEM */
7279                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
7280                 BUG_ON(ret); /* -ENOMEM */
7281                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7282                                                   eb->len, flag,
7283                                                   btrfs_header_level(eb), 0);
7284                 BUG_ON(ret); /* -ENOMEM */
7285                 wc->flags[level] |= flag;
7286         }
7287
7288         /*
7289          * the block is shared by multiple trees, so it's not good to
7290          * keep the tree lock
7291          */
7292         if (path->locks[level] && level > 0) {
7293                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7294                 path->locks[level] = 0;
7295         }
7296         return 0;
7297 }
7298
7299 /*
7300  * helper to process tree block pointer.
7301  *
7302  * when wc->stage == DROP_REFERENCE, this function checks
7303  * reference count of the block pointed to. if the block
7304  * is shared and we need update back refs for the subtree
7305  * rooted at the block, this function changes wc->stage to
7306  * UPDATE_BACKREF. if the block is shared and there is no
7307  * need to update back, this function drops the reference
7308  * to the block.
7309  *
7310  * NOTE: return value 1 means we should stop walking down.
7311  */
7312 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7313                                  struct btrfs_root *root,
7314                                  struct btrfs_path *path,
7315                                  struct walk_control *wc, int *lookup_info)
7316 {
7317         u64 bytenr;
7318         u64 generation;
7319         u64 parent;
7320         u32 blocksize;
7321         struct btrfs_key key;
7322         struct extent_buffer *next;
7323         int level = wc->level;
7324         int reada = 0;
7325         int ret = 0;
7326
7327         generation = btrfs_node_ptr_generation(path->nodes[level],
7328                                                path->slots[level]);
7329         /*
7330          * if the lower level block was created before the snapshot
7331          * was created, we know there is no need to update back refs
7332          * for the subtree
7333          */
7334         if (wc->stage == UPDATE_BACKREF &&
7335             generation <= root->root_key.offset) {
7336                 *lookup_info = 1;
7337                 return 1;
7338         }
7339
7340         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7341         blocksize = btrfs_level_size(root, level - 1);
7342
7343         next = btrfs_find_tree_block(root, bytenr, blocksize);
7344         if (!next) {
7345                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7346                 if (!next)
7347                         return -ENOMEM;
7348                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
7349                                                level - 1);
7350                 reada = 1;
7351         }
7352         btrfs_tree_lock(next);
7353         btrfs_set_lock_blocking(next);
7354
7355         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7356                                        &wc->refs[level - 1],
7357                                        &wc->flags[level - 1]);
7358         if (ret < 0) {
7359                 btrfs_tree_unlock(next);
7360                 return ret;
7361         }
7362
7363         if (unlikely(wc->refs[level - 1] == 0)) {
7364                 btrfs_err(root->fs_info, "Missing references.");
7365                 BUG();
7366         }
7367         *lookup_info = 0;
7368
7369         if (wc->stage == DROP_REFERENCE) {
7370                 if (wc->refs[level - 1] > 1) {
7371                         if (level == 1 &&
7372                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7373                                 goto skip;
7374
7375                         if (!wc->update_ref ||
7376                             generation <= root->root_key.offset)
7377                                 goto skip;
7378
7379                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7380                                               path->slots[level]);
7381                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7382                         if (ret < 0)
7383                                 goto skip;
7384
7385                         wc->stage = UPDATE_BACKREF;
7386                         wc->shared_level = level - 1;
7387                 }
7388         } else {
7389                 if (level == 1 &&
7390                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7391                         goto skip;
7392         }
7393
7394         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7395                 btrfs_tree_unlock(next);
7396                 free_extent_buffer(next);
7397                 next = NULL;
7398                 *lookup_info = 1;
7399         }
7400
7401         if (!next) {
7402                 if (reada && level == 1)
7403                         reada_walk_down(trans, root, wc, path);
7404                 next = read_tree_block(root, bytenr, blocksize, generation);
7405                 if (!next || !extent_buffer_uptodate(next)) {
7406                         free_extent_buffer(next);
7407                         return -EIO;
7408                 }
7409                 btrfs_tree_lock(next);
7410                 btrfs_set_lock_blocking(next);
7411         }
7412
7413         level--;
7414         BUG_ON(level != btrfs_header_level(next));
7415         path->nodes[level] = next;
7416         path->slots[level] = 0;
7417         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7418         wc->level = level;
7419         if (wc->level == 1)
7420                 wc->reada_slot = 0;
7421         return 0;
7422 skip:
7423         wc->refs[level - 1] = 0;
7424         wc->flags[level - 1] = 0;
7425         if (wc->stage == DROP_REFERENCE) {
7426                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7427                         parent = path->nodes[level]->start;
7428                 } else {
7429                         BUG_ON(root->root_key.objectid !=
7430                                btrfs_header_owner(path->nodes[level]));
7431                         parent = 0;
7432                 }
7433
7434                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7435                                 root->root_key.objectid, level - 1, 0, 0);
7436                 BUG_ON(ret); /* -ENOMEM */
7437         }
7438         btrfs_tree_unlock(next);
7439         free_extent_buffer(next);
7440         *lookup_info = 1;
7441         return 1;
7442 }
7443
7444 /*
7445  * helper to process tree block while walking up the tree.
7446  *
7447  * when wc->stage == DROP_REFERENCE, this function drops
7448  * reference count on the block.
7449  *
7450  * when wc->stage == UPDATE_BACKREF, this function changes
7451  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7452  * to UPDATE_BACKREF previously while processing the block.
7453  *
7454  * NOTE: return value 1 means we should stop walking up.
7455  */
7456 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7457                                  struct btrfs_root *root,
7458                                  struct btrfs_path *path,
7459                                  struct walk_control *wc)
7460 {
7461         int ret;
7462         int level = wc->level;
7463         struct extent_buffer *eb = path->nodes[level];
7464         u64 parent = 0;
7465
7466         if (wc->stage == UPDATE_BACKREF) {
7467                 BUG_ON(wc->shared_level < level);
7468                 if (level < wc->shared_level)
7469                         goto out;
7470
7471                 ret = find_next_key(path, level + 1, &wc->update_progress);
7472                 if (ret > 0)
7473                         wc->update_ref = 0;
7474
7475                 wc->stage = DROP_REFERENCE;
7476                 wc->shared_level = -1;
7477                 path->slots[level] = 0;
7478
7479                 /*
7480                  * check reference count again if the block isn't locked.
7481                  * we should start walking down the tree again if reference
7482                  * count is one.
7483                  */
7484                 if (!path->locks[level]) {
7485                         BUG_ON(level == 0);
7486                         btrfs_tree_lock(eb);
7487                         btrfs_set_lock_blocking(eb);
7488                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7489
7490                         ret = btrfs_lookup_extent_info(trans, root,
7491                                                        eb->start, level, 1,
7492                                                        &wc->refs[level],
7493                                                        &wc->flags[level]);
7494                         if (ret < 0) {
7495                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7496                                 path->locks[level] = 0;
7497                                 return ret;
7498                         }
7499                         BUG_ON(wc->refs[level] == 0);
7500                         if (wc->refs[level] == 1) {
7501                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7502                                 path->locks[level] = 0;
7503                                 return 1;
7504                         }
7505                 }
7506         }
7507
7508         /* wc->stage == DROP_REFERENCE */
7509         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7510
7511         if (wc->refs[level] == 1) {
7512                 if (level == 0) {
7513                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7514                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7515                                                     wc->for_reloc);
7516                         else
7517                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7518                                                     wc->for_reloc);
7519                         BUG_ON(ret); /* -ENOMEM */
7520                 }
7521                 /* make block locked assertion in clean_tree_block happy */
7522                 if (!path->locks[level] &&
7523                     btrfs_header_generation(eb) == trans->transid) {
7524                         btrfs_tree_lock(eb);
7525                         btrfs_set_lock_blocking(eb);
7526                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7527                 }
7528                 clean_tree_block(trans, root, eb);
7529         }
7530
7531         if (eb == root->node) {
7532                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7533                         parent = eb->start;
7534                 else
7535                         BUG_ON(root->root_key.objectid !=
7536                                btrfs_header_owner(eb));
7537         } else {
7538                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7539                         parent = path->nodes[level + 1]->start;
7540                 else
7541                         BUG_ON(root->root_key.objectid !=
7542                                btrfs_header_owner(path->nodes[level + 1]));
7543         }
7544
7545         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7546 out:
7547         wc->refs[level] = 0;
7548         wc->flags[level] = 0;
7549         return 0;
7550 }
7551
7552 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7553                                    struct btrfs_root *root,
7554                                    struct btrfs_path *path,
7555                                    struct walk_control *wc)
7556 {
7557         int level = wc->level;
7558         int lookup_info = 1;
7559         int ret;
7560
7561         while (level >= 0) {
7562                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7563                 if (ret > 0)
7564                         break;
7565
7566                 if (level == 0)
7567                         break;
7568
7569                 if (path->slots[level] >=
7570                     btrfs_header_nritems(path->nodes[level]))
7571                         break;
7572
7573                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7574                 if (ret > 0) {
7575                         path->slots[level]++;
7576                         continue;
7577                 } else if (ret < 0)
7578                         return ret;
7579                 level = wc->level;
7580         }
7581         return 0;
7582 }
7583
7584 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7585                                  struct btrfs_root *root,
7586                                  struct btrfs_path *path,
7587                                  struct walk_control *wc, int max_level)
7588 {
7589         int level = wc->level;
7590         int ret;
7591
7592         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7593         while (level < max_level && path->nodes[level]) {
7594                 wc->level = level;
7595                 if (path->slots[level] + 1 <
7596                     btrfs_header_nritems(path->nodes[level])) {
7597                         path->slots[level]++;
7598                         return 0;
7599                 } else {
7600                         ret = walk_up_proc(trans, root, path, wc);
7601                         if (ret > 0)
7602                                 return 0;
7603
7604                         if (path->locks[level]) {
7605                                 btrfs_tree_unlock_rw(path->nodes[level],
7606                                                      path->locks[level]);
7607                                 path->locks[level] = 0;
7608                         }
7609                         free_extent_buffer(path->nodes[level]);
7610                         path->nodes[level] = NULL;
7611                         level++;
7612                 }
7613         }
7614         return 1;
7615 }
7616
7617 /*
7618  * drop a subvolume tree.
7619  *
7620  * this function traverses the tree freeing any blocks that only
7621  * referenced by the tree.
7622  *
7623  * when a shared tree block is found. this function decreases its
7624  * reference count by one. if update_ref is true, this function
7625  * also make sure backrefs for the shared block and all lower level
7626  * blocks are properly updated.
7627  *
7628  * If called with for_reloc == 0, may exit early with -EAGAIN
7629  */
7630 int btrfs_drop_snapshot(struct btrfs_root *root,
7631                          struct btrfs_block_rsv *block_rsv, int update_ref,
7632                          int for_reloc)
7633 {
7634         struct btrfs_path *path;
7635         struct btrfs_trans_handle *trans;
7636         struct btrfs_root *tree_root = root->fs_info->tree_root;
7637         struct btrfs_root_item *root_item = &root->root_item;
7638         struct walk_control *wc;
7639         struct btrfs_key key;
7640         int err = 0;
7641         int ret;
7642         int level;
7643         bool root_dropped = false;
7644
7645         path = btrfs_alloc_path();
7646         if (!path) {
7647                 err = -ENOMEM;
7648                 goto out;
7649         }
7650
7651         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7652         if (!wc) {
7653                 btrfs_free_path(path);
7654                 err = -ENOMEM;
7655                 goto out;
7656         }
7657
7658         trans = btrfs_start_transaction(tree_root, 0);
7659         if (IS_ERR(trans)) {
7660                 err = PTR_ERR(trans);
7661                 goto out_free;
7662         }
7663
7664         if (block_rsv)
7665                 trans->block_rsv = block_rsv;
7666
7667         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7668                 level = btrfs_header_level(root->node);
7669                 path->nodes[level] = btrfs_lock_root_node(root);
7670                 btrfs_set_lock_blocking(path->nodes[level]);
7671                 path->slots[level] = 0;
7672                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7673                 memset(&wc->update_progress, 0,
7674                        sizeof(wc->update_progress));
7675         } else {
7676                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7677                 memcpy(&wc->update_progress, &key,
7678                        sizeof(wc->update_progress));
7679
7680                 level = root_item->drop_level;
7681                 BUG_ON(level == 0);
7682                 path->lowest_level = level;
7683                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7684                 path->lowest_level = 0;
7685                 if (ret < 0) {
7686                         err = ret;
7687                         goto out_end_trans;
7688                 }
7689                 WARN_ON(ret > 0);
7690
7691                 /*
7692                  * unlock our path, this is safe because only this
7693                  * function is allowed to delete this snapshot
7694                  */
7695                 btrfs_unlock_up_safe(path, 0);
7696
7697                 level = btrfs_header_level(root->node);
7698                 while (1) {
7699                         btrfs_tree_lock(path->nodes[level]);
7700                         btrfs_set_lock_blocking(path->nodes[level]);
7701                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7702
7703                         ret = btrfs_lookup_extent_info(trans, root,
7704                                                 path->nodes[level]->start,
7705                                                 level, 1, &wc->refs[level],
7706                                                 &wc->flags[level]);
7707                         if (ret < 0) {
7708                                 err = ret;
7709                                 goto out_end_trans;
7710                         }
7711                         BUG_ON(wc->refs[level] == 0);
7712
7713                         if (level == root_item->drop_level)
7714                                 break;
7715
7716                         btrfs_tree_unlock(path->nodes[level]);
7717                         path->locks[level] = 0;
7718                         WARN_ON(wc->refs[level] != 1);
7719                         level--;
7720                 }
7721         }
7722
7723         wc->level = level;
7724         wc->shared_level = -1;
7725         wc->stage = DROP_REFERENCE;
7726         wc->update_ref = update_ref;
7727         wc->keep_locks = 0;
7728         wc->for_reloc = for_reloc;
7729         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7730
7731         while (1) {
7732
7733                 ret = walk_down_tree(trans, root, path, wc);
7734                 if (ret < 0) {
7735                         err = ret;
7736                         break;
7737                 }
7738
7739                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7740                 if (ret < 0) {
7741                         err = ret;
7742                         break;
7743                 }
7744
7745                 if (ret > 0) {
7746                         BUG_ON(wc->stage != DROP_REFERENCE);
7747                         break;
7748                 }
7749
7750                 if (wc->stage == DROP_REFERENCE) {
7751                         level = wc->level;
7752                         btrfs_node_key(path->nodes[level],
7753                                        &root_item->drop_progress,
7754                                        path->slots[level]);
7755                         root_item->drop_level = level;
7756                 }
7757
7758                 BUG_ON(wc->level == 0);
7759                 if (btrfs_should_end_transaction(trans, tree_root) ||
7760                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
7761                         ret = btrfs_update_root(trans, tree_root,
7762                                                 &root->root_key,
7763                                                 root_item);
7764                         if (ret) {
7765                                 btrfs_abort_transaction(trans, tree_root, ret);
7766                                 err = ret;
7767                                 goto out_end_trans;
7768                         }
7769
7770                         btrfs_end_transaction_throttle(trans, tree_root);
7771                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
7772                                 pr_debug("BTRFS: drop snapshot early exit\n");
7773                                 err = -EAGAIN;
7774                                 goto out_free;
7775                         }
7776
7777                         trans = btrfs_start_transaction(tree_root, 0);
7778                         if (IS_ERR(trans)) {
7779                                 err = PTR_ERR(trans);
7780                                 goto out_free;
7781                         }
7782                         if (block_rsv)
7783                                 trans->block_rsv = block_rsv;
7784                 }
7785         }
7786         btrfs_release_path(path);
7787         if (err)
7788                 goto out_end_trans;
7789
7790         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7791         if (ret) {
7792                 btrfs_abort_transaction(trans, tree_root, ret);
7793                 goto out_end_trans;
7794         }
7795
7796         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7797                 ret = btrfs_find_root(tree_root, &root->root_key, path,
7798                                       NULL, NULL);
7799                 if (ret < 0) {
7800                         btrfs_abort_transaction(trans, tree_root, ret);
7801                         err = ret;
7802                         goto out_end_trans;
7803                 } else if (ret > 0) {
7804                         /* if we fail to delete the orphan item this time
7805                          * around, it'll get picked up the next time.
7806                          *
7807                          * The most common failure here is just -ENOENT.
7808                          */
7809                         btrfs_del_orphan_item(trans, tree_root,
7810                                               root->root_key.objectid);
7811                 }
7812         }
7813
7814         if (root->in_radix) {
7815                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
7816         } else {
7817                 free_extent_buffer(root->node);
7818                 free_extent_buffer(root->commit_root);
7819                 btrfs_put_fs_root(root);
7820         }
7821         root_dropped = true;
7822 out_end_trans:
7823         btrfs_end_transaction_throttle(trans, tree_root);
7824 out_free:
7825         kfree(wc);
7826         btrfs_free_path(path);
7827 out:
7828         /*
7829          * So if we need to stop dropping the snapshot for whatever reason we
7830          * need to make sure to add it back to the dead root list so that we
7831          * keep trying to do the work later.  This also cleans up roots if we
7832          * don't have it in the radix (like when we recover after a power fail
7833          * or unmount) so we don't leak memory.
7834          */
7835         if (!for_reloc && root_dropped == false)
7836                 btrfs_add_dead_root(root);
7837         if (err)
7838                 btrfs_std_error(root->fs_info, err);
7839         return err;
7840 }
7841
7842 /*
7843  * drop subtree rooted at tree block 'node'.
7844  *
7845  * NOTE: this function will unlock and release tree block 'node'
7846  * only used by relocation code
7847  */
7848 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7849                         struct btrfs_root *root,
7850                         struct extent_buffer *node,
7851                         struct extent_buffer *parent)
7852 {
7853         struct btrfs_path *path;
7854         struct walk_control *wc;
7855         int level;
7856         int parent_level;
7857         int ret = 0;
7858         int wret;
7859
7860         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7861
7862         path = btrfs_alloc_path();
7863         if (!path)
7864                 return -ENOMEM;
7865
7866         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7867         if (!wc) {
7868                 btrfs_free_path(path);
7869                 return -ENOMEM;
7870         }
7871
7872         btrfs_assert_tree_locked(parent);
7873         parent_level = btrfs_header_level(parent);
7874         extent_buffer_get(parent);
7875         path->nodes[parent_level] = parent;
7876         path->slots[parent_level] = btrfs_header_nritems(parent);
7877
7878         btrfs_assert_tree_locked(node);
7879         level = btrfs_header_level(node);
7880         path->nodes[level] = node;
7881         path->slots[level] = 0;
7882         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7883
7884         wc->refs[parent_level] = 1;
7885         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7886         wc->level = level;
7887         wc->shared_level = -1;
7888         wc->stage = DROP_REFERENCE;
7889         wc->update_ref = 0;
7890         wc->keep_locks = 1;
7891         wc->for_reloc = 1;
7892         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7893
7894         while (1) {
7895                 wret = walk_down_tree(trans, root, path, wc);
7896                 if (wret < 0) {
7897                         ret = wret;
7898                         break;
7899                 }
7900
7901                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7902                 if (wret < 0)
7903                         ret = wret;
7904                 if (wret != 0)
7905                         break;
7906         }
7907
7908         kfree(wc);
7909         btrfs_free_path(path);
7910         return ret;
7911 }
7912
7913 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7914 {
7915         u64 num_devices;
7916         u64 stripped;
7917
7918         /*
7919          * if restripe for this chunk_type is on pick target profile and
7920          * return, otherwise do the usual balance
7921          */
7922         stripped = get_restripe_target(root->fs_info, flags);
7923         if (stripped)
7924                 return extended_to_chunk(stripped);
7925
7926         /*
7927          * we add in the count of missing devices because we want
7928          * to make sure that any RAID levels on a degraded FS
7929          * continue to be honored.
7930          */
7931         num_devices = root->fs_info->fs_devices->rw_devices +
7932                 root->fs_info->fs_devices->missing_devices;
7933
7934         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7935                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7936                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7937
7938         if (num_devices == 1) {
7939                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7940                 stripped = flags & ~stripped;
7941
7942                 /* turn raid0 into single device chunks */
7943                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7944                         return stripped;
7945
7946                 /* turn mirroring into duplication */
7947                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7948                              BTRFS_BLOCK_GROUP_RAID10))
7949                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7950         } else {
7951                 /* they already had raid on here, just return */
7952                 if (flags & stripped)
7953                         return flags;
7954
7955                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7956                 stripped = flags & ~stripped;
7957
7958                 /* switch duplicated blocks with raid1 */
7959                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7960                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7961
7962                 /* this is drive concat, leave it alone */
7963         }
7964
7965         return flags;
7966 }
7967
7968 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7969 {
7970         struct btrfs_space_info *sinfo = cache->space_info;
7971         u64 num_bytes;
7972         u64 min_allocable_bytes;
7973         int ret = -ENOSPC;
7974
7975
7976         /*
7977          * We need some metadata space and system metadata space for
7978          * allocating chunks in some corner cases until we force to set
7979          * it to be readonly.
7980          */
7981         if ((sinfo->flags &
7982              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7983             !force)
7984                 min_allocable_bytes = 1 * 1024 * 1024;
7985         else
7986                 min_allocable_bytes = 0;
7987
7988         spin_lock(&sinfo->lock);
7989         spin_lock(&cache->lock);
7990
7991         if (cache->ro) {
7992                 ret = 0;
7993                 goto out;
7994         }
7995
7996         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7997                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7998
7999         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8000             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8001             min_allocable_bytes <= sinfo->total_bytes) {
8002                 sinfo->bytes_readonly += num_bytes;
8003                 cache->ro = 1;
8004                 ret = 0;
8005         }
8006 out:
8007         spin_unlock(&cache->lock);
8008         spin_unlock(&sinfo->lock);
8009         return ret;
8010 }
8011
8012 int btrfs_set_block_group_ro(struct btrfs_root *root,
8013                              struct btrfs_block_group_cache *cache)
8014
8015 {
8016         struct btrfs_trans_handle *trans;
8017         u64 alloc_flags;
8018         int ret;
8019
8020         BUG_ON(cache->ro);
8021
8022         trans = btrfs_join_transaction(root);
8023         if (IS_ERR(trans))
8024                 return PTR_ERR(trans);
8025
8026         alloc_flags = update_block_group_flags(root, cache->flags);
8027         if (alloc_flags != cache->flags) {
8028                 ret = do_chunk_alloc(trans, root, alloc_flags,
8029                                      CHUNK_ALLOC_FORCE);
8030                 if (ret < 0)
8031                         goto out;
8032         }
8033
8034         ret = set_block_group_ro(cache, 0);
8035         if (!ret)
8036                 goto out;
8037         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8038         ret = do_chunk_alloc(trans, root, alloc_flags,
8039                              CHUNK_ALLOC_FORCE);
8040         if (ret < 0)
8041                 goto out;
8042         ret = set_block_group_ro(cache, 0);
8043 out:
8044         btrfs_end_transaction(trans, root);
8045         return ret;
8046 }
8047
8048 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8049                             struct btrfs_root *root, u64 type)
8050 {
8051         u64 alloc_flags = get_alloc_profile(root, type);
8052         return do_chunk_alloc(trans, root, alloc_flags,
8053                               CHUNK_ALLOC_FORCE);
8054 }
8055
8056 /*
8057  * helper to account the unused space of all the readonly block group in the
8058  * list. takes mirrors into account.
8059  */
8060 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
8061 {
8062         struct btrfs_block_group_cache *block_group;
8063         u64 free_bytes = 0;
8064         int factor;
8065
8066         list_for_each_entry(block_group, groups_list, list) {
8067                 spin_lock(&block_group->lock);
8068
8069                 if (!block_group->ro) {
8070                         spin_unlock(&block_group->lock);
8071                         continue;
8072                 }
8073
8074                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8075                                           BTRFS_BLOCK_GROUP_RAID10 |
8076                                           BTRFS_BLOCK_GROUP_DUP))
8077                         factor = 2;
8078                 else
8079                         factor = 1;
8080
8081                 free_bytes += (block_group->key.offset -
8082                                btrfs_block_group_used(&block_group->item)) *
8083                                factor;
8084
8085                 spin_unlock(&block_group->lock);
8086         }
8087
8088         return free_bytes;
8089 }
8090
8091 /*
8092  * helper to account the unused space of all the readonly block group in the
8093  * space_info. takes mirrors into account.
8094  */
8095 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8096 {
8097         int i;
8098         u64 free_bytes = 0;
8099
8100         spin_lock(&sinfo->lock);
8101
8102         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
8103                 if (!list_empty(&sinfo->block_groups[i]))
8104                         free_bytes += __btrfs_get_ro_block_group_free_space(
8105                                                 &sinfo->block_groups[i]);
8106
8107         spin_unlock(&sinfo->lock);
8108
8109         return free_bytes;
8110 }
8111
8112 void btrfs_set_block_group_rw(struct btrfs_root *root,
8113                               struct btrfs_block_group_cache *cache)
8114 {
8115         struct btrfs_space_info *sinfo = cache->space_info;
8116         u64 num_bytes;
8117
8118         BUG_ON(!cache->ro);
8119
8120         spin_lock(&sinfo->lock);
8121         spin_lock(&cache->lock);
8122         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8123                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8124         sinfo->bytes_readonly -= num_bytes;
8125         cache->ro = 0;
8126         spin_unlock(&cache->lock);
8127         spin_unlock(&sinfo->lock);
8128 }
8129
8130 /*
8131  * checks to see if its even possible to relocate this block group.
8132  *
8133  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8134  * ok to go ahead and try.
8135  */
8136 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8137 {
8138         struct btrfs_block_group_cache *block_group;
8139         struct btrfs_space_info *space_info;
8140         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8141         struct btrfs_device *device;
8142         struct btrfs_trans_handle *trans;
8143         u64 min_free;
8144         u64 dev_min = 1;
8145         u64 dev_nr = 0;
8146         u64 target;
8147         int index;
8148         int full = 0;
8149         int ret = 0;
8150
8151         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8152
8153         /* odd, couldn't find the block group, leave it alone */
8154         if (!block_group)
8155                 return -1;
8156
8157         min_free = btrfs_block_group_used(&block_group->item);
8158
8159         /* no bytes used, we're good */
8160         if (!min_free)
8161                 goto out;
8162
8163         space_info = block_group->space_info;
8164         spin_lock(&space_info->lock);
8165
8166         full = space_info->full;
8167
8168         /*
8169          * if this is the last block group we have in this space, we can't
8170          * relocate it unless we're able to allocate a new chunk below.
8171          *
8172          * Otherwise, we need to make sure we have room in the space to handle
8173          * all of the extents from this block group.  If we can, we're good
8174          */
8175         if ((space_info->total_bytes != block_group->key.offset) &&
8176             (space_info->bytes_used + space_info->bytes_reserved +
8177              space_info->bytes_pinned + space_info->bytes_readonly +
8178              min_free < space_info->total_bytes)) {
8179                 spin_unlock(&space_info->lock);
8180                 goto out;
8181         }
8182         spin_unlock(&space_info->lock);
8183
8184         /*
8185          * ok we don't have enough space, but maybe we have free space on our
8186          * devices to allocate new chunks for relocation, so loop through our
8187          * alloc devices and guess if we have enough space.  if this block
8188          * group is going to be restriped, run checks against the target
8189          * profile instead of the current one.
8190          */
8191         ret = -1;
8192
8193         /*
8194          * index:
8195          *      0: raid10
8196          *      1: raid1
8197          *      2: dup
8198          *      3: raid0
8199          *      4: single
8200          */
8201         target = get_restripe_target(root->fs_info, block_group->flags);
8202         if (target) {
8203                 index = __get_raid_index(extended_to_chunk(target));
8204         } else {
8205                 /*
8206                  * this is just a balance, so if we were marked as full
8207                  * we know there is no space for a new chunk
8208                  */
8209                 if (full)
8210                         goto out;
8211
8212                 index = get_block_group_index(block_group);
8213         }
8214
8215         if (index == BTRFS_RAID_RAID10) {
8216                 dev_min = 4;
8217                 /* Divide by 2 */
8218                 min_free >>= 1;
8219         } else if (index == BTRFS_RAID_RAID1) {
8220                 dev_min = 2;
8221         } else if (index == BTRFS_RAID_DUP) {
8222                 /* Multiply by 2 */
8223                 min_free <<= 1;
8224         } else if (index == BTRFS_RAID_RAID0) {
8225                 dev_min = fs_devices->rw_devices;
8226                 do_div(min_free, dev_min);
8227         }
8228
8229         /* We need to do this so that we can look at pending chunks */
8230         trans = btrfs_join_transaction(root);
8231         if (IS_ERR(trans)) {
8232                 ret = PTR_ERR(trans);
8233                 goto out;
8234         }
8235
8236         mutex_lock(&root->fs_info->chunk_mutex);
8237         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8238                 u64 dev_offset;
8239
8240                 /*
8241                  * check to make sure we can actually find a chunk with enough
8242                  * space to fit our block group in.
8243                  */
8244                 if (device->total_bytes > device->bytes_used + min_free &&
8245                     !device->is_tgtdev_for_dev_replace) {
8246                         ret = find_free_dev_extent(trans, device, min_free,
8247                                                    &dev_offset, NULL);
8248                         if (!ret)
8249                                 dev_nr++;
8250
8251                         if (dev_nr >= dev_min)
8252                                 break;
8253
8254                         ret = -1;
8255                 }
8256         }
8257         mutex_unlock(&root->fs_info->chunk_mutex);
8258         btrfs_end_transaction(trans, root);
8259 out:
8260         btrfs_put_block_group(block_group);
8261         return ret;
8262 }
8263
8264 static int find_first_block_group(struct btrfs_root *root,
8265                 struct btrfs_path *path, struct btrfs_key *key)
8266 {
8267         int ret = 0;
8268         struct btrfs_key found_key;
8269         struct extent_buffer *leaf;
8270         int slot;
8271
8272         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8273         if (ret < 0)
8274                 goto out;
8275
8276         while (1) {
8277                 slot = path->slots[0];
8278                 leaf = path->nodes[0];
8279                 if (slot >= btrfs_header_nritems(leaf)) {
8280                         ret = btrfs_next_leaf(root, path);
8281                         if (ret == 0)
8282                                 continue;
8283                         if (ret < 0)
8284                                 goto out;
8285                         break;
8286                 }
8287                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8288
8289                 if (found_key.objectid >= key->objectid &&
8290                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8291                         ret = 0;
8292                         goto out;
8293                 }
8294                 path->slots[0]++;
8295         }
8296 out:
8297         return ret;
8298 }
8299
8300 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8301 {
8302         struct btrfs_block_group_cache *block_group;
8303         u64 last = 0;
8304
8305         while (1) {
8306                 struct inode *inode;
8307
8308                 block_group = btrfs_lookup_first_block_group(info, last);
8309                 while (block_group) {
8310                         spin_lock(&block_group->lock);
8311                         if (block_group->iref)
8312                                 break;
8313                         spin_unlock(&block_group->lock);
8314                         block_group = next_block_group(info->tree_root,
8315                                                        block_group);
8316                 }
8317                 if (!block_group) {
8318                         if (last == 0)
8319                                 break;
8320                         last = 0;
8321                         continue;
8322                 }
8323
8324                 inode = block_group->inode;
8325                 block_group->iref = 0;
8326                 block_group->inode = NULL;
8327                 spin_unlock(&block_group->lock);
8328                 iput(inode);
8329                 last = block_group->key.objectid + block_group->key.offset;
8330                 btrfs_put_block_group(block_group);
8331         }
8332 }
8333
8334 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8335 {
8336         struct btrfs_block_group_cache *block_group;
8337         struct btrfs_space_info *space_info;
8338         struct btrfs_caching_control *caching_ctl;
8339         struct rb_node *n;
8340
8341         down_write(&info->extent_commit_sem);
8342         while (!list_empty(&info->caching_block_groups)) {
8343                 caching_ctl = list_entry(info->caching_block_groups.next,
8344                                          struct btrfs_caching_control, list);
8345                 list_del(&caching_ctl->list);
8346                 put_caching_control(caching_ctl);
8347         }
8348         up_write(&info->extent_commit_sem);
8349
8350         spin_lock(&info->block_group_cache_lock);
8351         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8352                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8353                                        cache_node);
8354                 rb_erase(&block_group->cache_node,
8355                          &info->block_group_cache_tree);
8356                 spin_unlock(&info->block_group_cache_lock);
8357
8358                 down_write(&block_group->space_info->groups_sem);
8359                 list_del(&block_group->list);
8360                 up_write(&block_group->space_info->groups_sem);
8361
8362                 if (block_group->cached == BTRFS_CACHE_STARTED)
8363                         wait_block_group_cache_done(block_group);
8364
8365                 /*
8366                  * We haven't cached this block group, which means we could
8367                  * possibly have excluded extents on this block group.
8368                  */
8369                 if (block_group->cached == BTRFS_CACHE_NO ||
8370                     block_group->cached == BTRFS_CACHE_ERROR)
8371                         free_excluded_extents(info->extent_root, block_group);
8372
8373                 btrfs_remove_free_space_cache(block_group);
8374                 btrfs_put_block_group(block_group);
8375
8376                 spin_lock(&info->block_group_cache_lock);
8377         }
8378         spin_unlock(&info->block_group_cache_lock);
8379
8380         /* now that all the block groups are freed, go through and
8381          * free all the space_info structs.  This is only called during
8382          * the final stages of unmount, and so we know nobody is
8383          * using them.  We call synchronize_rcu() once before we start,
8384          * just to be on the safe side.
8385          */
8386         synchronize_rcu();
8387
8388         release_global_block_rsv(info);
8389
8390         while (!list_empty(&info->space_info)) {
8391                 int i;
8392
8393                 space_info = list_entry(info->space_info.next,
8394                                         struct btrfs_space_info,
8395                                         list);
8396                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8397                         if (WARN_ON(space_info->bytes_pinned > 0 ||
8398                             space_info->bytes_reserved > 0 ||
8399                             space_info->bytes_may_use > 0)) {
8400                                 dump_space_info(space_info, 0, 0);
8401                         }
8402                 }
8403                 list_del(&space_info->list);
8404                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
8405                         struct kobject *kobj;
8406                         kobj = &space_info->block_group_kobjs[i];
8407                         if (kobj->parent) {
8408                                 kobject_del(kobj);
8409                                 kobject_put(kobj);
8410                         }
8411                 }
8412                 kobject_del(&space_info->kobj);
8413                 kobject_put(&space_info->kobj);
8414         }
8415         return 0;
8416 }
8417
8418 static void __link_block_group(struct btrfs_space_info *space_info,
8419                                struct btrfs_block_group_cache *cache)
8420 {
8421         int index = get_block_group_index(cache);
8422
8423         down_write(&space_info->groups_sem);
8424         if (list_empty(&space_info->block_groups[index])) {
8425                 struct kobject *kobj = &space_info->block_group_kobjs[index];
8426                 int ret;
8427
8428                 kobject_get(&space_info->kobj); /* put in release */
8429                 ret = kobject_add(kobj, &space_info->kobj, "%s",
8430                                   get_raid_name(index));
8431                 if (ret) {
8432                         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
8433                         kobject_put(&space_info->kobj);
8434                 }
8435         }
8436         list_add_tail(&cache->list, &space_info->block_groups[index]);
8437         up_write(&space_info->groups_sem);
8438 }
8439
8440 int btrfs_read_block_groups(struct btrfs_root *root)
8441 {
8442         struct btrfs_path *path;
8443         int ret;
8444         struct btrfs_block_group_cache *cache;
8445         struct btrfs_fs_info *info = root->fs_info;
8446         struct btrfs_space_info *space_info;
8447         struct btrfs_key key;
8448         struct btrfs_key found_key;
8449         struct extent_buffer *leaf;
8450         int need_clear = 0;
8451         u64 cache_gen;
8452
8453         root = info->extent_root;
8454         key.objectid = 0;
8455         key.offset = 0;
8456         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8457         path = btrfs_alloc_path();
8458         if (!path)
8459                 return -ENOMEM;
8460         path->reada = 1;
8461
8462         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
8463         if (btrfs_test_opt(root, SPACE_CACHE) &&
8464             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
8465                 need_clear = 1;
8466         if (btrfs_test_opt(root, CLEAR_CACHE))
8467                 need_clear = 1;
8468
8469         while (1) {
8470                 ret = find_first_block_group(root, path, &key);
8471                 if (ret > 0)
8472                         break;
8473                 if (ret != 0)
8474                         goto error;
8475                 leaf = path->nodes[0];
8476                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8477                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
8478                 if (!cache) {
8479                         ret = -ENOMEM;
8480                         goto error;
8481                 }
8482                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8483                                                 GFP_NOFS);
8484                 if (!cache->free_space_ctl) {
8485                         kfree(cache);
8486                         ret = -ENOMEM;
8487                         goto error;
8488                 }
8489
8490                 atomic_set(&cache->count, 1);
8491                 spin_lock_init(&cache->lock);
8492                 cache->fs_info = info;
8493                 INIT_LIST_HEAD(&cache->list);
8494                 INIT_LIST_HEAD(&cache->cluster_list);
8495
8496                 if (need_clear) {
8497                         /*
8498                          * When we mount with old space cache, we need to
8499                          * set BTRFS_DC_CLEAR and set dirty flag.
8500                          *
8501                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8502                          *    truncate the old free space cache inode and
8503                          *    setup a new one.
8504                          * b) Setting 'dirty flag' makes sure that we flush
8505                          *    the new space cache info onto disk.
8506                          */
8507                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8508                         if (btrfs_test_opt(root, SPACE_CACHE))
8509                                 cache->dirty = 1;
8510                 }
8511
8512                 read_extent_buffer(leaf, &cache->item,
8513                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8514                                    sizeof(cache->item));
8515                 memcpy(&cache->key, &found_key, sizeof(found_key));
8516
8517                 key.objectid = found_key.objectid + found_key.offset;
8518                 btrfs_release_path(path);
8519                 cache->flags = btrfs_block_group_flags(&cache->item);
8520                 cache->sectorsize = root->sectorsize;
8521                 cache->full_stripe_len = btrfs_full_stripe_len(root,
8522                                                &root->fs_info->mapping_tree,
8523                                                found_key.objectid);
8524                 btrfs_init_free_space_ctl(cache);
8525
8526                 /*
8527                  * We need to exclude the super stripes now so that the space
8528                  * info has super bytes accounted for, otherwise we'll think
8529                  * we have more space than we actually do.
8530                  */
8531                 ret = exclude_super_stripes(root, cache);
8532                 if (ret) {
8533                         /*
8534                          * We may have excluded something, so call this just in
8535                          * case.
8536                          */
8537                         free_excluded_extents(root, cache);
8538                         kfree(cache->free_space_ctl);
8539                         kfree(cache);
8540                         goto error;
8541                 }
8542
8543                 /*
8544                  * check for two cases, either we are full, and therefore
8545                  * don't need to bother with the caching work since we won't
8546                  * find any space, or we are empty, and we can just add all
8547                  * the space in and be done with it.  This saves us _alot_ of
8548                  * time, particularly in the full case.
8549                  */
8550                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8551                         cache->last_byte_to_unpin = (u64)-1;
8552                         cache->cached = BTRFS_CACHE_FINISHED;
8553                         free_excluded_extents(root, cache);
8554                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8555                         cache->last_byte_to_unpin = (u64)-1;
8556                         cache->cached = BTRFS_CACHE_FINISHED;
8557                         add_new_free_space(cache, root->fs_info,
8558                                            found_key.objectid,
8559                                            found_key.objectid +
8560                                            found_key.offset);
8561                         free_excluded_extents(root, cache);
8562                 }
8563
8564                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8565                 if (ret) {
8566                         btrfs_remove_free_space_cache(cache);
8567                         btrfs_put_block_group(cache);
8568                         goto error;
8569                 }
8570
8571                 ret = update_space_info(info, cache->flags, found_key.offset,
8572                                         btrfs_block_group_used(&cache->item),
8573                                         &space_info);
8574                 if (ret) {
8575                         btrfs_remove_free_space_cache(cache);
8576                         spin_lock(&info->block_group_cache_lock);
8577                         rb_erase(&cache->cache_node,
8578                                  &info->block_group_cache_tree);
8579                         spin_unlock(&info->block_group_cache_lock);
8580                         btrfs_put_block_group(cache);
8581                         goto error;
8582                 }
8583
8584                 cache->space_info = space_info;
8585                 spin_lock(&cache->space_info->lock);
8586                 cache->space_info->bytes_readonly += cache->bytes_super;
8587                 spin_unlock(&cache->space_info->lock);
8588
8589                 __link_block_group(space_info, cache);
8590
8591                 set_avail_alloc_bits(root->fs_info, cache->flags);
8592                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8593                         set_block_group_ro(cache, 1);
8594         }
8595
8596         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8597                 if (!(get_alloc_profile(root, space_info->flags) &
8598                       (BTRFS_BLOCK_GROUP_RAID10 |
8599                        BTRFS_BLOCK_GROUP_RAID1 |
8600                        BTRFS_BLOCK_GROUP_RAID5 |
8601                        BTRFS_BLOCK_GROUP_RAID6 |
8602                        BTRFS_BLOCK_GROUP_DUP)))
8603                         continue;
8604                 /*
8605                  * avoid allocating from un-mirrored block group if there are
8606                  * mirrored block groups.
8607                  */
8608                 list_for_each_entry(cache,
8609                                 &space_info->block_groups[BTRFS_RAID_RAID0],
8610                                 list)
8611                         set_block_group_ro(cache, 1);
8612                 list_for_each_entry(cache,
8613                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
8614                                 list)
8615                         set_block_group_ro(cache, 1);
8616         }
8617
8618         init_global_block_rsv(info);
8619         ret = 0;
8620 error:
8621         btrfs_free_path(path);
8622         return ret;
8623 }
8624
8625 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8626                                        struct btrfs_root *root)
8627 {
8628         struct btrfs_block_group_cache *block_group, *tmp;
8629         struct btrfs_root *extent_root = root->fs_info->extent_root;
8630         struct btrfs_block_group_item item;
8631         struct btrfs_key key;
8632         int ret = 0;
8633
8634         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8635                                  new_bg_list) {
8636                 list_del_init(&block_group->new_bg_list);
8637
8638                 if (ret)
8639                         continue;
8640
8641                 spin_lock(&block_group->lock);
8642                 memcpy(&item, &block_group->item, sizeof(item));
8643                 memcpy(&key, &block_group->key, sizeof(key));
8644                 spin_unlock(&block_group->lock);
8645
8646                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8647                                         sizeof(item));
8648                 if (ret)
8649                         btrfs_abort_transaction(trans, extent_root, ret);
8650                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
8651                                                key.objectid, key.offset);
8652                 if (ret)
8653                         btrfs_abort_transaction(trans, extent_root, ret);
8654         }
8655 }
8656
8657 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8658                            struct btrfs_root *root, u64 bytes_used,
8659                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8660                            u64 size)
8661 {
8662         int ret;
8663         struct btrfs_root *extent_root;
8664         struct btrfs_block_group_cache *cache;
8665
8666         extent_root = root->fs_info->extent_root;
8667
8668         root->fs_info->last_trans_log_full_commit = trans->transid;
8669
8670         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8671         if (!cache)
8672                 return -ENOMEM;
8673         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8674                                         GFP_NOFS);
8675         if (!cache->free_space_ctl) {
8676                 kfree(cache);
8677                 return -ENOMEM;
8678         }
8679
8680         cache->key.objectid = chunk_offset;
8681         cache->key.offset = size;
8682         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8683         cache->sectorsize = root->sectorsize;
8684         cache->fs_info = root->fs_info;
8685         cache->full_stripe_len = btrfs_full_stripe_len(root,
8686                                                &root->fs_info->mapping_tree,
8687                                                chunk_offset);
8688
8689         atomic_set(&cache->count, 1);
8690         spin_lock_init(&cache->lock);
8691         INIT_LIST_HEAD(&cache->list);
8692         INIT_LIST_HEAD(&cache->cluster_list);
8693         INIT_LIST_HEAD(&cache->new_bg_list);
8694
8695         btrfs_init_free_space_ctl(cache);
8696
8697         btrfs_set_block_group_used(&cache->item, bytes_used);
8698         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8699         cache->flags = type;
8700         btrfs_set_block_group_flags(&cache->item, type);
8701
8702         cache->last_byte_to_unpin = (u64)-1;
8703         cache->cached = BTRFS_CACHE_FINISHED;
8704         ret = exclude_super_stripes(root, cache);
8705         if (ret) {
8706                 /*
8707                  * We may have excluded something, so call this just in
8708                  * case.
8709                  */
8710                 free_excluded_extents(root, cache);
8711                 kfree(cache->free_space_ctl);
8712                 kfree(cache);
8713                 return ret;
8714         }
8715
8716         add_new_free_space(cache, root->fs_info, chunk_offset,
8717                            chunk_offset + size);
8718
8719         free_excluded_extents(root, cache);
8720
8721         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8722         if (ret) {
8723                 btrfs_remove_free_space_cache(cache);
8724                 btrfs_put_block_group(cache);
8725                 return ret;
8726         }
8727
8728         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8729                                 &cache->space_info);
8730         if (ret) {
8731                 btrfs_remove_free_space_cache(cache);
8732                 spin_lock(&root->fs_info->block_group_cache_lock);
8733                 rb_erase(&cache->cache_node,
8734                          &root->fs_info->block_group_cache_tree);
8735                 spin_unlock(&root->fs_info->block_group_cache_lock);
8736                 btrfs_put_block_group(cache);
8737                 return ret;
8738         }
8739         update_global_block_rsv(root->fs_info);
8740
8741         spin_lock(&cache->space_info->lock);
8742         cache->space_info->bytes_readonly += cache->bytes_super;
8743         spin_unlock(&cache->space_info->lock);
8744
8745         __link_block_group(cache->space_info, cache);
8746
8747         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8748
8749         set_avail_alloc_bits(extent_root->fs_info, type);
8750
8751         return 0;
8752 }
8753
8754 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8755 {
8756         u64 extra_flags = chunk_to_extended(flags) &
8757                                 BTRFS_EXTENDED_PROFILE_MASK;
8758
8759         write_seqlock(&fs_info->profiles_lock);
8760         if (flags & BTRFS_BLOCK_GROUP_DATA)
8761                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8762         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8763                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8764         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8765                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8766         write_sequnlock(&fs_info->profiles_lock);
8767 }
8768
8769 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8770                              struct btrfs_root *root, u64 group_start)
8771 {
8772         struct btrfs_path *path;
8773         struct btrfs_block_group_cache *block_group;
8774         struct btrfs_free_cluster *cluster;
8775         struct btrfs_root *tree_root = root->fs_info->tree_root;
8776         struct btrfs_key key;
8777         struct inode *inode;
8778         int ret;
8779         int index;
8780         int factor;
8781
8782         root = root->fs_info->extent_root;
8783
8784         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8785         BUG_ON(!block_group);
8786         BUG_ON(!block_group->ro);
8787
8788         /*
8789          * Free the reserved super bytes from this block group before
8790          * remove it.
8791          */
8792         free_excluded_extents(root, block_group);
8793
8794         memcpy(&key, &block_group->key, sizeof(key));
8795         index = get_block_group_index(block_group);
8796         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8797                                   BTRFS_BLOCK_GROUP_RAID1 |
8798                                   BTRFS_BLOCK_GROUP_RAID10))
8799                 factor = 2;
8800         else
8801                 factor = 1;
8802
8803         /* make sure this block group isn't part of an allocation cluster */
8804         cluster = &root->fs_info->data_alloc_cluster;
8805         spin_lock(&cluster->refill_lock);
8806         btrfs_return_cluster_to_free_space(block_group, cluster);
8807         spin_unlock(&cluster->refill_lock);
8808
8809         /*
8810          * make sure this block group isn't part of a metadata
8811          * allocation cluster
8812          */
8813         cluster = &root->fs_info->meta_alloc_cluster;
8814         spin_lock(&cluster->refill_lock);
8815         btrfs_return_cluster_to_free_space(block_group, cluster);
8816         spin_unlock(&cluster->refill_lock);
8817
8818         path = btrfs_alloc_path();
8819         if (!path) {
8820                 ret = -ENOMEM;
8821                 goto out;
8822         }
8823
8824         inode = lookup_free_space_inode(tree_root, block_group, path);
8825         if (!IS_ERR(inode)) {
8826                 ret = btrfs_orphan_add(trans, inode);
8827                 if (ret) {
8828                         btrfs_add_delayed_iput(inode);
8829                         goto out;
8830                 }
8831                 clear_nlink(inode);
8832                 /* One for the block groups ref */
8833                 spin_lock(&block_group->lock);
8834                 if (block_group->iref) {
8835                         block_group->iref = 0;
8836                         block_group->inode = NULL;
8837                         spin_unlock(&block_group->lock);
8838                         iput(inode);
8839                 } else {
8840                         spin_unlock(&block_group->lock);
8841                 }
8842                 /* One for our lookup ref */
8843                 btrfs_add_delayed_iput(inode);
8844         }
8845
8846         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8847         key.offset = block_group->key.objectid;
8848         key.type = 0;
8849
8850         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8851         if (ret < 0)
8852                 goto out;
8853         if (ret > 0)
8854                 btrfs_release_path(path);
8855         if (ret == 0) {
8856                 ret = btrfs_del_item(trans, tree_root, path);
8857                 if (ret)
8858                         goto out;
8859                 btrfs_release_path(path);
8860         }
8861
8862         spin_lock(&root->fs_info->block_group_cache_lock);
8863         rb_erase(&block_group->cache_node,
8864                  &root->fs_info->block_group_cache_tree);
8865
8866         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8867                 root->fs_info->first_logical_byte = (u64)-1;
8868         spin_unlock(&root->fs_info->block_group_cache_lock);
8869
8870         down_write(&block_group->space_info->groups_sem);
8871         /*
8872          * we must use list_del_init so people can check to see if they
8873          * are still on the list after taking the semaphore
8874          */
8875         list_del_init(&block_group->list);
8876         if (list_empty(&block_group->space_info->block_groups[index])) {
8877                 kobject_del(&block_group->space_info->block_group_kobjs[index]);
8878                 kobject_put(&block_group->space_info->block_group_kobjs[index]);
8879                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8880         }
8881         up_write(&block_group->space_info->groups_sem);
8882
8883         if (block_group->cached == BTRFS_CACHE_STARTED)
8884                 wait_block_group_cache_done(block_group);
8885
8886         btrfs_remove_free_space_cache(block_group);
8887
8888         spin_lock(&block_group->space_info->lock);
8889         block_group->space_info->total_bytes -= block_group->key.offset;
8890         block_group->space_info->bytes_readonly -= block_group->key.offset;
8891         block_group->space_info->disk_total -= block_group->key.offset * factor;
8892         spin_unlock(&block_group->space_info->lock);
8893
8894         memcpy(&key, &block_group->key, sizeof(key));
8895
8896         btrfs_clear_space_info_full(root->fs_info);
8897
8898         btrfs_put_block_group(block_group);
8899         btrfs_put_block_group(block_group);
8900
8901         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8902         if (ret > 0)
8903                 ret = -EIO;
8904         if (ret < 0)
8905                 goto out;
8906
8907         ret = btrfs_del_item(trans, root, path);
8908 out:
8909         btrfs_free_path(path);
8910         return ret;
8911 }
8912
8913 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8914 {
8915         struct btrfs_space_info *space_info;
8916         struct btrfs_super_block *disk_super;
8917         u64 features;
8918         u64 flags;
8919         int mixed = 0;
8920         int ret;
8921
8922         disk_super = fs_info->super_copy;
8923         if (!btrfs_super_root(disk_super))
8924                 return 1;
8925
8926         features = btrfs_super_incompat_flags(disk_super);
8927         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8928                 mixed = 1;
8929
8930         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8931         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8932         if (ret)
8933                 goto out;
8934
8935         if (mixed) {
8936                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8937                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8938         } else {
8939                 flags = BTRFS_BLOCK_GROUP_METADATA;
8940                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8941                 if (ret)
8942                         goto out;
8943
8944                 flags = BTRFS_BLOCK_GROUP_DATA;
8945                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8946         }
8947 out:
8948         return ret;
8949 }
8950
8951 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8952 {
8953         return unpin_extent_range(root, start, end);
8954 }
8955
8956 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8957                                u64 num_bytes, u64 *actual_bytes)
8958 {
8959         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8960 }
8961
8962 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8963 {
8964         struct btrfs_fs_info *fs_info = root->fs_info;
8965         struct btrfs_block_group_cache *cache = NULL;
8966         u64 group_trimmed;
8967         u64 start;
8968         u64 end;
8969         u64 trimmed = 0;
8970         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8971         int ret = 0;
8972
8973         /*
8974          * try to trim all FS space, our block group may start from non-zero.
8975          */
8976         if (range->len == total_bytes)
8977                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8978         else
8979                 cache = btrfs_lookup_block_group(fs_info, range->start);
8980
8981         while (cache) {
8982                 if (cache->key.objectid >= (range->start + range->len)) {
8983                         btrfs_put_block_group(cache);
8984                         break;
8985                 }
8986
8987                 start = max(range->start, cache->key.objectid);
8988                 end = min(range->start + range->len,
8989                                 cache->key.objectid + cache->key.offset);
8990
8991                 if (end - start >= range->minlen) {
8992                         if (!block_group_cache_done(cache)) {
8993                                 ret = cache_block_group(cache, 0);
8994                                 if (ret) {
8995                                         btrfs_put_block_group(cache);
8996                                         break;
8997                                 }
8998                                 ret = wait_block_group_cache_done(cache);
8999                                 if (ret) {
9000                                         btrfs_put_block_group(cache);
9001                                         break;
9002                                 }
9003                         }
9004                         ret = btrfs_trim_block_group(cache,
9005                                                      &group_trimmed,
9006                                                      start,
9007                                                      end,
9008                                                      range->minlen);
9009
9010                         trimmed += group_trimmed;
9011                         if (ret) {
9012                                 btrfs_put_block_group(cache);
9013                                 break;
9014                         }
9015                 }
9016
9017                 cache = next_block_group(fs_info->tree_root, cache);
9018         }
9019
9020         range->len = trimmed;
9021         return ret;
9022 }