fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include <linux/percpu_counter.h>
  28 #include "hash.h"
  29 #include "tree-log.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "volumes.h"
  33 #include "raid56.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36 #include "free-space-tree.h"
  37 #include "math.h"
  38 #include "sysfs.h"
  39 #include "qgroup.h"
  40
  41 #undef SCRAMBLE_DELAYED_REFS
  42
  43 /*
  44  * control flags for do_chunk_alloc's force field
  45  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  46  * if we really need one.
  47  *
  48  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  49  * if we have very few chunks already allocated.  This is
  50  * used as part of the clustering code to help make sure
  51  * we have a good pool of storage to cluster in, without
  52  * filling the FS with empty chunks
  53  *
  54  * CHUNK_ALLOC_FORCE means it must try to allocate one
  55  *
  56  */
  57 enum {
  58         CHUNK_ALLOC_NO_FORCE = 0,
  59         CHUNK_ALLOC_LIMITED = 1,
  60         CHUNK_ALLOC_FORCE = 2,
  61 };
  62
  63 static int update_block_group(struct btrfs_trans_handle *trans,
  64                               struct btrfs_fs_info *fs_info, u64 bytenr,
  65                               u64 num_bytes, int alloc);
  66 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  67                                struct btrfs_fs_info *fs_info,
  68                                 struct btrfs_delayed_ref_node *node, u64 parent,
  69                                 u64 root_objectid, u64 owner_objectid,
  70                                 u64 owner_offset, int refs_to_drop,
  71                                 struct btrfs_delayed_extent_op *extra_op);
  72 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  73                                     struct extent_buffer *leaf,
  74                                     struct btrfs_extent_item *ei);
  75 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  76                                       struct btrfs_fs_info *fs_info,
  77                                       u64 parent, u64 root_objectid,
  78                                       u64 flags, u64 owner, u64 offset,
  79                                       struct btrfs_key *ins, int ref_mod);
  80 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  81                                      struct btrfs_fs_info *fs_info,
  82                                      u64 parent, u64 root_objectid,
  83                                      u64 flags, struct btrfs_disk_key *key,
  84                                      int level, struct btrfs_key *ins);
  85 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  86                           struct btrfs_fs_info *fs_info, u64 flags,
  87                           int force);
  88 static int find_next_key(struct btrfs_path *path, int level,
  89                          struct btrfs_key *key);
  90 static void dump_space_info(struct btrfs_fs_info *fs_info,
  91                             struct btrfs_space_info *info, u64 bytes,
  92                             int dump_block_groups);
  93 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
  94                                     u64 ram_bytes, u64 num_bytes, int delalloc);
  95 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
  96                                      u64 num_bytes, int delalloc);
  97 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  98                                u64 num_bytes);
  99 static int __reserve_metadata_bytes(struct btrfs_root *root,
 100                                     struct btrfs_space_info *space_info,
 101                                     u64 orig_bytes,
 102                                     enum btrfs_reserve_flush_enum flush);
 103 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
 104                                      struct btrfs_space_info *space_info,
 105                                      u64 num_bytes);
 106 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 107                                      struct btrfs_space_info *space_info,
 108                                      u64 num_bytes);
 109
 110 static noinline int
 111 block_group_cache_done(struct btrfs_block_group_cache *cache)
 112 {
 113         smp_mb();
 114         return cache->cached == BTRFS_CACHE_FINISHED ||
 115                 cache->cached == BTRFS_CACHE_ERROR;
 116 }
 117
 118 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 119 {
 120         return (cache->flags & bits) == bits;
 121 }
 122
 123 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 124 {
 125         atomic_inc(&cache->count);
 126 }
 127
 128 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 129 {
 130         if (atomic_dec_and_test(&cache->count)) {
 131                 WARN_ON(cache->pinned > 0);
 132                 WARN_ON(cache->reserved > 0);
 133                 kfree(cache->free_space_ctl);
 134                 kfree(cache);
 135         }
 136 }
 137
 138 /*
 139  * this adds the block group to the fs_info rb tree for the block group
 140  * cache
 141  */
 142 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 143                                 struct btrfs_block_group_cache *block_group)
 144 {
 145         struct rb_node **p;
 146         struct rb_node *parent = NULL;
 147         struct btrfs_block_group_cache *cache;
 148
 149         spin_lock(&info->block_group_cache_lock);
 150         p = &info->block_group_cache_tree.rb_node;
 151
 152         while (*p) {
 153                 parent = *p;
 154                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 155                                  cache_node);
 156                 if (block_group->key.objectid < cache->key.objectid) {
 157                         p = &(*p)->rb_left;
 158                 } else if (block_group->key.objectid > cache->key.objectid) {
 159                         p = &(*p)->rb_right;
 160                 } else {
 161                         spin_unlock(&info->block_group_cache_lock);
 162                         return -EEXIST;
 163                 }
 164         }
 165
 166         rb_link_node(&block_group->cache_node, parent, p);
 167         rb_insert_color(&block_group->cache_node,
 168                         &info->block_group_cache_tree);
 169
 170         if (info->first_logical_byte > block_group->key.objectid)
 171                 info->first_logical_byte = block_group->key.objectid;
 172
 173         spin_unlock(&info->block_group_cache_lock);
 174
 175         return 0;
 176 }
 177
 178 /*
 179  * This will return the block group at or after bytenr if contains is 0, else
 180  * it will return the block group that contains the bytenr
 181  */
 182 static struct btrfs_block_group_cache *
 183 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 184                               int contains)
 185 {
 186         struct btrfs_block_group_cache *cache, *ret = NULL;
 187         struct rb_node *n;
 188         u64 end, start;
 189
 190         spin_lock(&info->block_group_cache_lock);
 191         n = info->block_group_cache_tree.rb_node;
 192
 193         while (n) {
 194                 cache = rb_entry(n, struct btrfs_block_group_cache,
 195                                  cache_node);
 196                 end = cache->key.objectid + cache->key.offset - 1;
 197                 start = cache->key.objectid;
 198
 199                 if (bytenr < start) {
 200                         if (!contains && (!ret || start < ret->key.objectid))
 201                                 ret = cache;
 202                         n = n->rb_left;
 203                 } else if (bytenr > start) {
 204                         if (contains && bytenr <= end) {
 205                                 ret = cache;
 206                                 break;
 207                         }
 208                         n = n->rb_right;
 209                 } else {
 210                         ret = cache;
 211                         break;
 212                 }
 213         }
 214         if (ret) {
 215                 btrfs_get_block_group(ret);
 216                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 217                         info->first_logical_byte = ret->key.objectid;
 218         }
 219         spin_unlock(&info->block_group_cache_lock);
 220
 221         return ret;
 222 }
 223
 224 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 225                                u64 start, u64 num_bytes)
 226 {
 227         u64 end = start + num_bytes - 1;
 228         set_extent_bits(&fs_info->freed_extents[0],
 229                         start, end, EXTENT_UPTODATE);
 230         set_extent_bits(&fs_info->freed_extents[1],
 231                         start, end, EXTENT_UPTODATE);
 232         return 0;
 233 }
 234
 235 static void free_excluded_extents(struct btrfs_fs_info *fs_info,
 236                                   struct btrfs_block_group_cache *cache)
 237 {
 238         u64 start, end;
 239
 240         start = cache->key.objectid;
 241         end = start + cache->key.offset - 1;
 242
 243         clear_extent_bits(&fs_info->freed_extents[0],
 244                           start, end, EXTENT_UPTODATE);
 245         clear_extent_bits(&fs_info->freed_extents[1],
 246                           start, end, EXTENT_UPTODATE);
 247 }
 248
 249 static int exclude_super_stripes(struct btrfs_fs_info *fs_info,
 250                                  struct btrfs_block_group_cache *cache)
 251 {
 252         u64 bytenr;
 253         u64 *logical;
 254         int stripe_len;
 255         int i, nr, ret;
 256
 257         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 258                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 259                 cache->bytes_super += stripe_len;
 260                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 261                                           stripe_len);
 262                 if (ret)
 263                         return ret;
 264         }
 265
 266         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 267                 bytenr = btrfs_sb_offset(i);
 268                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 269                                        bytenr, 0, &logical, &nr, &stripe_len);
 270                 if (ret)
 271                         return ret;
 272
 273                 while (nr--) {
 274                         u64 start, len;
 275
 276                         if (logical[nr] > cache->key.objectid +
 277                             cache->key.offset)
 278                                 continue;
 279
 280                         if (logical[nr] + stripe_len <= cache->key.objectid)
 281                                 continue;
 282
 283                         start = logical[nr];
 284                         if (start < cache->key.objectid) {
 285                                 start = cache->key.objectid;
 286                                 len = (logical[nr] + stripe_len) - start;
 287                         } else {
 288                                 len = min_t(u64, stripe_len,
 289                                             cache->key.objectid +
 290                                             cache->key.offset - start);
 291                         }
 292
 293                         cache->bytes_super += len;
 294                         ret = add_excluded_extent(fs_info, start, len);
 295                         if (ret) {
 296                                 kfree(logical);
 297                                 return ret;
 298                         }
 299                 }
 300
 301                 kfree(logical);
 302         }
 303         return 0;
 304 }
 305
 306 static struct btrfs_caching_control *
 307 get_caching_control(struct btrfs_block_group_cache *cache)
 308 {
 309         struct btrfs_caching_control *ctl;
 310
 311         spin_lock(&cache->lock);
 312         if (!cache->caching_ctl) {
 313                 spin_unlock(&cache->lock);
 314                 return NULL;
 315         }
 316
 317         ctl = cache->caching_ctl;
 318         atomic_inc(&ctl->count);
 319         spin_unlock(&cache->lock);
 320         return ctl;
 321 }
 322
 323 static void put_caching_control(struct btrfs_caching_control *ctl)
 324 {
 325         if (atomic_dec_and_test(&ctl->count))
 326                 kfree(ctl);
 327 }
 328
 329 #ifdef CONFIG_BTRFS_DEBUG
 330 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 331 {
 332         struct btrfs_fs_info *fs_info = block_group->fs_info;
 333         u64 start = block_group->key.objectid;
 334         u64 len = block_group->key.offset;
 335         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 336                 fs_info->nodesize : fs_info->sectorsize;
 337         u64 step = chunk << 1;
 338
 339         while (len > chunk) {
 340                 btrfs_remove_free_space(block_group, start, chunk);
 341                 start += step;
 342                 if (len < step)
 343                         len = 0;
 344                 else
 345                         len -= step;
 346         }
 347 }
 348 #endif
 349
 350 /*
 351  * this is only called by cache_block_group, since we could have freed extents
 352  * we need to check the pinned_extents for any extents that can't be used yet
 353  * since their free space will be released as soon as the transaction commits.
 354  */
 355 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 356                        struct btrfs_fs_info *info, u64 start, u64 end)
 357 {
 358         u64 extent_start, extent_end, size, total_added = 0;
 359         int ret;
 360
 361         while (start < end) {
 362                 ret = find_first_extent_bit(info->pinned_extents, start,
 363                                             &extent_start, &extent_end,
 364                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 365                                             NULL);
 366                 if (ret)
 367                         break;
 368
 369                 if (extent_start <= start) {
 370                         start = extent_end + 1;
 371                 } else if (extent_start > start && extent_start < end) {
 372                         size = extent_start - start;
 373                         total_added += size;
 374                         ret = btrfs_add_free_space(block_group, start,
 375                                                    size);
 376                         BUG_ON(ret); /* -ENOMEM or logic error */
 377                         start = extent_end + 1;
 378                 } else {
 379                         break;
 380                 }
 381         }
 382
 383         if (start < end) {
 384                 size = end - start;
 385                 total_added += size;
 386                 ret = btrfs_add_free_space(block_group, start, size);
 387                 BUG_ON(ret); /* -ENOMEM or logic error */
 388         }
 389
 390         return total_added;
 391 }
 392
 393 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 394 {
 395         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 396         struct btrfs_fs_info *fs_info = block_group->fs_info;
 397         struct btrfs_root *extent_root = fs_info->extent_root;
 398         struct btrfs_path *path;
 399         struct extent_buffer *leaf;
 400         struct btrfs_key key;
 401         u64 total_found = 0;
 402         u64 last = 0;
 403         u32 nritems;
 404         int ret;
 405         bool wakeup = true;
 406
 407         path = btrfs_alloc_path();
 408         if (!path)
 409                 return -ENOMEM;
 410
 411         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 412
 413 #ifdef CONFIG_BTRFS_DEBUG
 414         /*
 415          * If we're fragmenting we don't want to make anybody think we can
 416          * allocate from this block group until we've had a chance to fragment
 417          * the free space.
 418          */
 419         if (btrfs_should_fragment_free_space(block_group))
 420                 wakeup = false;
 421 #endif
 422         /*
 423          * We don't want to deadlock with somebody trying to allocate a new
 424          * extent for the extent root while also trying to search the extent
 425          * root to add free space.  So we skip locking and search the commit
 426          * root, since its read-only
 427          */
 428         path->skip_locking = 1;
 429         path->search_commit_root = 1;
 430         path->reada = READA_FORWARD;
 431
 432         key.objectid = last;
 433         key.offset = 0;
 434         key.type = BTRFS_EXTENT_ITEM_KEY;
 435
 436 next:
 437         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 438         if (ret < 0)
 439                 goto out;
 440
 441         leaf = path->nodes[0];
 442         nritems = btrfs_header_nritems(leaf);
 443
 444         while (1) {
 445                 if (btrfs_fs_closing(fs_info) > 1) {
 446                         last = (u64)-1;
 447                         break;
 448                 }
 449
 450                 if (path->slots[0] < nritems) {
 451                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 452                 } else {
 453                         ret = find_next_key(path, 0, &key);
 454                         if (ret)
 455                                 break;
 456
 457                         if (need_resched() ||
 458                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 459                                 if (wakeup)
 460                                         caching_ctl->progress = last;
 461                                 btrfs_release_path(path);
 462                                 up_read(&fs_info->commit_root_sem);
 463                                 mutex_unlock(&caching_ctl->mutex);
 464                                 cond_resched();
 465                                 mutex_lock(&caching_ctl->mutex);
 466                                 down_read(&fs_info->commit_root_sem);
 467                                 goto next;
 468                         }
 469
 470                         ret = btrfs_next_leaf(extent_root, path);
 471                         if (ret < 0)
 472                                 goto out;
 473                         if (ret)
 474                                 break;
 475                         leaf = path->nodes[0];
 476                         nritems = btrfs_header_nritems(leaf);
 477                         continue;
 478                 }
 479
 480                 if (key.objectid < last) {
 481                         key.objectid = last;
 482                         key.offset = 0;
 483                         key.type = BTRFS_EXTENT_ITEM_KEY;
 484
 485                         if (wakeup)
 486                                 caching_ctl->progress = last;
 487                         btrfs_release_path(path);
 488                         goto next;
 489                 }
 490
 491                 if (key.objectid < block_group->key.objectid) {
 492                         path->slots[0]++;
 493                         continue;
 494                 }
 495
 496                 if (key.objectid >= block_group->key.objectid +
 497                     block_group->key.offset)
 498                         break;
 499
 500                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 501                     key.type == BTRFS_METADATA_ITEM_KEY) {
 502                         total_found += add_new_free_space(block_group,
 503                                                           fs_info, last,
 504                                                           key.objectid);
 505                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 506                                 last = key.objectid +
 507                                         fs_info->nodesize;
 508                         else
 509                                 last = key.objectid + key.offset;
 510
 511                         if (total_found > CACHING_CTL_WAKE_UP) {
 512                                 total_found = 0;
 513                                 if (wakeup)
 514                                         wake_up(&caching_ctl->wait);
 515                         }
 516                 }
 517                 path->slots[0]++;
 518         }
 519         ret = 0;
 520
 521         total_found += add_new_free_space(block_group, fs_info, last,
 522                                           block_group->key.objectid +
 523                                           block_group->key.offset);
 524         caching_ctl->progress = (u64)-1;
 525
 526 out:
 527         btrfs_free_path(path);
 528         return ret;
 529 }
 530
 531 static noinline void caching_thread(struct btrfs_work *work)
 532 {
 533         struct btrfs_block_group_cache *block_group;
 534         struct btrfs_fs_info *fs_info;
 535         struct btrfs_caching_control *caching_ctl;
 536         struct btrfs_root *extent_root;
 537         int ret;
 538
 539         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 540         block_group = caching_ctl->block_group;
 541         fs_info = block_group->fs_info;
 542         extent_root = fs_info->extent_root;
 543
 544         mutex_lock(&caching_ctl->mutex);
 545         down_read(&fs_info->commit_root_sem);
 546
 547         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 548                 ret = load_free_space_tree(caching_ctl);
 549         else
 550                 ret = load_extent_tree_free(caching_ctl);
 551
 552         spin_lock(&block_group->lock);
 553         block_group->caching_ctl = NULL;
 554         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 555         spin_unlock(&block_group->lock);
 556
 557 #ifdef CONFIG_BTRFS_DEBUG
 558         if (btrfs_should_fragment_free_space(block_group)) {
 559                 u64 bytes_used;
 560
 561                 spin_lock(&block_group->space_info->lock);
 562                 spin_lock(&block_group->lock);
 563                 bytes_used = block_group->key.offset -
 564                         btrfs_block_group_used(&block_group->item);
 565                 block_group->space_info->bytes_used += bytes_used >> 1;
 566                 spin_unlock(&block_group->lock);
 567                 spin_unlock(&block_group->space_info->lock);
 568                 fragment_free_space(block_group);
 569         }
 570 #endif
 571
 572         caching_ctl->progress = (u64)-1;
 573
 574         up_read(&fs_info->commit_root_sem);
 575         free_excluded_extents(fs_info, block_group);
 576         mutex_unlock(&caching_ctl->mutex);
 577
 578         wake_up(&caching_ctl->wait);
 579
 580         put_caching_control(caching_ctl);
 581         btrfs_put_block_group(block_group);
 582 }
 583
 584 static int cache_block_group(struct btrfs_block_group_cache *cache,
 585                              int load_cache_only)
 586 {
 587         DEFINE_WAIT(wait);
 588         struct btrfs_fs_info *fs_info = cache->fs_info;
 589         struct btrfs_caching_control *caching_ctl;
 590         int ret = 0;
 591
 592         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 593         if (!caching_ctl)
 594                 return -ENOMEM;
 595
 596         INIT_LIST_HEAD(&caching_ctl->list);
 597         mutex_init(&caching_ctl->mutex);
 598         init_waitqueue_head(&caching_ctl->wait);
 599         caching_ctl->block_group = cache;
 600         caching_ctl->progress = cache->key.objectid;
 601         atomic_set(&caching_ctl->count, 1);
 602         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 603                         caching_thread, NULL, NULL);
 604
 605         spin_lock(&cache->lock);
 606         /*
 607          * This should be a rare occasion, but this could happen I think in the
 608          * case where one thread starts to load the space cache info, and then
 609          * some other thread starts a transaction commit which tries to do an
 610          * allocation while the other thread is still loading the space cache
 611          * info.  The previous loop should have kept us from choosing this block
 612          * group, but if we've moved to the state where we will wait on caching
 613          * block groups we need to first check if we're doing a fast load here,
 614          * so we can wait for it to finish, otherwise we could end up allocating
 615          * from a block group who's cache gets evicted for one reason or
 616          * another.
 617          */
 618         while (cache->cached == BTRFS_CACHE_FAST) {
 619                 struct btrfs_caching_control *ctl;
 620
 621                 ctl = cache->caching_ctl;
 622                 atomic_inc(&ctl->count);
 623                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 624                 spin_unlock(&cache->lock);
 625
 626                 schedule();
 627
 628                 finish_wait(&ctl->wait, &wait);
 629                 put_caching_control(ctl);
 630                 spin_lock(&cache->lock);
 631         }
 632
 633         if (cache->cached != BTRFS_CACHE_NO) {
 634                 spin_unlock(&cache->lock);
 635                 kfree(caching_ctl);
 636                 return 0;
 637         }
 638         WARN_ON(cache->caching_ctl);
 639         cache->caching_ctl = caching_ctl;
 640         cache->cached = BTRFS_CACHE_FAST;
 641         spin_unlock(&cache->lock);
 642
 643         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 644                 mutex_lock(&caching_ctl->mutex);
 645                 ret = load_free_space_cache(fs_info, cache);
 646
 647                 spin_lock(&cache->lock);
 648                 if (ret == 1) {
 649                         cache->caching_ctl = NULL;
 650                         cache->cached = BTRFS_CACHE_FINISHED;
 651                         cache->last_byte_to_unpin = (u64)-1;
 652                         caching_ctl->progress = (u64)-1;
 653                 } else {
 654                         if (load_cache_only) {
 655                                 cache->caching_ctl = NULL;
 656                                 cache->cached = BTRFS_CACHE_NO;
 657                         } else {
 658                                 cache->cached = BTRFS_CACHE_STARTED;
 659                                 cache->has_caching_ctl = 1;
 660                         }
 661                 }
 662                 spin_unlock(&cache->lock);
 663 #ifdef CONFIG_BTRFS_DEBUG
 664                 if (ret == 1 &&
 665                     btrfs_should_fragment_free_space(cache)) {
 666                         u64 bytes_used;
 667
 668                         spin_lock(&cache->space_info->lock);
 669                         spin_lock(&cache->lock);
 670                         bytes_used = cache->key.offset -
 671                                 btrfs_block_group_used(&cache->item);
 672                         cache->space_info->bytes_used += bytes_used >> 1;
 673                         spin_unlock(&cache->lock);
 674                         spin_unlock(&cache->space_info->lock);
 675                         fragment_free_space(cache);
 676                 }
 677 #endif
 678                 mutex_unlock(&caching_ctl->mutex);
 679
 680                 wake_up(&caching_ctl->wait);
 681                 if (ret == 1) {
 682                         put_caching_control(caching_ctl);
 683                         free_excluded_extents(fs_info, cache);
 684                         return 0;
 685                 }
 686         } else {
 687                 /*
 688                  * We're either using the free space tree or no caching at all.
 689                  * Set cached to the appropriate value and wakeup any waiters.
 690                  */
 691                 spin_lock(&cache->lock);
 692                 if (load_cache_only) {
 693                         cache->caching_ctl = NULL;
 694                         cache->cached = BTRFS_CACHE_NO;
 695                 } else {
 696                         cache->cached = BTRFS_CACHE_STARTED;
 697                         cache->has_caching_ctl = 1;
 698                 }
 699                 spin_unlock(&cache->lock);
 700                 wake_up(&caching_ctl->wait);
 701         }
 702
 703         if (load_cache_only) {
 704                 put_caching_control(caching_ctl);
 705                 return 0;
 706         }
 707
 708         down_write(&fs_info->commit_root_sem);
 709         atomic_inc(&caching_ctl->count);
 710         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 711         up_write(&fs_info->commit_root_sem);
 712
 713         btrfs_get_block_group(cache);
 714
 715         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 716
 717         return ret;
 718 }
 719
 720 /*
 721  * return the block group that starts at or after bytenr
 722  */
 723 static struct btrfs_block_group_cache *
 724 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 725 {
 726         return block_group_cache_tree_search(info, bytenr, 0);
 727 }
 728
 729 /*
 730  * return the block group that contains the given bytenr
 731  */
 732 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 733                                                  struct btrfs_fs_info *info,
 734                                                  u64 bytenr)
 735 {
 736         return block_group_cache_tree_search(info, bytenr, 1);
 737 }
 738
 739 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 740                                                   u64 flags)
 741 {
 742         struct list_head *head = &info->space_info;
 743         struct btrfs_space_info *found;
 744
 745         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 746
 747         rcu_read_lock();
 748         list_for_each_entry_rcu(found, head, list) {
 749                 if (found->flags & flags) {
 750                         rcu_read_unlock();
 751                         return found;
 752                 }
 753         }
 754         rcu_read_unlock();
 755         return NULL;
 756 }
 757
 758 /*
 759  * after adding space to the filesystem, we need to clear the full flags
 760  * on all the space infos.
 761  */
 762 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 763 {
 764         struct list_head *head = &info->space_info;
 765         struct btrfs_space_info *found;
 766
 767         rcu_read_lock();
 768         list_for_each_entry_rcu(found, head, list)
 769                 found->full = 0;
 770         rcu_read_unlock();
 771 }
 772
 773 /* simple helper to search for an existing data extent at a given offset */
 774 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 775 {
 776         int ret;
 777         struct btrfs_key key;
 778         struct btrfs_path *path;
 779
 780         path = btrfs_alloc_path();
 781         if (!path)
 782                 return -ENOMEM;
 783
 784         key.objectid = start;
 785         key.offset = len;
 786         key.type = BTRFS_EXTENT_ITEM_KEY;
 787         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 788         btrfs_free_path(path);
 789         return ret;
 790 }
 791
 792 /*
 793  * helper function to lookup reference count and flags of a tree block.
 794  *
 795  * the head node for delayed ref is used to store the sum of all the
 796  * reference count modifications queued up in the rbtree. the head
 797  * node may also store the extent flags to set. This way you can check
 798  * to see what the reference count and extent flags would be if all of
 799  * the delayed refs are not processed.
 800  */
 801 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 802                              struct btrfs_fs_info *fs_info, u64 bytenr,
 803                              u64 offset, int metadata, u64 *refs, u64 *flags)
 804 {
 805         struct btrfs_delayed_ref_head *head;
 806         struct btrfs_delayed_ref_root *delayed_refs;
 807         struct btrfs_path *path;
 808         struct btrfs_extent_item *ei;
 809         struct extent_buffer *leaf;
 810         struct btrfs_key key;
 811         u32 item_size;
 812         u64 num_refs;
 813         u64 extent_flags;
 814         int ret;
 815
 816         /*
 817          * If we don't have skinny metadata, don't bother doing anything
 818          * different
 819          */
 820         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 821                 offset = fs_info->nodesize;
 822                 metadata = 0;
 823         }
 824
 825         path = btrfs_alloc_path();
 826         if (!path)
 827                 return -ENOMEM;
 828
 829         if (!trans) {
 830                 path->skip_locking = 1;
 831                 path->search_commit_root = 1;
 832         }
 833
 834 search_again:
 835         key.objectid = bytenr;
 836         key.offset = offset;
 837         if (metadata)
 838                 key.type = BTRFS_METADATA_ITEM_KEY;
 839         else
 840                 key.type = BTRFS_EXTENT_ITEM_KEY;
 841
 842         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 843         if (ret < 0)
 844                 goto out_free;
 845
 846         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 847                 if (path->slots[0]) {
 848                         path->slots[0]--;
 849                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 850                                               path->slots[0]);
 851                         if (key.objectid == bytenr &&
 852                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 853                             key.offset == fs_info->nodesize)
 854                                 ret = 0;
 855                 }
 856         }
 857
 858         if (ret == 0) {
 859                 leaf = path->nodes[0];
 860                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 861                 if (item_size >= sizeof(*ei)) {
 862                         ei = btrfs_item_ptr(leaf, path->slots[0],
 863                                             struct btrfs_extent_item);
 864                         num_refs = btrfs_extent_refs(leaf, ei);
 865                         extent_flags = btrfs_extent_flags(leaf, ei);
 866                 } else {
 867 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 868                         struct btrfs_extent_item_v0 *ei0;
 869                         BUG_ON(item_size != sizeof(*ei0));
 870                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 871                                              struct btrfs_extent_item_v0);
 872                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 873                         /* FIXME: this isn't correct for data */
 874                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 875 #else
 876                         BUG();
 877 #endif
 878                 }
 879                 BUG_ON(num_refs == 0);
 880         } else {
 881                 num_refs = 0;
 882                 extent_flags = 0;
 883                 ret = 0;
 884         }
 885
 886         if (!trans)
 887                 goto out;
 888
 889         delayed_refs = &trans->transaction->delayed_refs;
 890         spin_lock(&delayed_refs->lock);
 891         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 892         if (head) {
 893                 if (!mutex_trylock(&head->mutex)) {
 894                         atomic_inc(&head->node.refs);
 895                         spin_unlock(&delayed_refs->lock);
 896
 897                         btrfs_release_path(path);
 898
 899                         /*
 900                          * Mutex was contended, block until it's released and try
 901                          * again
 902                          */
 903                         mutex_lock(&head->mutex);
 904                         mutex_unlock(&head->mutex);
 905                         btrfs_put_delayed_ref(&head->node);
 906                         goto search_again;
 907                 }
 908                 spin_lock(&head->lock);
 909                 if (head->extent_op && head->extent_op->update_flags)
 910                         extent_flags |= head->extent_op->flags_to_set;
 911                 else
 912                         BUG_ON(num_refs == 0);
 913
 914                 num_refs += head->node.ref_mod;
 915                 spin_unlock(&head->lock);
 916                 mutex_unlock(&head->mutex);
 917         }
 918         spin_unlock(&delayed_refs->lock);
 919 out:
 920         WARN_ON(num_refs == 0);
 921         if (refs)
 922                 *refs = num_refs;
 923         if (flags)
 924                 *flags = extent_flags;
 925 out_free:
 926         btrfs_free_path(path);
 927         return ret;
 928 }
 929
 930 /*
 931  * Back reference rules.  Back refs have three main goals:
 932  *
 933  * 1) differentiate between all holders of references to an extent so that
 934  *    when a reference is dropped we can make sure it was a valid reference
 935  *    before freeing the extent.
 936  *
 937  * 2) Provide enough information to quickly find the holders of an extent
 938  *    if we notice a given block is corrupted or bad.
 939  *
 940  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 941  *    maintenance.  This is actually the same as #2, but with a slightly
 942  *    different use case.
 943  *
 944  * There are two kinds of back refs. The implicit back refs is optimized
 945  * for pointers in non-shared tree blocks. For a given pointer in a block,
 946  * back refs of this kind provide information about the block's owner tree
 947  * and the pointer's key. These information allow us to find the block by
 948  * b-tree searching. The full back refs is for pointers in tree blocks not
 949  * referenced by their owner trees. The location of tree block is recorded
 950  * in the back refs. Actually the full back refs is generic, and can be
 951  * used in all cases the implicit back refs is used. The major shortcoming
 952  * of the full back refs is its overhead. Every time a tree block gets
 953  * COWed, we have to update back refs entry for all pointers in it.
 954  *
 955  * For a newly allocated tree block, we use implicit back refs for
 956  * pointers in it. This means most tree related operations only involve
 957  * implicit back refs. For a tree block created in old transaction, the
 958  * only way to drop a reference to it is COW it. So we can detect the
 959  * event that tree block loses its owner tree's reference and do the
 960  * back refs conversion.
 961  *
 962  * When a tree block is COWed through a tree, there are four cases:
 963  *
 964  * The reference count of the block is one and the tree is the block's
 965  * owner tree. Nothing to do in this case.
 966  *
 967  * The reference count of the block is one and the tree is not the
 968  * block's owner tree. In this case, full back refs is used for pointers
 969  * in the block. Remove these full back refs, add implicit back refs for
 970  * every pointers in the new block.
 971  *
 972  * The reference count of the block is greater than one and the tree is
 973  * the block's owner tree. In this case, implicit back refs is used for
 974  * pointers in the block. Add full back refs for every pointers in the
 975  * block, increase lower level extents' reference counts. The original
 976  * implicit back refs are entailed to the new block.
 977  *
 978  * The reference count of the block is greater than one and the tree is
 979  * not the block's owner tree. Add implicit back refs for every pointer in
 980  * the new block, increase lower level extents' reference count.
 981  *
 982  * Back Reference Key composing:
 983  *
 984  * The key objectid corresponds to the first byte in the extent,
 985  * The key type is used to differentiate between types of back refs.
 986  * There are different meanings of the key offset for different types
 987  * of back refs.
 988  *
 989  * File extents can be referenced by:
 990  *
 991  * - multiple snapshots, subvolumes, or different generations in one subvol
 992  * - different files inside a single subvolume
 993  * - different offsets inside a file (bookend extents in file.c)
 994  *
 995  * The extent ref structure for the implicit back refs has fields for:
 996  *
 997  * - Objectid of the subvolume root
 998  * - objectid of the file holding the reference
 999  * - original offset in the file
1000  * - how many bookend extents
1001  *
1002  * The key offset for the implicit back refs is hash of the first
1003  * three fields.
1004  *
1005  * The extent ref structure for the full back refs has field for:
1006  *
1007  * - number of pointers in the tree leaf
1008  *
1009  * The key offset for the implicit back refs is the first byte of
1010  * the tree leaf
1011  *
1012  * When a file extent is allocated, The implicit back refs is used.
1013  * the fields are filled in:
1014  *
1015  *     (root_key.objectid, inode objectid, offset in file, 1)
1016  *
1017  * When a file extent is removed file truncation, we find the
1018  * corresponding implicit back refs and check the following fields:
1019  *
1020  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1021  *
1022  * Btree extents can be referenced by:
1023  *
1024  * - Different subvolumes
1025  *
1026  * Both the implicit back refs and the full back refs for tree blocks
1027  * only consist of key. The key offset for the implicit back refs is
1028  * objectid of block's owner tree. The key offset for the full back refs
1029  * is the first byte of parent block.
1030  *
1031  * When implicit back refs is used, information about the lowest key and
1032  * level of the tree block are required. These information are stored in
1033  * tree block info structure.
1034  */
1035
1036 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1037 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1038                                   struct btrfs_fs_info *fs_info,
1039                                   struct btrfs_path *path,
1040                                   u64 owner, u32 extra_size)
1041 {
1042         struct btrfs_root *root = fs_info->extent_root;
1043         struct btrfs_extent_item *item;
1044         struct btrfs_extent_item_v0 *ei0;
1045         struct btrfs_extent_ref_v0 *ref0;
1046         struct btrfs_tree_block_info *bi;
1047         struct extent_buffer *leaf;
1048         struct btrfs_key key;
1049         struct btrfs_key found_key;
1050         u32 new_size = sizeof(*item);
1051         u64 refs;
1052         int ret;
1053
1054         leaf = path->nodes[0];
1055         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1056
1057         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1058         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1059                              struct btrfs_extent_item_v0);
1060         refs = btrfs_extent_refs_v0(leaf, ei0);
1061
1062         if (owner == (u64)-1) {
1063                 while (1) {
1064                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1065                                 ret = btrfs_next_leaf(root, path);
1066                                 if (ret < 0)
1067                                         return ret;
1068                                 BUG_ON(ret > 0); /* Corruption */
1069                                 leaf = path->nodes[0];
1070                         }
1071                         btrfs_item_key_to_cpu(leaf, &found_key,
1072                                               path->slots[0]);
1073                         BUG_ON(key.objectid != found_key.objectid);
1074                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1075                                 path->slots[0]++;
1076                                 continue;
1077                         }
1078                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1079                                               struct btrfs_extent_ref_v0);
1080                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1081                         break;
1082                 }
1083         }
1084         btrfs_release_path(path);
1085
1086         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1087                 new_size += sizeof(*bi);
1088
1089         new_size -= sizeof(*ei0);
1090         ret = btrfs_search_slot(trans, root, &key, path,
1091                                 new_size + extra_size, 1);
1092         if (ret < 0)
1093                 return ret;
1094         BUG_ON(ret); /* Corruption */
1095
1096         btrfs_extend_item(fs_info, path, new_size);
1097
1098         leaf = path->nodes[0];
1099         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1100         btrfs_set_extent_refs(leaf, item, refs);
1101         /* FIXME: get real generation */
1102         btrfs_set_extent_generation(leaf, item, 0);
1103         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1104                 btrfs_set_extent_flags(leaf, item,
1105                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1106                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1107                 bi = (struct btrfs_tree_block_info *)(item + 1);
1108                 /* FIXME: get first key of the block */
1109                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1110                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1111         } else {
1112                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1113         }
1114         btrfs_mark_buffer_dirty(leaf);
1115         return 0;
1116 }
1117 #endif
1118
1119 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1120 {
1121         u32 high_crc = ~(u32)0;
1122         u32 low_crc = ~(u32)0;
1123         __le64 lenum;
1124
1125         lenum = cpu_to_le64(root_objectid);
1126         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1127         lenum = cpu_to_le64(owner);
1128         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1129         lenum = cpu_to_le64(offset);
1130         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1131
1132         return ((u64)high_crc << 31) ^ (u64)low_crc;
1133 }
1134
1135 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1136                                      struct btrfs_extent_data_ref *ref)
1137 {
1138         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1139                                     btrfs_extent_data_ref_objectid(leaf, ref),
1140                                     btrfs_extent_data_ref_offset(leaf, ref));
1141 }
1142
1143 static int match_extent_data_ref(struct extent_buffer *leaf,
1144                                  struct btrfs_extent_data_ref *ref,
1145                                  u64 root_objectid, u64 owner, u64 offset)
1146 {
1147         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1148             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1149             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1150                 return 0;
1151         return 1;
1152 }
1153
1154 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1155                                            struct btrfs_fs_info *fs_info,
1156                                            struct btrfs_path *path,
1157                                            u64 bytenr, u64 parent,
1158                                            u64 root_objectid,
1159                                            u64 owner, u64 offset)
1160 {
1161         struct btrfs_root *root = fs_info->extent_root;
1162         struct btrfs_key key;
1163         struct btrfs_extent_data_ref *ref;
1164         struct extent_buffer *leaf;
1165         u32 nritems;
1166         int ret;
1167         int recow;
1168         int err = -ENOENT;
1169
1170         key.objectid = bytenr;
1171         if (parent) {
1172                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1173                 key.offset = parent;
1174         } else {
1175                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1176                 key.offset = hash_extent_data_ref(root_objectid,
1177                                                   owner, offset);
1178         }
1179 again:
1180         recow = 0;
1181         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1182         if (ret < 0) {
1183                 err = ret;
1184                 goto fail;
1185         }
1186
1187         if (parent) {
1188                 if (!ret)
1189                         return 0;
1190 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1191                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1192                 btrfs_release_path(path);
1193                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1194                 if (ret < 0) {
1195                         err = ret;
1196                         goto fail;
1197                 }
1198                 if (!ret)
1199                         return 0;
1200 #endif
1201                 goto fail;
1202         }
1203
1204         leaf = path->nodes[0];
1205         nritems = btrfs_header_nritems(leaf);
1206         while (1) {
1207                 if (path->slots[0] >= nritems) {
1208                         ret = btrfs_next_leaf(root, path);
1209                         if (ret < 0)
1210                                 err = ret;
1211                         if (ret)
1212                                 goto fail;
1213
1214                         leaf = path->nodes[0];
1215                         nritems = btrfs_header_nritems(leaf);
1216                         recow = 1;
1217                 }
1218
1219                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1220                 if (key.objectid != bytenr ||
1221                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1222                         goto fail;
1223
1224                 ref = btrfs_item_ptr(leaf, path->slots[0],
1225                                      struct btrfs_extent_data_ref);
1226
1227                 if (match_extent_data_ref(leaf, ref, root_objectid,
1228                                           owner, offset)) {
1229                         if (recow) {
1230                                 btrfs_release_path(path);
1231                                 goto again;
1232                         }
1233                         err = 0;
1234                         break;
1235                 }
1236                 path->slots[0]++;
1237         }
1238 fail:
1239         return err;
1240 }
1241
1242 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1243                                            struct btrfs_fs_info *fs_info,
1244                                            struct btrfs_path *path,
1245                                            u64 bytenr, u64 parent,
1246                                            u64 root_objectid, u64 owner,
1247                                            u64 offset, int refs_to_add)
1248 {
1249         struct btrfs_root *root = fs_info->extent_root;
1250         struct btrfs_key key;
1251         struct extent_buffer *leaf;
1252         u32 size;
1253         u32 num_refs;
1254         int ret;
1255
1256         key.objectid = bytenr;
1257         if (parent) {
1258                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1259                 key.offset = parent;
1260                 size = sizeof(struct btrfs_shared_data_ref);
1261         } else {
1262                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1263                 key.offset = hash_extent_data_ref(root_objectid,
1264                                                   owner, offset);
1265                 size = sizeof(struct btrfs_extent_data_ref);
1266         }
1267
1268         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1269         if (ret && ret != -EEXIST)
1270                 goto fail;
1271
1272         leaf = path->nodes[0];
1273         if (parent) {
1274                 struct btrfs_shared_data_ref *ref;
1275                 ref = btrfs_item_ptr(leaf, path->slots[0],
1276                                      struct btrfs_shared_data_ref);
1277                 if (ret == 0) {
1278                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1279                 } else {
1280                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1281                         num_refs += refs_to_add;
1282                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1283                 }
1284         } else {
1285                 struct btrfs_extent_data_ref *ref;
1286                 while (ret == -EEXIST) {
1287                         ref = btrfs_item_ptr(leaf, path->slots[0],
1288                                              struct btrfs_extent_data_ref);
1289                         if (match_extent_data_ref(leaf, ref, root_objectid,
1290                                                   owner, offset))
1291                                 break;
1292                         btrfs_release_path(path);
1293                         key.offset++;
1294                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1295                                                       size);
1296                         if (ret && ret != -EEXIST)
1297                                 goto fail;
1298
1299                         leaf = path->nodes[0];
1300                 }
1301                 ref = btrfs_item_ptr(leaf, path->slots[0],
1302                                      struct btrfs_extent_data_ref);
1303                 if (ret == 0) {
1304                         btrfs_set_extent_data_ref_root(leaf, ref,
1305                                                        root_objectid);
1306                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1307                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1308                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1309                 } else {
1310                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1311                         num_refs += refs_to_add;
1312                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1313                 }
1314         }
1315         btrfs_mark_buffer_dirty(leaf);
1316         ret = 0;
1317 fail:
1318         btrfs_release_path(path);
1319         return ret;
1320 }
1321
1322 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1323                                            struct btrfs_fs_info *fs_info,
1324                                            struct btrfs_path *path,
1325                                            int refs_to_drop, int *last_ref)
1326 {
1327         struct btrfs_key key;
1328         struct btrfs_extent_data_ref *ref1 = NULL;
1329         struct btrfs_shared_data_ref *ref2 = NULL;
1330         struct extent_buffer *leaf;
1331         u32 num_refs = 0;
1332         int ret = 0;
1333
1334         leaf = path->nodes[0];
1335         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1336
1337         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1338                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1339                                       struct btrfs_extent_data_ref);
1340                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1341         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1342                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1343                                       struct btrfs_shared_data_ref);
1344                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1345 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1346         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1347                 struct btrfs_extent_ref_v0 *ref0;
1348                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1349                                       struct btrfs_extent_ref_v0);
1350                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1351 #endif
1352         } else {
1353                 BUG();
1354         }
1355
1356         BUG_ON(num_refs < refs_to_drop);
1357         num_refs -= refs_to_drop;
1358
1359         if (num_refs == 0) {
1360                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1361                 *last_ref = 1;
1362         } else {
1363                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1364                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1365                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1366                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1367 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1368                 else {
1369                         struct btrfs_extent_ref_v0 *ref0;
1370                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1371                                         struct btrfs_extent_ref_v0);
1372                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1373                 }
1374 #endif
1375                 btrfs_mark_buffer_dirty(leaf);
1376         }
1377         return ret;
1378 }
1379
1380 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1381                                           struct btrfs_extent_inline_ref *iref)
1382 {
1383         struct btrfs_key key;
1384         struct extent_buffer *leaf;
1385         struct btrfs_extent_data_ref *ref1;
1386         struct btrfs_shared_data_ref *ref2;
1387         u32 num_refs = 0;
1388
1389         leaf = path->nodes[0];
1390         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1391         if (iref) {
1392                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1393                     BTRFS_EXTENT_DATA_REF_KEY) {
1394                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1395                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1396                 } else {
1397                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1398                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1399                 }
1400         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1401                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1402                                       struct btrfs_extent_data_ref);
1403                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1404         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1405                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1406                                       struct btrfs_shared_data_ref);
1407                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1408 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1409         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1410                 struct btrfs_extent_ref_v0 *ref0;
1411                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1412                                       struct btrfs_extent_ref_v0);
1413                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1414 #endif
1415         } else {
1416                 WARN_ON(1);
1417         }
1418         return num_refs;
1419 }
1420
1421 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1422                                           struct btrfs_fs_info *fs_info,
1423                                           struct btrfs_path *path,
1424                                           u64 bytenr, u64 parent,
1425                                           u64 root_objectid)
1426 {
1427         struct btrfs_root *root = fs_info->extent_root;
1428         struct btrfs_key key;
1429         int ret;
1430
1431         key.objectid = bytenr;
1432         if (parent) {
1433                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1434                 key.offset = parent;
1435         } else {
1436                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1437                 key.offset = root_objectid;
1438         }
1439
1440         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1441         if (ret > 0)
1442                 ret = -ENOENT;
1443 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1444         if (ret == -ENOENT && parent) {
1445                 btrfs_release_path(path);
1446                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1447                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1448                 if (ret > 0)
1449                         ret = -ENOENT;
1450         }
1451 #endif
1452         return ret;
1453 }
1454
1455 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1456                                           struct btrfs_fs_info *fs_info,
1457                                           struct btrfs_path *path,
1458                                           u64 bytenr, u64 parent,
1459                                           u64 root_objectid)
1460 {
1461         struct btrfs_key key;
1462         int ret;
1463
1464         key.objectid = bytenr;
1465         if (parent) {
1466                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1467                 key.offset = parent;
1468         } else {
1469                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1470                 key.offset = root_objectid;
1471         }
1472
1473         ret = btrfs_insert_empty_item(trans, fs_info->extent_root,
1474                                       path, &key, 0);
1475         btrfs_release_path(path);
1476         return ret;
1477 }
1478
1479 static inline int extent_ref_type(u64 parent, u64 owner)
1480 {
1481         int type;
1482         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1483                 if (parent > 0)
1484                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1485                 else
1486                         type = BTRFS_TREE_BLOCK_REF_KEY;
1487         } else {
1488                 if (parent > 0)
1489                         type = BTRFS_SHARED_DATA_REF_KEY;
1490                 else
1491                         type = BTRFS_EXTENT_DATA_REF_KEY;
1492         }
1493         return type;
1494 }
1495
1496 static int find_next_key(struct btrfs_path *path, int level,
1497                          struct btrfs_key *key)
1498
1499 {
1500         for (; level < BTRFS_MAX_LEVEL; level++) {
1501                 if (!path->nodes[level])
1502                         break;
1503                 if (path->slots[level] + 1 >=
1504                     btrfs_header_nritems(path->nodes[level]))
1505                         continue;
1506                 if (level == 0)
1507                         btrfs_item_key_to_cpu(path->nodes[level], key,
1508                                               path->slots[level] + 1);
1509                 else
1510                         btrfs_node_key_to_cpu(path->nodes[level], key,
1511                                               path->slots[level] + 1);
1512                 return 0;
1513         }
1514         return 1;
1515 }
1516
1517 /*
1518  * look for inline back ref. if back ref is found, *ref_ret is set
1519  * to the address of inline back ref, and 0 is returned.
1520  *
1521  * if back ref isn't found, *ref_ret is set to the address where it
1522  * should be inserted, and -ENOENT is returned.
1523  *
1524  * if insert is true and there are too many inline back refs, the path
1525  * points to the extent item, and -EAGAIN is returned.
1526  *
1527  * NOTE: inline back refs are ordered in the same way that back ref
1528  *       items in the tree are ordered.
1529  */
1530 static noinline_for_stack
1531 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1532                                  struct btrfs_fs_info *fs_info,
1533                                  struct btrfs_path *path,
1534                                  struct btrfs_extent_inline_ref **ref_ret,
1535                                  u64 bytenr, u64 num_bytes,
1536                                  u64 parent, u64 root_objectid,
1537                                  u64 owner, u64 offset, int insert)
1538 {
1539         struct btrfs_root *root = fs_info->extent_root;
1540         struct btrfs_key key;
1541         struct extent_buffer *leaf;
1542         struct btrfs_extent_item *ei;
1543         struct btrfs_extent_inline_ref *iref;
1544         u64 flags;
1545         u64 item_size;
1546         unsigned long ptr;
1547         unsigned long end;
1548         int extra_size;
1549         int type;
1550         int want;
1551         int ret;
1552         int err = 0;
1553         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1554
1555         key.objectid = bytenr;
1556         key.type = BTRFS_EXTENT_ITEM_KEY;
1557         key.offset = num_bytes;
1558
1559         want = extent_ref_type(parent, owner);
1560         if (insert) {
1561                 extra_size = btrfs_extent_inline_ref_size(want);
1562                 path->keep_locks = 1;
1563         } else
1564                 extra_size = -1;
1565
1566         /*
1567          * Owner is our parent level, so we can just add one to get the level
1568          * for the block we are interested in.
1569          */
1570         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1571                 key.type = BTRFS_METADATA_ITEM_KEY;
1572                 key.offset = owner;
1573         }
1574
1575 again:
1576         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1577         if (ret < 0) {
1578                 err = ret;
1579                 goto out;
1580         }
1581
1582         /*
1583          * We may be a newly converted file system which still has the old fat
1584          * extent entries for metadata, so try and see if we have one of those.
1585          */
1586         if (ret > 0 && skinny_metadata) {
1587                 skinny_metadata = false;
1588                 if (path->slots[0]) {
1589                         path->slots[0]--;
1590                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1591                                               path->slots[0]);
1592                         if (key.objectid == bytenr &&
1593                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1594                             key.offset == num_bytes)
1595                                 ret = 0;
1596                 }
1597                 if (ret) {
1598                         key.objectid = bytenr;
1599                         key.type = BTRFS_EXTENT_ITEM_KEY;
1600                         key.offset = num_bytes;
1601                         btrfs_release_path(path);
1602                         goto again;
1603                 }
1604         }
1605
1606         if (ret && !insert) {
1607                 err = -ENOENT;
1608                 goto out;
1609         } else if (WARN_ON(ret)) {
1610                 err = -EIO;
1611                 goto out;
1612         }
1613
1614         leaf = path->nodes[0];
1615         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1616 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1617         if (item_size < sizeof(*ei)) {
1618                 if (!insert) {
1619                         err = -ENOENT;
1620                         goto out;
1621                 }
1622                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1623                                              extra_size);
1624                 if (ret < 0) {
1625                         err = ret;
1626                         goto out;
1627                 }
1628                 leaf = path->nodes[0];
1629                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1630         }
1631 #endif
1632         BUG_ON(item_size < sizeof(*ei));
1633
1634         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1635         flags = btrfs_extent_flags(leaf, ei);
1636
1637         ptr = (unsigned long)(ei + 1);
1638         end = (unsigned long)ei + item_size;
1639
1640         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1641                 ptr += sizeof(struct btrfs_tree_block_info);
1642                 BUG_ON(ptr > end);
1643         }
1644
1645         err = -ENOENT;
1646         while (1) {
1647                 if (ptr >= end) {
1648                         WARN_ON(ptr > end);
1649                         break;
1650                 }
1651                 iref = (struct btrfs_extent_inline_ref *)ptr;
1652                 type = btrfs_extent_inline_ref_type(leaf, iref);
1653                 if (want < type)
1654                         break;
1655                 if (want > type) {
1656                         ptr += btrfs_extent_inline_ref_size(type);
1657                         continue;
1658                 }
1659
1660                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1661                         struct btrfs_extent_data_ref *dref;
1662                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1663                         if (match_extent_data_ref(leaf, dref, root_objectid,
1664                                                   owner, offset)) {
1665                                 err = 0;
1666                                 break;
1667                         }
1668                         if (hash_extent_data_ref_item(leaf, dref) <
1669                             hash_extent_data_ref(root_objectid, owner, offset))
1670                                 break;
1671                 } else {
1672                         u64 ref_offset;
1673                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1674                         if (parent > 0) {
1675                                 if (parent == ref_offset) {
1676                                         err = 0;
1677                                         break;
1678                                 }
1679                                 if (ref_offset < parent)
1680                                         break;
1681                         } else {
1682                                 if (root_objectid == ref_offset) {
1683                                         err = 0;
1684                                         break;
1685                                 }
1686                                 if (ref_offset < root_objectid)
1687                                         break;
1688                         }
1689                 }
1690                 ptr += btrfs_extent_inline_ref_size(type);
1691         }
1692         if (err == -ENOENT && insert) {
1693                 if (item_size + extra_size >=
1694                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1695                         err = -EAGAIN;
1696                         goto out;
1697                 }
1698                 /*
1699                  * To add new inline back ref, we have to make sure
1700                  * there is no corresponding back ref item.
1701                  * For simplicity, we just do not add new inline back
1702                  * ref if there is any kind of item for this block
1703                  */
1704                 if (find_next_key(path, 0, &key) == 0 &&
1705                     key.objectid == bytenr &&
1706                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1707                         err = -EAGAIN;
1708                         goto out;
1709                 }
1710         }
1711         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1712 out:
1713         if (insert) {
1714                 path->keep_locks = 0;
1715                 btrfs_unlock_up_safe(path, 1);
1716         }
1717         return err;
1718 }
1719
1720 /*
1721  * helper to add new inline back ref
1722  */
1723 static noinline_for_stack
1724 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1725                                  struct btrfs_path *path,
1726                                  struct btrfs_extent_inline_ref *iref,
1727                                  u64 parent, u64 root_objectid,
1728                                  u64 owner, u64 offset, int refs_to_add,
1729                                  struct btrfs_delayed_extent_op *extent_op)
1730 {
1731         struct extent_buffer *leaf;
1732         struct btrfs_extent_item *ei;
1733         unsigned long ptr;
1734         unsigned long end;
1735         unsigned long item_offset;
1736         u64 refs;
1737         int size;
1738         int type;
1739
1740         leaf = path->nodes[0];
1741         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1742         item_offset = (unsigned long)iref - (unsigned long)ei;
1743
1744         type = extent_ref_type(parent, owner);
1745         size = btrfs_extent_inline_ref_size(type);
1746
1747         btrfs_extend_item(fs_info, path, size);
1748
1749         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1750         refs = btrfs_extent_refs(leaf, ei);
1751         refs += refs_to_add;
1752         btrfs_set_extent_refs(leaf, ei, refs);
1753         if (extent_op)
1754                 __run_delayed_extent_op(extent_op, leaf, ei);
1755
1756         ptr = (unsigned long)ei + item_offset;
1757         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1758         if (ptr < end - size)
1759                 memmove_extent_buffer(leaf, ptr + size, ptr,
1760                                       end - size - ptr);
1761
1762         iref = (struct btrfs_extent_inline_ref *)ptr;
1763         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1764         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1765                 struct btrfs_extent_data_ref *dref;
1766                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1767                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1768                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1769                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1770                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1771         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1772                 struct btrfs_shared_data_ref *sref;
1773                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1774                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1775                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1776         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1777                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1778         } else {
1779                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1780         }
1781         btrfs_mark_buffer_dirty(leaf);
1782 }
1783
1784 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1785                                  struct btrfs_fs_info *fs_info,
1786                                  struct btrfs_path *path,
1787                                  struct btrfs_extent_inline_ref **ref_ret,
1788                                  u64 bytenr, u64 num_bytes, u64 parent,
1789                                  u64 root_objectid, u64 owner, u64 offset)
1790 {
1791         int ret;
1792
1793         ret = lookup_inline_extent_backref(trans, fs_info, path, ref_ret,
1794                                            bytenr, num_bytes, parent,
1795                                            root_objectid, owner, offset, 0);
1796         if (ret != -ENOENT)
1797                 return ret;
1798
1799         btrfs_release_path(path);
1800         *ref_ret = NULL;
1801
1802         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1803                 ret = lookup_tree_block_ref(trans, fs_info, path, bytenr,
1804                                             parent, root_objectid);
1805         } else {
1806                 ret = lookup_extent_data_ref(trans, fs_info, path, bytenr,
1807                                              parent, root_objectid, owner,
1808                                              offset);
1809         }
1810         return ret;
1811 }
1812
1813 /*
1814  * helper to update/remove inline back ref
1815  */
1816 static noinline_for_stack
1817 void update_inline_extent_backref(struct btrfs_fs_info *fs_info,
1818                                   struct btrfs_path *path,
1819                                   struct btrfs_extent_inline_ref *iref,
1820                                   int refs_to_mod,
1821                                   struct btrfs_delayed_extent_op *extent_op,
1822                                   int *last_ref)
1823 {
1824         struct extent_buffer *leaf;
1825         struct btrfs_extent_item *ei;
1826         struct btrfs_extent_data_ref *dref = NULL;
1827         struct btrfs_shared_data_ref *sref = NULL;
1828         unsigned long ptr;
1829         unsigned long end;
1830         u32 item_size;
1831         int size;
1832         int type;
1833         u64 refs;
1834
1835         leaf = path->nodes[0];
1836         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1837         refs = btrfs_extent_refs(leaf, ei);
1838         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1839         refs += refs_to_mod;
1840         btrfs_set_extent_refs(leaf, ei, refs);
1841         if (extent_op)
1842                 __run_delayed_extent_op(extent_op, leaf, ei);
1843
1844         type = btrfs_extent_inline_ref_type(leaf, iref);
1845
1846         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1847                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1848                 refs = btrfs_extent_data_ref_count(leaf, dref);
1849         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1850                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1851                 refs = btrfs_shared_data_ref_count(leaf, sref);
1852         } else {
1853                 refs = 1;
1854                 BUG_ON(refs_to_mod != -1);
1855         }
1856
1857         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1858         refs += refs_to_mod;
1859
1860         if (refs > 0) {
1861                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1862                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1863                 else
1864                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1865         } else {
1866                 *last_ref = 1;
1867                 size =  btrfs_extent_inline_ref_size(type);
1868                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1869                 ptr = (unsigned long)iref;
1870                 end = (unsigned long)ei + item_size;
1871                 if (ptr + size < end)
1872                         memmove_extent_buffer(leaf, ptr, ptr + size,
1873                                               end - ptr - size);
1874                 item_size -= size;
1875                 btrfs_truncate_item(fs_info, path, item_size, 1);
1876         }
1877         btrfs_mark_buffer_dirty(leaf);
1878 }
1879
1880 static noinline_for_stack
1881 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1882                                  struct btrfs_fs_info *fs_info,
1883                                  struct btrfs_path *path,
1884                                  u64 bytenr, u64 num_bytes, u64 parent,
1885                                  u64 root_objectid, u64 owner,
1886                                  u64 offset, int refs_to_add,
1887                                  struct btrfs_delayed_extent_op *extent_op)
1888 {
1889         struct btrfs_extent_inline_ref *iref;
1890         int ret;
1891
1892         ret = lookup_inline_extent_backref(trans, fs_info, path, &iref,
1893                                            bytenr, num_bytes, parent,
1894                                            root_objectid, owner, offset, 1);
1895         if (ret == 0) {
1896                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1897                 update_inline_extent_backref(fs_info, path, iref,
1898                                              refs_to_add, extent_op, NULL);
1899         } else if (ret == -ENOENT) {
1900                 setup_inline_extent_backref(fs_info, path, iref, parent,
1901                                             root_objectid, owner, offset,
1902                                             refs_to_add, extent_op);
1903                 ret = 0;
1904         }
1905         return ret;
1906 }
1907
1908 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1909                                  struct btrfs_fs_info *fs_info,
1910                                  struct btrfs_path *path,
1911                                  u64 bytenr, u64 parent, u64 root_objectid,
1912                                  u64 owner, u64 offset, int refs_to_add)
1913 {
1914         int ret;
1915         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1916                 BUG_ON(refs_to_add != 1);
1917                 ret = insert_tree_block_ref(trans, fs_info, path, bytenr,
1918                                             parent, root_objectid);
1919         } else {
1920                 ret = insert_extent_data_ref(trans, fs_info, path, bytenr,
1921                                              parent, root_objectid,
1922                                              owner, offset, refs_to_add);
1923         }
1924         return ret;
1925 }
1926
1927 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1928                                  struct btrfs_fs_info *fs_info,
1929                                  struct btrfs_path *path,
1930                                  struct btrfs_extent_inline_ref *iref,
1931                                  int refs_to_drop, int is_data, int *last_ref)
1932 {
1933         int ret = 0;
1934
1935         BUG_ON(!is_data && refs_to_drop != 1);
1936         if (iref) {
1937                 update_inline_extent_backref(fs_info, path, iref,
1938                                              -refs_to_drop, NULL, last_ref);
1939         } else if (is_data) {
1940                 ret = remove_extent_data_ref(trans, fs_info, path, refs_to_drop,
1941                                              last_ref);
1942         } else {
1943                 *last_ref = 1;
1944                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1945         }
1946         return ret;
1947 }
1948
1949 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1950 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1951                                u64 *discarded_bytes)
1952 {
1953         int j, ret = 0;
1954         u64 bytes_left, end;
1955         u64 aligned_start = ALIGN(start, 1 << 9);
1956
1957         if (WARN_ON(start != aligned_start)) {
1958                 len -= aligned_start - start;
1959                 len = round_down(len, 1 << 9);
1960                 start = aligned_start;
1961         }
1962
1963         *discarded_bytes = 0;
1964
1965         if (!len)
1966                 return 0;
1967
1968         end = start + len;
1969         bytes_left = len;
1970
1971         /* Skip any superblocks on this device. */
1972         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1973                 u64 sb_start = btrfs_sb_offset(j);
1974                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1975                 u64 size = sb_start - start;
1976
1977                 if (!in_range(sb_start, start, bytes_left) &&
1978                     !in_range(sb_end, start, bytes_left) &&
1979                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1980                         continue;
1981
1982                 /*
1983                  * Superblock spans beginning of range.  Adjust start and
1984                  * try again.
1985                  */
1986                 if (sb_start <= start) {
1987                         start += sb_end - start;
1988                         if (start > end) {
1989                                 bytes_left = 0;
1990                                 break;
1991                         }
1992                         bytes_left = end - start;
1993                         continue;
1994                 }
1995
1996                 if (size) {
1997                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1998                                                    GFP_NOFS, 0);
1999                         if (!ret)
2000                                 *discarded_bytes += size;
2001                         else if (ret != -EOPNOTSUPP)
2002                                 return ret;
2003                 }
2004
2005                 start = sb_end;
2006                 if (start > end) {
2007                         bytes_left = 0;
2008                         break;
2009                 }
2010                 bytes_left = end - start;
2011         }
2012
2013         if (bytes_left) {
2014                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2015                                            GFP_NOFS, 0);
2016                 if (!ret)
2017                         *discarded_bytes += bytes_left;
2018         }
2019         return ret;
2020 }
2021
2022 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2023                          u64 num_bytes, u64 *actual_bytes)
2024 {
2025         int ret;
2026         u64 discarded_bytes = 0;
2027         struct btrfs_bio *bbio = NULL;
2028
2029
2030         /*
2031          * Avoid races with device replace and make sure our bbio has devices
2032          * associated to its stripes that don't go away while we are discarding.
2033          */
2034         btrfs_bio_counter_inc_blocked(fs_info);
2035         /* Tell the block device(s) that the sectors can be discarded */
2036         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2037                               &bbio, 0);
2038         /* Error condition is -ENOMEM */
2039         if (!ret) {
2040                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2041                 int i;
2042
2043
2044                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2045                         u64 bytes;
2046                         if (!stripe->dev->can_discard)
2047                                 continue;
2048
2049                         ret = btrfs_issue_discard(stripe->dev->bdev,
2050                                                   stripe->physical,
2051                                                   stripe->length,
2052                                                   &bytes);
2053                         if (!ret)
2054                                 discarded_bytes += bytes;
2055                         else if (ret != -EOPNOTSUPP)
2056                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2057
2058                         /*
2059                          * Just in case we get back EOPNOTSUPP for some reason,
2060                          * just ignore the return value so we don't screw up
2061                          * people calling discard_extent.
2062                          */
2063                         ret = 0;
2064                 }
2065                 btrfs_put_bbio(bbio);
2066         }
2067         btrfs_bio_counter_dec(fs_info);
2068
2069         if (actual_bytes)
2070                 *actual_bytes = discarded_bytes;
2071
2072
2073         if (ret == -EOPNOTSUPP)
2074                 ret = 0;
2075         return ret;
2076 }
2077
2078 /* Can return -ENOMEM */
2079 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2080                          struct btrfs_fs_info *fs_info,
2081                          u64 bytenr, u64 num_bytes, u64 parent,
2082                          u64 root_objectid, u64 owner, u64 offset)
2083 {
2084         int ret;
2085
2086         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2087                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2088
2089         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2090                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2091                                         num_bytes,
2092                                         parent, root_objectid, (int)owner,
2093                                         BTRFS_ADD_DELAYED_REF, NULL);
2094         } else {
2095                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2096                                         num_bytes, parent, root_objectid,
2097                                         owner, offset, 0,
2098                                         BTRFS_ADD_DELAYED_REF);
2099         }
2100         return ret;
2101 }
2102
2103 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2104                                   struct btrfs_fs_info *fs_info,
2105                                   struct btrfs_delayed_ref_node *node,
2106                                   u64 parent, u64 root_objectid,
2107                                   u64 owner, u64 offset, int refs_to_add,
2108                                   struct btrfs_delayed_extent_op *extent_op)
2109 {
2110         struct btrfs_path *path;
2111         struct extent_buffer *leaf;
2112         struct btrfs_extent_item *item;
2113         struct btrfs_key key;
2114         u64 bytenr = node->bytenr;
2115         u64 num_bytes = node->num_bytes;
2116         u64 refs;
2117         int ret;
2118
2119         path = btrfs_alloc_path();
2120         if (!path)
2121                 return -ENOMEM;
2122
2123         path->reada = READA_FORWARD;
2124         path->leave_spinning = 1;
2125         /* this will setup the path even if it fails to insert the back ref */
2126         ret = insert_inline_extent_backref(trans, fs_info, path, bytenr,
2127                                            num_bytes, parent, root_objectid,
2128                                            owner, offset,
2129                                            refs_to_add, extent_op);
2130         if ((ret < 0 && ret != -EAGAIN) || !ret)
2131                 goto out;
2132
2133         /*
2134          * Ok we had -EAGAIN which means we didn't have space to insert and
2135          * inline extent ref, so just update the reference count and add a
2136          * normal backref.
2137          */
2138         leaf = path->nodes[0];
2139         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2140         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2141         refs = btrfs_extent_refs(leaf, item);
2142         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2143         if (extent_op)
2144                 __run_delayed_extent_op(extent_op, leaf, item);
2145
2146         btrfs_mark_buffer_dirty(leaf);
2147         btrfs_release_path(path);
2148
2149         path->reada = READA_FORWARD;
2150         path->leave_spinning = 1;
2151         /* now insert the actual backref */
2152         ret = insert_extent_backref(trans, fs_info, path, bytenr, parent,
2153                                     root_objectid, owner, offset, refs_to_add);
2154         if (ret)
2155                 btrfs_abort_transaction(trans, ret);
2156 out:
2157         btrfs_free_path(path);
2158         return ret;
2159 }
2160
2161 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2162                                 struct btrfs_fs_info *fs_info,
2163                                 struct btrfs_delayed_ref_node *node,
2164                                 struct btrfs_delayed_extent_op *extent_op,
2165                                 int insert_reserved)
2166 {
2167         int ret = 0;
2168         struct btrfs_delayed_data_ref *ref;
2169         struct btrfs_key ins;
2170         u64 parent = 0;
2171         u64 ref_root = 0;
2172         u64 flags = 0;
2173
2174         ins.objectid = node->bytenr;
2175         ins.offset = node->num_bytes;
2176         ins.type = BTRFS_EXTENT_ITEM_KEY;
2177
2178         ref = btrfs_delayed_node_to_data_ref(node);
2179         trace_run_delayed_data_ref(fs_info, node, ref, node->action);
2180
2181         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2182                 parent = ref->parent;
2183         ref_root = ref->root;
2184
2185         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2186                 if (extent_op)
2187                         flags |= extent_op->flags_to_set;
2188                 ret = alloc_reserved_file_extent(trans, fs_info,
2189                                                  parent, ref_root, flags,
2190                                                  ref->objectid, ref->offset,
2191                                                  &ins, node->ref_mod);
2192         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2193                 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent,
2194                                              ref_root, ref->objectid,
2195                                              ref->offset, node->ref_mod,
2196                                              extent_op);
2197         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2198                 ret = __btrfs_free_extent(trans, fs_info, node, parent,
2199                                           ref_root, ref->objectid,
2200                                           ref->offset, node->ref_mod,
2201                                           extent_op);
2202         } else {
2203                 BUG();
2204         }
2205         return ret;
2206 }
2207
2208 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2209                                     struct extent_buffer *leaf,
2210                                     struct btrfs_extent_item *ei)
2211 {
2212         u64 flags = btrfs_extent_flags(leaf, ei);
2213         if (extent_op->update_flags) {
2214                 flags |= extent_op->flags_to_set;
2215                 btrfs_set_extent_flags(leaf, ei, flags);
2216         }
2217
2218         if (extent_op->update_key) {
2219                 struct btrfs_tree_block_info *bi;
2220                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2221                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2222                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2223         }
2224 }
2225
2226 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2227                                  struct btrfs_fs_info *fs_info,
2228                                  struct btrfs_delayed_ref_node *node,
2229                                  struct btrfs_delayed_extent_op *extent_op)
2230 {
2231         struct btrfs_key key;
2232         struct btrfs_path *path;
2233         struct btrfs_extent_item *ei;
2234         struct extent_buffer *leaf;
2235         u32 item_size;
2236         int ret;
2237         int err = 0;
2238         int metadata = !extent_op->is_data;
2239
2240         if (trans->aborted)
2241                 return 0;
2242
2243         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2244                 metadata = 0;
2245
2246         path = btrfs_alloc_path();
2247         if (!path)
2248                 return -ENOMEM;
2249
2250         key.objectid = node->bytenr;
2251
2252         if (metadata) {
2253                 key.type = BTRFS_METADATA_ITEM_KEY;
2254                 key.offset = extent_op->level;
2255         } else {
2256                 key.type = BTRFS_EXTENT_ITEM_KEY;
2257                 key.offset = node->num_bytes;
2258         }
2259
2260 again:
2261         path->reada = READA_FORWARD;
2262         path->leave_spinning = 1;
2263         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2264         if (ret < 0) {
2265                 err = ret;
2266                 goto out;
2267         }
2268         if (ret > 0) {
2269                 if (metadata) {
2270                         if (path->slots[0] > 0) {
2271                                 path->slots[0]--;
2272                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2273                                                       path->slots[0]);
2274                                 if (key.objectid == node->bytenr &&
2275                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2276                                     key.offset == node->num_bytes)
2277                                         ret = 0;
2278                         }
2279                         if (ret > 0) {
2280                                 btrfs_release_path(path);
2281                                 metadata = 0;
2282
2283                                 key.objectid = node->bytenr;
2284                                 key.offset = node->num_bytes;
2285                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2286                                 goto again;
2287                         }
2288                 } else {
2289                         err = -EIO;
2290                         goto out;
2291                 }
2292         }
2293
2294         leaf = path->nodes[0];
2295         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2296 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2297         if (item_size < sizeof(*ei)) {
2298                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2299                 if (ret < 0) {
2300                         err = ret;
2301                         goto out;
2302                 }
2303                 leaf = path->nodes[0];
2304                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2305         }
2306 #endif
2307         BUG_ON(item_size < sizeof(*ei));
2308         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2309         __run_delayed_extent_op(extent_op, leaf, ei);
2310
2311         btrfs_mark_buffer_dirty(leaf);
2312 out:
2313         btrfs_free_path(path);
2314         return err;
2315 }
2316
2317 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2318                                 struct btrfs_fs_info *fs_info,
2319                                 struct btrfs_delayed_ref_node *node,
2320                                 struct btrfs_delayed_extent_op *extent_op,
2321                                 int insert_reserved)
2322 {
2323         int ret = 0;
2324         struct btrfs_delayed_tree_ref *ref;
2325         struct btrfs_key ins;
2326         u64 parent = 0;
2327         u64 ref_root = 0;
2328         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2329
2330         ref = btrfs_delayed_node_to_tree_ref(node);
2331         trace_run_delayed_tree_ref(fs_info, node, ref, node->action);
2332
2333         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2334                 parent = ref->parent;
2335         ref_root = ref->root;
2336
2337         ins.objectid = node->bytenr;
2338         if (skinny_metadata) {
2339                 ins.offset = ref->level;
2340                 ins.type = BTRFS_METADATA_ITEM_KEY;
2341         } else {
2342                 ins.offset = node->num_bytes;
2343                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2344         }
2345
2346         if (node->ref_mod != 1) {
2347                 btrfs_err(fs_info,
2348         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2349                           node->bytenr, node->ref_mod, node->action, ref_root,
2350                           parent);
2351                 return -EIO;
2352         }
2353         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2354                 BUG_ON(!extent_op || !extent_op->update_flags);
2355                 ret = alloc_reserved_tree_block(trans, fs_info,
2356                                                 parent, ref_root,
2357                                                 extent_op->flags_to_set,
2358                                                 &extent_op->key,
2359                                                 ref->level, &ins);
2360         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2361                 ret = __btrfs_inc_extent_ref(trans, fs_info, node,
2362                                              parent, ref_root,
2363                                              ref->level, 0, 1,
2364                                              extent_op);
2365         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2366                 ret = __btrfs_free_extent(trans, fs_info, node,
2367                                           parent, ref_root,
2368                                           ref->level, 0, 1, extent_op);
2369         } else {
2370                 BUG();
2371         }
2372         return ret;
2373 }
2374
2375 /* helper function to actually process a single delayed ref entry */
2376 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2377                                struct btrfs_fs_info *fs_info,
2378                                struct btrfs_delayed_ref_node *node,
2379                                struct btrfs_delayed_extent_op *extent_op,
2380                                int insert_reserved)
2381 {
2382         int ret = 0;
2383
2384         if (trans->aborted) {
2385                 if (insert_reserved)
2386                         btrfs_pin_extent(fs_info, node->bytenr,
2387                                          node->num_bytes, 1);
2388                 return 0;
2389         }
2390
2391         if (btrfs_delayed_ref_is_head(node)) {
2392                 struct btrfs_delayed_ref_head *head;
2393                 /*
2394                  * we've hit the end of the chain and we were supposed
2395                  * to insert this extent into the tree.  But, it got
2396                  * deleted before we ever needed to insert it, so all
2397                  * we have to do is clean up the accounting
2398                  */
2399                 BUG_ON(extent_op);
2400                 head = btrfs_delayed_node_to_head(node);
2401                 trace_run_delayed_ref_head(fs_info, node, head, node->action);
2402
2403                 if (insert_reserved) {
2404                         btrfs_pin_extent(fs_info, node->bytenr,
2405                                          node->num_bytes, 1);
2406                         if (head->is_data) {
2407                                 ret = btrfs_del_csums(trans, fs_info,
2408                                                       node->bytenr,
2409                                                       node->num_bytes);
2410                         }
2411                 }
2412
2413                 /* Also free its reserved qgroup space */
2414                 btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2415                                               head->qgroup_reserved);
2416                 return ret;
2417         }
2418
2419         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2420             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2421                 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op,
2422                                            insert_reserved);
2423         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2424                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2425                 ret = run_delayed_data_ref(trans, fs_info, node, extent_op,
2426                                            insert_reserved);
2427         else
2428                 BUG();
2429         return ret;
2430 }
2431
2432 static inline struct btrfs_delayed_ref_node *
2433 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2434 {
2435         struct btrfs_delayed_ref_node *ref;
2436
2437         if (list_empty(&head->ref_list))
2438                 return NULL;
2439
2440         /*
2441          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2442          * This is to prevent a ref count from going down to zero, which deletes
2443          * the extent item from the extent tree, when there still are references
2444          * to add, which would fail because they would not find the extent item.
2445          */
2446         if (!list_empty(&head->ref_add_list))
2447                 return list_first_entry(&head->ref_add_list,
2448                                 struct btrfs_delayed_ref_node, add_list);
2449
2450         ref = list_first_entry(&head->ref_list, struct btrfs_delayed_ref_node,
2451                                list);
2452         ASSERT(list_empty(&ref->add_list));
2453         return ref;
2454 }
2455
2456 /*
2457  * Returns 0 on success or if called with an already aborted transaction.
2458  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2459  */
2460 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2461                                              struct btrfs_fs_info *fs_info,
2462                                              unsigned long nr)
2463 {
2464         struct btrfs_delayed_ref_root *delayed_refs;
2465         struct btrfs_delayed_ref_node *ref;
2466         struct btrfs_delayed_ref_head *locked_ref = NULL;
2467         struct btrfs_delayed_extent_op *extent_op;
2468         ktime_t start = ktime_get();
2469         int ret;
2470         unsigned long count = 0;
2471         unsigned long actual_count = 0;
2472         int must_insert_reserved = 0;
2473
2474         delayed_refs = &trans->transaction->delayed_refs;
2475         while (1) {
2476                 if (!locked_ref) {
2477                         if (count >= nr)
2478                                 break;
2479
2480                         spin_lock(&delayed_refs->lock);
2481                         locked_ref = btrfs_select_ref_head(trans);
2482                         if (!locked_ref) {
2483                                 spin_unlock(&delayed_refs->lock);
2484                                 break;
2485                         }
2486
2487                         /* grab the lock that says we are going to process
2488                          * all the refs for this head */
2489                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2490                         spin_unlock(&delayed_refs->lock);
2491                         /*
2492                          * we may have dropped the spin lock to get the head
2493                          * mutex lock, and that might have given someone else
2494                          * time to free the head.  If that's true, it has been
2495                          * removed from our list and we can move on.
2496                          */
2497                         if (ret == -EAGAIN) {
2498                                 locked_ref = NULL;
2499                                 count++;
2500                                 continue;
2501                         }
2502                 }
2503
2504                 /*
2505                  * We need to try and merge add/drops of the same ref since we
2506                  * can run into issues with relocate dropping the implicit ref
2507                  * and then it being added back again before the drop can
2508                  * finish.  If we merged anything we need to re-loop so we can
2509                  * get a good ref.
2510                  * Or we can get node references of the same type that weren't
2511                  * merged when created due to bumps in the tree mod seq, and
2512                  * we need to merge them to prevent adding an inline extent
2513                  * backref before dropping it (triggering a BUG_ON at
2514                  * insert_inline_extent_backref()).
2515                  */
2516                 spin_lock(&locked_ref->lock);
2517                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2518                                          locked_ref);
2519
2520                 /*
2521                  * locked_ref is the head node, so we have to go one
2522                  * node back for any delayed ref updates
2523                  */
2524                 ref = select_delayed_ref(locked_ref);
2525
2526                 if (ref && ref->seq &&
2527                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2528                         spin_unlock(&locked_ref->lock);
2529                         spin_lock(&delayed_refs->lock);
2530                         locked_ref->processing = 0;
2531                         delayed_refs->num_heads_ready++;
2532                         spin_unlock(&delayed_refs->lock);
2533                         btrfs_delayed_ref_unlock(locked_ref);
2534                         locked_ref = NULL;
2535                         cond_resched();
2536                         count++;
2537                         continue;
2538                 }
2539
2540                 /*
2541                  * record the must insert reserved flag before we
2542                  * drop the spin lock.
2543                  */
2544                 must_insert_reserved = locked_ref->must_insert_reserved;
2545                 locked_ref->must_insert_reserved = 0;
2546
2547                 extent_op = locked_ref->extent_op;
2548                 locked_ref->extent_op = NULL;
2549
2550                 if (!ref) {
2551
2552
2553                         /* All delayed refs have been processed, Go ahead
2554                          * and send the head node to run_one_delayed_ref,
2555                          * so that any accounting fixes can happen
2556                          */
2557                         ref = &locked_ref->node;
2558
2559                         if (extent_op && must_insert_reserved) {
2560                                 btrfs_free_delayed_extent_op(extent_op);
2561                                 extent_op = NULL;
2562                         }
2563
2564                         if (extent_op) {
2565                                 spin_unlock(&locked_ref->lock);
2566                                 ret = run_delayed_extent_op(trans, fs_info,
2567                                                             ref, extent_op);
2568                                 btrfs_free_delayed_extent_op(extent_op);
2569
2570                                 if (ret) {
2571                                         /*
2572                                          * Need to reset must_insert_reserved if
2573                                          * there was an error so the abort stuff
2574                                          * can cleanup the reserved space
2575                                          * properly.
2576                                          */
2577                                         if (must_insert_reserved)
2578                                                 locked_ref->must_insert_reserved = 1;
2579                                         spin_lock(&delayed_refs->lock);
2580                                         locked_ref->processing = 0;
2581                                         delayed_refs->num_heads_ready++;
2582                                         spin_unlock(&delayed_refs->lock);
2583                                         btrfs_debug(fs_info,
2584                                                     "run_delayed_extent_op returned %d",
2585                                                     ret);
2586                                         btrfs_delayed_ref_unlock(locked_ref);
2587                                         return ret;
2588                                 }
2589                                 continue;
2590                         }
2591
2592                         /*
2593                          * Need to drop our head ref lock and re-acquire the
2594                          * delayed ref lock and then re-check to make sure
2595                          * nobody got added.
2596                          */
2597                         spin_unlock(&locked_ref->lock);
2598                         spin_lock(&delayed_refs->lock);
2599                         spin_lock(&locked_ref->lock);
2600                         if (!list_empty(&locked_ref->ref_list) ||
2601                             locked_ref->extent_op) {
2602                                 spin_unlock(&locked_ref->lock);
2603                                 spin_unlock(&delayed_refs->lock);
2604                                 continue;
2605                         }
2606                         ref->in_tree = 0;
2607                         delayed_refs->num_heads--;
2608                         rb_erase(&locked_ref->href_node,
2609                                  &delayed_refs->href_root);
2610                         spin_unlock(&delayed_refs->lock);
2611                 } else {
2612                         actual_count++;
2613                         ref->in_tree = 0;
2614                         list_del(&ref->list);
2615                         if (!list_empty(&ref->add_list))
2616                                 list_del(&ref->add_list);
2617                 }
2618                 atomic_dec(&delayed_refs->num_entries);
2619
2620                 if (!btrfs_delayed_ref_is_head(ref)) {
2621                         /*
2622                          * when we play the delayed ref, also correct the
2623                          * ref_mod on head
2624                          */
2625                         switch (ref->action) {
2626                         case BTRFS_ADD_DELAYED_REF:
2627                         case BTRFS_ADD_DELAYED_EXTENT:
2628                                 locked_ref->node.ref_mod -= ref->ref_mod;
2629                                 break;
2630                         case BTRFS_DROP_DELAYED_REF:
2631                                 locked_ref->node.ref_mod += ref->ref_mod;
2632                                 break;
2633                         default:
2634                                 WARN_ON(1);
2635                         }
2636                 }
2637                 spin_unlock(&locked_ref->lock);
2638
2639                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2640                                           must_insert_reserved);
2641
2642                 btrfs_free_delayed_extent_op(extent_op);
2643                 if (ret) {
2644                         spin_lock(&delayed_refs->lock);
2645                         locked_ref->processing = 0;
2646                         delayed_refs->num_heads_ready++;
2647                         spin_unlock(&delayed_refs->lock);
2648                         btrfs_delayed_ref_unlock(locked_ref);
2649                         btrfs_put_delayed_ref(ref);
2650                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2651                                     ret);
2652                         return ret;
2653                 }
2654
2655                 /*
2656                  * If this node is a head, that means all the refs in this head
2657                  * have been dealt with, and we will pick the next head to deal
2658                  * with, so we must unlock the head and drop it from the cluster
2659                  * list before we release it.
2660                  */
2661                 if (btrfs_delayed_ref_is_head(ref)) {
2662                         if (locked_ref->is_data &&
2663                             locked_ref->total_ref_mod < 0) {
2664                                 spin_lock(&delayed_refs->lock);
2665                                 delayed_refs->pending_csums -= ref->num_bytes;
2666                                 spin_unlock(&delayed_refs->lock);
2667                         }
2668                         btrfs_delayed_ref_unlock(locked_ref);
2669                         locked_ref = NULL;
2670                 }
2671                 btrfs_put_delayed_ref(ref);
2672                 count++;
2673                 cond_resched();
2674         }
2675
2676         /*
2677          * We don't want to include ref heads since we can have empty ref heads
2678          * and those will drastically skew our runtime down since we just do
2679          * accounting, no actual extent tree updates.
2680          */
2681         if (actual_count > 0) {
2682                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2683                 u64 avg;
2684
2685                 /*
2686                  * We weigh the current average higher than our current runtime
2687                  * to avoid large swings in the average.
2688                  */
2689                 spin_lock(&delayed_refs->lock);
2690                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2691                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2692                 spin_unlock(&delayed_refs->lock);
2693         }
2694         return 0;
2695 }
2696
2697 #ifdef SCRAMBLE_DELAYED_REFS
2698 /*
2699  * Normally delayed refs get processed in ascending bytenr order. This
2700  * correlates in most cases to the order added. To expose dependencies on this
2701  * order, we start to process the tree in the middle instead of the beginning
2702  */
2703 static u64 find_middle(struct rb_root *root)
2704 {
2705         struct rb_node *n = root->rb_node;
2706         struct btrfs_delayed_ref_node *entry;
2707         int alt = 1;
2708         u64 middle;
2709         u64 first = 0, last = 0;
2710
2711         n = rb_first(root);
2712         if (n) {
2713                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2714                 first = entry->bytenr;
2715         }
2716         n = rb_last(root);
2717         if (n) {
2718                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2719                 last = entry->bytenr;
2720         }
2721         n = root->rb_node;
2722
2723         while (n) {
2724                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2725                 WARN_ON(!entry->in_tree);
2726
2727                 middle = entry->bytenr;
2728
2729                 if (alt)
2730                         n = n->rb_left;
2731                 else
2732                         n = n->rb_right;
2733
2734                 alt = 1 - alt;
2735         }
2736         return middle;
2737 }
2738 #endif
2739
2740 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2741 {
2742         u64 num_bytes;
2743
2744         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2745                              sizeof(struct btrfs_extent_inline_ref));
2746         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2747                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2748
2749         /*
2750          * We don't ever fill up leaves all the way so multiply by 2 just to be
2751          * closer to what we're really going to want to use.
2752          */
2753         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2754 }
2755
2756 /*
2757  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2758  * would require to store the csums for that many bytes.
2759  */
2760 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2761 {
2762         u64 csum_size;
2763         u64 num_csums_per_leaf;
2764         u64 num_csums;
2765
2766         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2767         num_csums_per_leaf = div64_u64(csum_size,
2768                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2769         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2770         num_csums += num_csums_per_leaf - 1;
2771         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2772         return num_csums;
2773 }
2774
2775 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2776                                        struct btrfs_fs_info *fs_info)
2777 {
2778         struct btrfs_block_rsv *global_rsv;
2779         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2780         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2781         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2782         u64 num_bytes, num_dirty_bgs_bytes;
2783         int ret = 0;
2784
2785         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2786         num_heads = heads_to_leaves(fs_info, num_heads);
2787         if (num_heads > 1)
2788                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2789         num_bytes <<= 1;
2790         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2791                                                         fs_info->nodesize;
2792         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2793                                                              num_dirty_bgs);
2794         global_rsv = &fs_info->global_block_rsv;
2795
2796         /*
2797          * If we can't allocate any more chunks lets make sure we have _lots_ of
2798          * wiggle room since running delayed refs can create more delayed refs.
2799          */
2800         if (global_rsv->space_info->full) {
2801                 num_dirty_bgs_bytes <<= 1;
2802                 num_bytes <<= 1;
2803         }
2804
2805         spin_lock(&global_rsv->lock);
2806         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2807                 ret = 1;
2808         spin_unlock(&global_rsv->lock);
2809         return ret;
2810 }
2811
2812 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2813                                        struct btrfs_fs_info *fs_info)
2814 {
2815         u64 num_entries =
2816                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2817         u64 avg_runtime;
2818         u64 val;
2819
2820         smp_mb();
2821         avg_runtime = fs_info->avg_delayed_ref_runtime;
2822         val = num_entries * avg_runtime;
2823         if (val >= NSEC_PER_SEC)
2824                 return 1;
2825         if (val >= NSEC_PER_SEC / 2)
2826                 return 2;
2827
2828         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2829 }
2830
2831 struct async_delayed_refs {
2832         struct btrfs_root *root;
2833         u64 transid;
2834         int count;
2835         int error;
2836         int sync;
2837         struct completion wait;
2838         struct btrfs_work work;
2839 };
2840
2841 static inline struct async_delayed_refs *
2842 to_async_delayed_refs(struct btrfs_work *work)
2843 {
2844         return container_of(work, struct async_delayed_refs, work);
2845 }
2846
2847 static void delayed_ref_async_start(struct btrfs_work *work)
2848 {
2849         struct async_delayed_refs *async = to_async_delayed_refs(work);
2850         struct btrfs_trans_handle *trans;
2851         struct btrfs_fs_info *fs_info = async->root->fs_info;
2852         int ret;
2853
2854         /* if the commit is already started, we don't need to wait here */
2855         if (btrfs_transaction_blocked(fs_info))
2856                 goto done;
2857
2858         trans = btrfs_join_transaction(async->root);
2859         if (IS_ERR(trans)) {
2860                 async->error = PTR_ERR(trans);
2861                 goto done;
2862         }
2863
2864         /*
2865          * trans->sync means that when we call end_transaction, we won't
2866          * wait on delayed refs
2867          */
2868         trans->sync = true;
2869
2870         /* Don't bother flushing if we got into a different transaction */
2871         if (trans->transid > async->transid)
2872                 goto end;
2873
2874         ret = btrfs_run_delayed_refs(trans, fs_info, async->count);
2875         if (ret)
2876                 async->error = ret;
2877 end:
2878         ret = btrfs_end_transaction(trans);
2879         if (ret && !async->error)
2880                 async->error = ret;
2881 done:
2882         if (async->sync)
2883                 complete(&async->wait);
2884         else
2885                 kfree(async);
2886 }
2887
2888 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2889                                  unsigned long count, u64 transid, int wait)
2890 {
2891         struct async_delayed_refs *async;
2892         int ret;
2893
2894         async = kmalloc(sizeof(*async), GFP_NOFS);
2895         if (!async)
2896                 return -ENOMEM;
2897
2898         async->root = fs_info->tree_root;
2899         async->count = count;
2900         async->error = 0;
2901         async->transid = transid;
2902         if (wait)
2903                 async->sync = 1;
2904         else
2905                 async->sync = 0;
2906         init_completion(&async->wait);
2907
2908         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2909                         delayed_ref_async_start, NULL, NULL);
2910
2911         btrfs_queue_work(fs_info->extent_workers, &async->work);
2912
2913         if (wait) {
2914                 wait_for_completion(&async->wait);
2915                 ret = async->error;
2916                 kfree(async);
2917                 return ret;
2918         }
2919         return 0;
2920 }
2921
2922 /*
2923  * this starts processing the delayed reference count updates and
2924  * extent insertions we have queued up so far.  count can be
2925  * 0, which means to process everything in the tree at the start
2926  * of the run (but not newly added entries), or it can be some target
2927  * number you'd like to process.
2928  *
2929  * Returns 0 on success or if called with an aborted transaction
2930  * Returns <0 on error and aborts the transaction
2931  */
2932 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2933                            struct btrfs_fs_info *fs_info, unsigned long count)
2934 {
2935         struct rb_node *node;
2936         struct btrfs_delayed_ref_root *delayed_refs;
2937         struct btrfs_delayed_ref_head *head;
2938         int ret;
2939         int run_all = count == (unsigned long)-1;
2940         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2941
2942         /* We'll clean this up in btrfs_cleanup_transaction */
2943         if (trans->aborted)
2944                 return 0;
2945
2946         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2947                 return 0;
2948
2949         delayed_refs = &trans->transaction->delayed_refs;
2950         if (count == 0)
2951                 count = atomic_read(&delayed_refs->num_entries) * 2;
2952
2953 again:
2954 #ifdef SCRAMBLE_DELAYED_REFS
2955         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2956 #endif
2957         trans->can_flush_pending_bgs = false;
2958         ret = __btrfs_run_delayed_refs(trans, fs_info, count);
2959         if (ret < 0) {
2960                 btrfs_abort_transaction(trans, ret);
2961                 return ret;
2962         }
2963
2964         if (run_all) {
2965                 if (!list_empty(&trans->new_bgs))
2966                         btrfs_create_pending_block_groups(trans, fs_info);
2967
2968                 spin_lock(&delayed_refs->lock);
2969                 node = rb_first(&delayed_refs->href_root);
2970                 if (!node) {
2971                         spin_unlock(&delayed_refs->lock);
2972                         goto out;
2973                 }
2974
2975                 while (node) {
2976                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2977                                         href_node);
2978                         if (btrfs_delayed_ref_is_head(&head->node)) {
2979                                 struct btrfs_delayed_ref_node *ref;
2980
2981                                 ref = &head->node;
2982                                 atomic_inc(&ref->refs);
2983
2984                                 spin_unlock(&delayed_refs->lock);
2985                                 /*
2986                                  * Mutex was contended, block until it's
2987                                  * released and try again
2988                                  */
2989                                 mutex_lock(&head->mutex);
2990                                 mutex_unlock(&head->mutex);
2991
2992                                 btrfs_put_delayed_ref(ref);
2993                                 cond_resched();
2994                                 goto again;
2995                         } else {
2996                                 WARN_ON(1);
2997                         }
2998                         node = rb_next(node);
2999                 }
3000                 spin_unlock(&delayed_refs->lock);
3001                 cond_resched();
3002                 goto again;
3003         }
3004 out:
3005         assert_qgroups_uptodate(trans);
3006         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3007         return 0;
3008 }
3009
3010 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3011                                 struct btrfs_fs_info *fs_info,
3012                                 u64 bytenr, u64 num_bytes, u64 flags,
3013                                 int level, int is_data)
3014 {
3015         struct btrfs_delayed_extent_op *extent_op;
3016         int ret;
3017
3018         extent_op = btrfs_alloc_delayed_extent_op();
3019         if (!extent_op)
3020                 return -ENOMEM;
3021
3022         extent_op->flags_to_set = flags;
3023         extent_op->update_flags = true;
3024         extent_op->update_key = false;
3025         extent_op->is_data = is_data ? true : false;
3026         extent_op->level = level;
3027
3028         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3029                                           num_bytes, extent_op);
3030         if (ret)
3031                 btrfs_free_delayed_extent_op(extent_op);
3032         return ret;
3033 }
3034
3035 static noinline int check_delayed_ref(struct btrfs_root *root,
3036                                       struct btrfs_path *path,
3037                                       u64 objectid, u64 offset, u64 bytenr)
3038 {
3039         struct btrfs_delayed_ref_head *head;
3040         struct btrfs_delayed_ref_node *ref;
3041         struct btrfs_delayed_data_ref *data_ref;
3042         struct btrfs_delayed_ref_root *delayed_refs;
3043         struct btrfs_transaction *cur_trans;
3044         int ret = 0;
3045
3046         cur_trans = root->fs_info->running_transaction;
3047         if (!cur_trans)
3048                 return 0;
3049
3050         delayed_refs = &cur_trans->delayed_refs;
3051         spin_lock(&delayed_refs->lock);
3052         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3053         if (!head) {
3054                 spin_unlock(&delayed_refs->lock);
3055                 return 0;
3056         }
3057
3058         if (!mutex_trylock(&head->mutex)) {
3059                 atomic_inc(&head->node.refs);
3060                 spin_unlock(&delayed_refs->lock);
3061
3062                 btrfs_release_path(path);
3063
3064                 /*
3065                  * Mutex was contended, block until it's released and let
3066                  * caller try again
3067                  */
3068                 mutex_lock(&head->mutex);
3069                 mutex_unlock(&head->mutex);
3070                 btrfs_put_delayed_ref(&head->node);
3071                 return -EAGAIN;
3072         }
3073         spin_unlock(&delayed_refs->lock);
3074
3075         spin_lock(&head->lock);
3076         list_for_each_entry(ref, &head->ref_list, list) {
3077                 /* If it's a shared ref we know a cross reference exists */
3078                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3079                         ret = 1;
3080                         break;
3081                 }
3082
3083                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3084
3085                 /*
3086                  * If our ref doesn't match the one we're currently looking at
3087                  * then we have a cross reference.
3088                  */
3089                 if (data_ref->root != root->root_key.objectid ||
3090                     data_ref->objectid != objectid ||
3091                     data_ref->offset != offset) {
3092                         ret = 1;
3093                         break;
3094                 }
3095         }
3096         spin_unlock(&head->lock);
3097         mutex_unlock(&head->mutex);
3098         return ret;
3099 }
3100
3101 static noinline int check_committed_ref(struct btrfs_root *root,
3102                                         struct btrfs_path *path,
3103                                         u64 objectid, u64 offset, u64 bytenr)
3104 {
3105         struct btrfs_fs_info *fs_info = root->fs_info;
3106         struct btrfs_root *extent_root = fs_info->extent_root;
3107         struct extent_buffer *leaf;
3108         struct btrfs_extent_data_ref *ref;
3109         struct btrfs_extent_inline_ref *iref;
3110         struct btrfs_extent_item *ei;
3111         struct btrfs_key key;
3112         u32 item_size;
3113         int ret;
3114
3115         key.objectid = bytenr;
3116         key.offset = (u64)-1;
3117         key.type = BTRFS_EXTENT_ITEM_KEY;
3118
3119         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3120         if (ret < 0)
3121                 goto out;
3122         BUG_ON(ret == 0); /* Corruption */
3123
3124         ret = -ENOENT;
3125         if (path->slots[0] == 0)
3126                 goto out;
3127
3128         path->slots[0]--;
3129         leaf = path->nodes[0];
3130         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3131
3132         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3133                 goto out;
3134
3135         ret = 1;
3136         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3137 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3138         if (item_size < sizeof(*ei)) {
3139                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3140                 goto out;
3141         }
3142 #endif
3143         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3144
3145         if (item_size != sizeof(*ei) +
3146             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3147                 goto out;
3148
3149         if (btrfs_extent_generation(leaf, ei) <=
3150             btrfs_root_last_snapshot(&root->root_item))
3151                 goto out;
3152
3153         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3154         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3155             BTRFS_EXTENT_DATA_REF_KEY)
3156                 goto out;
3157
3158         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3159         if (btrfs_extent_refs(leaf, ei) !=
3160             btrfs_extent_data_ref_count(leaf, ref) ||
3161             btrfs_extent_data_ref_root(leaf, ref) !=
3162             root->root_key.objectid ||
3163             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3164             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3165                 goto out;
3166
3167         ret = 0;
3168 out:
3169         return ret;
3170 }
3171
3172 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3173                           u64 bytenr)
3174 {
3175         struct btrfs_path *path;
3176         int ret;
3177         int ret2;
3178
3179         path = btrfs_alloc_path();
3180         if (!path)
3181                 return -ENOENT;
3182
3183         do {
3184                 ret = check_committed_ref(root, path, objectid,
3185                                           offset, bytenr);
3186                 if (ret && ret != -ENOENT)
3187                         goto out;
3188
3189                 ret2 = check_delayed_ref(root, path, objectid,
3190                                          offset, bytenr);
3191         } while (ret2 == -EAGAIN);
3192
3193         if (ret2 && ret2 != -ENOENT) {
3194                 ret = ret2;
3195                 goto out;
3196         }
3197
3198         if (ret != -ENOENT || ret2 != -ENOENT)
3199                 ret = 0;
3200 out:
3201         btrfs_free_path(path);
3202         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3203                 WARN_ON(ret > 0);
3204         return ret;
3205 }
3206
3207 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3208                            struct btrfs_root *root,
3209                            struct extent_buffer *buf,
3210                            int full_backref, int inc)
3211 {
3212         struct btrfs_fs_info *fs_info = root->fs_info;
3213         u64 bytenr;
3214         u64 num_bytes;
3215         u64 parent;
3216         u64 ref_root;
3217         u32 nritems;
3218         struct btrfs_key key;
3219         struct btrfs_file_extent_item *fi;
3220         int i;
3221         int level;
3222         int ret = 0;
3223         int (*process_func)(struct btrfs_trans_handle *,
3224                             struct btrfs_fs_info *,
3225                             u64, u64, u64, u64, u64, u64);
3226
3227
3228         if (btrfs_is_testing(fs_info))
3229                 return 0;
3230
3231         ref_root = btrfs_header_owner(buf);
3232         nritems = btrfs_header_nritems(buf);
3233         level = btrfs_header_level(buf);
3234
3235         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3236                 return 0;
3237
3238         if (inc)
3239                 process_func = btrfs_inc_extent_ref;
3240         else
3241                 process_func = btrfs_free_extent;
3242
3243         if (full_backref)
3244                 parent = buf->start;
3245         else
3246                 parent = 0;
3247
3248         for (i = 0; i < nritems; i++) {
3249                 if (level == 0) {
3250                         btrfs_item_key_to_cpu(buf, &key, i);
3251                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3252                                 continue;
3253                         fi = btrfs_item_ptr(buf, i,
3254                                             struct btrfs_file_extent_item);
3255                         if (btrfs_file_extent_type(buf, fi) ==
3256                             BTRFS_FILE_EXTENT_INLINE)
3257                                 continue;
3258                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3259                         if (bytenr == 0)
3260                                 continue;
3261
3262                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3263                         key.offset -= btrfs_file_extent_offset(buf, fi);
3264                         ret = process_func(trans, fs_info, bytenr, num_bytes,
3265                                            parent, ref_root, key.objectid,
3266                                            key.offset);
3267                         if (ret)
3268                                 goto fail;
3269                 } else {
3270                         bytenr = btrfs_node_blockptr(buf, i);
3271                         num_bytes = fs_info->nodesize;
3272                         ret = process_func(trans, fs_info, bytenr, num_bytes,
3273                                            parent, ref_root, level - 1, 0);
3274                         if (ret)
3275                                 goto fail;
3276                 }
3277         }
3278         return 0;
3279 fail:
3280         return ret;
3281 }
3282
3283 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3284                   struct extent_buffer *buf, int full_backref)
3285 {
3286         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3287 }
3288
3289 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3290                   struct extent_buffer *buf, int full_backref)
3291 {
3292         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3293 }
3294
3295 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3296                                  struct btrfs_fs_info *fs_info,
3297                                  struct btrfs_path *path,
3298                                  struct btrfs_block_group_cache *cache)
3299 {
3300         int ret;
3301         struct btrfs_root *extent_root = fs_info->extent_root;
3302         unsigned long bi;
3303         struct extent_buffer *leaf;
3304
3305         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3306         if (ret) {
3307                 if (ret > 0)
3308                         ret = -ENOENT;
3309                 goto fail;
3310         }
3311
3312         leaf = path->nodes[0];
3313         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3314         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3315         btrfs_mark_buffer_dirty(leaf);
3316 fail:
3317         btrfs_release_path(path);
3318         return ret;
3319
3320 }
3321
3322 static struct btrfs_block_group_cache *
3323 next_block_group(struct btrfs_fs_info *fs_info,
3324                  struct btrfs_block_group_cache *cache)
3325 {
3326         struct rb_node *node;
3327
3328         spin_lock(&fs_info->block_group_cache_lock);
3329
3330         /* If our block group was removed, we need a full search. */
3331         if (RB_EMPTY_NODE(&cache->cache_node)) {
3332                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3333
3334                 spin_unlock(&fs_info->block_group_cache_lock);
3335                 btrfs_put_block_group(cache);
3336                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3337         }
3338         node = rb_next(&cache->cache_node);
3339         btrfs_put_block_group(cache);
3340         if (node) {
3341                 cache = rb_entry(node, struct btrfs_block_group_cache,
3342                                  cache_node);
3343                 btrfs_get_block_group(cache);
3344         } else
3345                 cache = NULL;
3346         spin_unlock(&fs_info->block_group_cache_lock);
3347         return cache;
3348 }
3349
3350 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3351                             struct btrfs_trans_handle *trans,
3352                             struct btrfs_path *path)
3353 {
3354         struct btrfs_fs_info *fs_info = block_group->fs_info;
3355         struct btrfs_root *root = fs_info->tree_root;
3356         struct inode *inode = NULL;
3357         u64 alloc_hint = 0;
3358         int dcs = BTRFS_DC_ERROR;
3359         u64 num_pages = 0;
3360         int retries = 0;
3361         int ret = 0;
3362
3363         /*
3364          * If this block group is smaller than 100 megs don't bother caching the
3365          * block group.
3366          */
3367         if (block_group->key.offset < (100 * SZ_1M)) {
3368                 spin_lock(&block_group->lock);
3369                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3370                 spin_unlock(&block_group->lock);
3371                 return 0;
3372         }
3373
3374         if (trans->aborted)
3375                 return 0;
3376 again:
3377         inode = lookup_free_space_inode(fs_info, block_group, path);
3378         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3379                 ret = PTR_ERR(inode);
3380                 btrfs_release_path(path);
3381                 goto out;
3382         }
3383
3384         if (IS_ERR(inode)) {
3385                 BUG_ON(retries);
3386                 retries++;
3387
3388                 if (block_group->ro)
3389                         goto out_free;
3390
3391                 ret = create_free_space_inode(fs_info, trans, block_group,
3392                                               path);
3393                 if (ret)
3394                         goto out_free;
3395                 goto again;
3396         }
3397
3398         /* We've already setup this transaction, go ahead and exit */
3399         if (block_group->cache_generation == trans->transid &&
3400             i_size_read(inode)) {
3401                 dcs = BTRFS_DC_SETUP;
3402                 goto out_put;
3403         }
3404
3405         /*
3406          * We want to set the generation to 0, that way if anything goes wrong
3407          * from here on out we know not to trust this cache when we load up next
3408          * time.
3409          */
3410         BTRFS_I(inode)->generation = 0;
3411         ret = btrfs_update_inode(trans, root, inode);
3412         if (ret) {
3413                 /*
3414                  * So theoretically we could recover from this, simply set the
3415                  * super cache generation to 0 so we know to invalidate the
3416                  * cache, but then we'd have to keep track of the block groups
3417                  * that fail this way so we know we _have_ to reset this cache
3418                  * before the next commit or risk reading stale cache.  So to
3419                  * limit our exposure to horrible edge cases lets just abort the
3420                  * transaction, this only happens in really bad situations
3421                  * anyway.
3422                  */
3423                 btrfs_abort_transaction(trans, ret);
3424                 goto out_put;
3425         }
3426         WARN_ON(ret);
3427
3428         if (i_size_read(inode) > 0) {
3429                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3430                                         &fs_info->global_block_rsv);
3431                 if (ret)
3432                         goto out_put;
3433
3434                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3435                 if (ret)
3436                         goto out_put;
3437         }
3438
3439         spin_lock(&block_group->lock);
3440         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3441             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3442                 /*
3443                  * don't bother trying to write stuff out _if_
3444                  * a) we're not cached,
3445                  * b) we're with nospace_cache mount option.
3446                  */
3447                 dcs = BTRFS_DC_WRITTEN;
3448                 spin_unlock(&block_group->lock);
3449                 goto out_put;
3450         }
3451         spin_unlock(&block_group->lock);
3452
3453         /*
3454          * We hit an ENOSPC when setting up the cache in this transaction, just
3455          * skip doing the setup, we've already cleared the cache so we're safe.
3456          */
3457         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3458                 ret = -ENOSPC;
3459                 goto out_put;
3460         }
3461
3462         /*
3463          * Try to preallocate enough space based on how big the block group is.
3464          * Keep in mind this has to include any pinned space which could end up
3465          * taking up quite a bit since it's not folded into the other space
3466          * cache.
3467          */
3468         num_pages = div_u64(block_group->key.offset, SZ_256M);
3469         if (!num_pages)
3470                 num_pages = 1;
3471
3472         num_pages *= 16;
3473         num_pages *= PAGE_SIZE;
3474
3475         ret = btrfs_check_data_free_space(inode, 0, num_pages);
3476         if (ret)
3477                 goto out_put;
3478
3479         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3480                                               num_pages, num_pages,
3481                                               &alloc_hint);
3482         /*
3483          * Our cache requires contiguous chunks so that we don't modify a bunch
3484          * of metadata or split extents when writing the cache out, which means
3485          * we can enospc if we are heavily fragmented in addition to just normal
3486          * out of space conditions.  So if we hit this just skip setting up any
3487          * other block groups for this transaction, maybe we'll unpin enough
3488          * space the next time around.
3489          */
3490         if (!ret)
3491                 dcs = BTRFS_DC_SETUP;
3492         else if (ret == -ENOSPC)
3493                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3494
3495 out_put:
3496         iput(inode);
3497 out_free:
3498         btrfs_release_path(path);
3499 out:
3500         spin_lock(&block_group->lock);
3501         if (!ret && dcs == BTRFS_DC_SETUP)
3502                 block_group->cache_generation = trans->transid;
3503         block_group->disk_cache_state = dcs;
3504         spin_unlock(&block_group->lock);
3505
3506         return ret;
3507 }
3508
3509 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3510                             struct btrfs_fs_info *fs_info)
3511 {
3512         struct btrfs_block_group_cache *cache, *tmp;
3513         struct btrfs_transaction *cur_trans = trans->transaction;
3514         struct btrfs_path *path;
3515
3516         if (list_empty(&cur_trans->dirty_bgs) ||
3517             !btrfs_test_opt(fs_info, SPACE_CACHE))
3518                 return 0;
3519
3520         path = btrfs_alloc_path();
3521         if (!path)
3522                 return -ENOMEM;
3523
3524         /* Could add new block groups, use _safe just in case */
3525         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3526                                  dirty_list) {
3527                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3528                         cache_save_setup(cache, trans, path);
3529         }
3530
3531         btrfs_free_path(path);
3532         return 0;
3533 }
3534
3535 /*
3536  * transaction commit does final block group cache writeback during a
3537  * critical section where nothing is allowed to change the FS.  This is
3538  * required in order for the cache to actually match the block group,
3539  * but can introduce a lot of latency into the commit.
3540  *
3541  * So, btrfs_start_dirty_block_groups is here to kick off block group
3542  * cache IO.  There's a chance we'll have to redo some of it if the
3543  * block group changes again during the commit, but it greatly reduces
3544  * the commit latency by getting rid of the easy block groups while
3545  * we're still allowing others to join the commit.
3546  */
3547 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3548                                    struct btrfs_fs_info *fs_info)
3549 {
3550         struct btrfs_block_group_cache *cache;
3551         struct btrfs_transaction *cur_trans = trans->transaction;
3552         int ret = 0;
3553         int should_put;
3554         struct btrfs_path *path = NULL;
3555         LIST_HEAD(dirty);
3556         struct list_head *io = &cur_trans->io_bgs;
3557         int num_started = 0;
3558         int loops = 0;
3559
3560         spin_lock(&cur_trans->dirty_bgs_lock);
3561         if (list_empty(&cur_trans->dirty_bgs)) {
3562                 spin_unlock(&cur_trans->dirty_bgs_lock);
3563                 return 0;
3564         }
3565         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3566         spin_unlock(&cur_trans->dirty_bgs_lock);
3567
3568 again:
3569         /*
3570          * make sure all the block groups on our dirty list actually
3571          * exist
3572          */
3573         btrfs_create_pending_block_groups(trans, fs_info);
3574
3575         if (!path) {
3576                 path = btrfs_alloc_path();
3577                 if (!path)
3578                         return -ENOMEM;
3579         }
3580
3581         /*
3582          * cache_write_mutex is here only to save us from balance or automatic
3583          * removal of empty block groups deleting this block group while we are
3584          * writing out the cache
3585          */
3586         mutex_lock(&trans->transaction->cache_write_mutex);
3587         while (!list_empty(&dirty)) {
3588                 cache = list_first_entry(&dirty,
3589                                          struct btrfs_block_group_cache,
3590                                          dirty_list);
3591                 /*
3592                  * this can happen if something re-dirties a block
3593                  * group that is already under IO.  Just wait for it to
3594                  * finish and then do it all again
3595                  */
3596                 if (!list_empty(&cache->io_list)) {
3597                         list_del_init(&cache->io_list);
3598                         btrfs_wait_cache_io(trans, cache, path);
3599                         btrfs_put_block_group(cache);
3600                 }
3601
3602
3603                 /*
3604                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3605                  * if it should update the cache_state.  Don't delete
3606                  * until after we wait.
3607                  *
3608                  * Since we're not running in the commit critical section
3609                  * we need the dirty_bgs_lock to protect from update_block_group
3610                  */
3611                 spin_lock(&cur_trans->dirty_bgs_lock);
3612                 list_del_init(&cache->dirty_list);
3613                 spin_unlock(&cur_trans->dirty_bgs_lock);
3614
3615                 should_put = 1;
3616
3617                 cache_save_setup(cache, trans, path);
3618
3619                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3620                         cache->io_ctl.inode = NULL;
3621                         ret = btrfs_write_out_cache(fs_info, trans,
3622                                                     cache, path);
3623                         if (ret == 0 && cache->io_ctl.inode) {
3624                                 num_started++;
3625                                 should_put = 0;
3626
3627                                 /*
3628                                  * the cache_write_mutex is protecting
3629                                  * the io_list
3630                                  */
3631                                 list_add_tail(&cache->io_list, io);
3632                         } else {
3633                                 /*
3634                                  * if we failed to write the cache, the
3635                                  * generation will be bad and life goes on
3636                                  */
3637                                 ret = 0;
3638                         }
3639                 }
3640                 if (!ret) {
3641                         ret = write_one_cache_group(trans, fs_info,
3642                                                     path, cache);
3643                         /*
3644                          * Our block group might still be attached to the list
3645                          * of new block groups in the transaction handle of some
3646                          * other task (struct btrfs_trans_handle->new_bgs). This
3647                          * means its block group item isn't yet in the extent
3648                          * tree. If this happens ignore the error, as we will
3649                          * try again later in the critical section of the
3650                          * transaction commit.
3651                          */
3652                         if (ret == -ENOENT) {
3653                                 ret = 0;
3654                                 spin_lock(&cur_trans->dirty_bgs_lock);
3655                                 if (list_empty(&cache->dirty_list)) {
3656                                         list_add_tail(&cache->dirty_list,
3657                                                       &cur_trans->dirty_bgs);
3658                                         btrfs_get_block_group(cache);
3659                                 }
3660                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3661                         } else if (ret) {
3662                                 btrfs_abort_transaction(trans, ret);
3663                         }
3664                 }
3665
3666                 /* if its not on the io list, we need to put the block group */
3667                 if (should_put)
3668                         btrfs_put_block_group(cache);
3669
3670                 if (ret)
3671                         break;
3672
3673                 /*
3674                  * Avoid blocking other tasks for too long. It might even save
3675                  * us from writing caches for block groups that are going to be
3676                  * removed.
3677                  */
3678                 mutex_unlock(&trans->transaction->cache_write_mutex);
3679                 mutex_lock(&trans->transaction->cache_write_mutex);
3680         }
3681         mutex_unlock(&trans->transaction->cache_write_mutex);
3682
3683         /*
3684          * go through delayed refs for all the stuff we've just kicked off
3685          * and then loop back (just once)
3686          */
3687         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
3688         if (!ret && loops == 0) {
3689                 loops++;
3690                 spin_lock(&cur_trans->dirty_bgs_lock);
3691                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3692                 /*
3693                  * dirty_bgs_lock protects us from concurrent block group
3694                  * deletes too (not just cache_write_mutex).
3695                  */
3696                 if (!list_empty(&dirty)) {
3697                         spin_unlock(&cur_trans->dirty_bgs_lock);
3698                         goto again;
3699                 }
3700                 spin_unlock(&cur_trans->dirty_bgs_lock);
3701         } else if (ret < 0) {
3702                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3703         }
3704
3705         btrfs_free_path(path);
3706         return ret;
3707 }
3708
3709 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3710                                    struct btrfs_fs_info *fs_info)
3711 {
3712         struct btrfs_block_group_cache *cache;
3713         struct btrfs_transaction *cur_trans = trans->transaction;
3714         int ret = 0;
3715         int should_put;
3716         struct btrfs_path *path;
3717         struct list_head *io = &cur_trans->io_bgs;
3718         int num_started = 0;
3719
3720         path = btrfs_alloc_path();
3721         if (!path)
3722                 return -ENOMEM;
3723
3724         /*
3725          * Even though we are in the critical section of the transaction commit,
3726          * we can still have concurrent tasks adding elements to this
3727          * transaction's list of dirty block groups. These tasks correspond to
3728          * endio free space workers started when writeback finishes for a
3729          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3730          * allocate new block groups as a result of COWing nodes of the root
3731          * tree when updating the free space inode. The writeback for the space
3732          * caches is triggered by an earlier call to
3733          * btrfs_start_dirty_block_groups() and iterations of the following
3734          * loop.
3735          * Also we want to do the cache_save_setup first and then run the
3736          * delayed refs to make sure we have the best chance at doing this all
3737          * in one shot.
3738          */
3739         spin_lock(&cur_trans->dirty_bgs_lock);
3740         while (!list_empty(&cur_trans->dirty_bgs)) {
3741                 cache = list_first_entry(&cur_trans->dirty_bgs,
3742                                          struct btrfs_block_group_cache,
3743                                          dirty_list);
3744
3745                 /*
3746                  * this can happen if cache_save_setup re-dirties a block
3747                  * group that is already under IO.  Just wait for it to
3748                  * finish and then do it all again
3749                  */
3750                 if (!list_empty(&cache->io_list)) {
3751                         spin_unlock(&cur_trans->dirty_bgs_lock);
3752                         list_del_init(&cache->io_list);
3753                         btrfs_wait_cache_io(trans, cache, path);
3754                         btrfs_put_block_group(cache);
3755                         spin_lock(&cur_trans->dirty_bgs_lock);
3756                 }
3757
3758                 /*
3759                  * don't remove from the dirty list until after we've waited
3760                  * on any pending IO
3761                  */
3762                 list_del_init(&cache->dirty_list);
3763                 spin_unlock(&cur_trans->dirty_bgs_lock);
3764                 should_put = 1;
3765
3766                 cache_save_setup(cache, trans, path);
3767
3768                 if (!ret)
3769                         ret = btrfs_run_delayed_refs(trans, fs_info,
3770                                                      (unsigned long) -1);
3771
3772                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3773                         cache->io_ctl.inode = NULL;
3774                         ret = btrfs_write_out_cache(fs_info, trans,
3775                                                     cache, path);
3776                         if (ret == 0 && cache->io_ctl.inode) {
3777                                 num_started++;
3778                                 should_put = 0;
3779                                 list_add_tail(&cache->io_list, io);
3780                         } else {
3781                                 /*
3782                                  * if we failed to write the cache, the
3783                                  * generation will be bad and life goes on
3784                                  */
3785                                 ret = 0;
3786                         }
3787                 }
3788                 if (!ret) {
3789                         ret = write_one_cache_group(trans, fs_info,
3790                                                     path, cache);
3791                         /*
3792                          * One of the free space endio workers might have
3793                          * created a new block group while updating a free space
3794                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3795                          * and hasn't released its transaction handle yet, in
3796                          * which case the new block group is still attached to
3797                          * its transaction handle and its creation has not
3798                          * finished yet (no block group item in the extent tree
3799                          * yet, etc). If this is the case, wait for all free
3800                          * space endio workers to finish and retry. This is a
3801                          * a very rare case so no need for a more efficient and
3802                          * complex approach.
3803                          */
3804                         if (ret == -ENOENT) {
3805                                 wait_event(cur_trans->writer_wait,
3806                                    atomic_read(&cur_trans->num_writers) == 1);
3807                                 ret = write_one_cache_group(trans, fs_info,
3808                                                             path, cache);
3809                         }
3810                         if (ret)
3811                                 btrfs_abort_transaction(trans, ret);
3812                 }
3813
3814                 /* if its not on the io list, we need to put the block group */
3815                 if (should_put)
3816                         btrfs_put_block_group(cache);
3817                 spin_lock(&cur_trans->dirty_bgs_lock);
3818         }
3819         spin_unlock(&cur_trans->dirty_bgs_lock);
3820
3821         while (!list_empty(io)) {
3822                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3823                                          io_list);
3824                 list_del_init(&cache->io_list);
3825                 btrfs_wait_cache_io(trans, cache, path);
3826                 btrfs_put_block_group(cache);
3827         }
3828
3829         btrfs_free_path(path);
3830         return ret;
3831 }
3832
3833 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3834 {
3835         struct btrfs_block_group_cache *block_group;
3836         int readonly = 0;
3837
3838         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3839         if (!block_group || block_group->ro)
3840                 readonly = 1;
3841         if (block_group)
3842                 btrfs_put_block_group(block_group);
3843         return readonly;
3844 }
3845
3846 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3847 {
3848         struct btrfs_block_group_cache *bg;
3849         bool ret = true;
3850
3851         bg = btrfs_lookup_block_group(fs_info, bytenr);
3852         if (!bg)
3853                 return false;
3854
3855         spin_lock(&bg->lock);
3856         if (bg->ro)
3857                 ret = false;
3858         else
3859                 atomic_inc(&bg->nocow_writers);
3860         spin_unlock(&bg->lock);
3861
3862         /* no put on block group, done by btrfs_dec_nocow_writers */
3863         if (!ret)
3864                 btrfs_put_block_group(bg);
3865
3866         return ret;
3867
3868 }
3869
3870 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3871 {
3872         struct btrfs_block_group_cache *bg;
3873
3874         bg = btrfs_lookup_block_group(fs_info, bytenr);
3875         ASSERT(bg);
3876         if (atomic_dec_and_test(&bg->nocow_writers))
3877                 wake_up_atomic_t(&bg->nocow_writers);
3878         /*
3879          * Once for our lookup and once for the lookup done by a previous call
3880          * to btrfs_inc_nocow_writers()
3881          */
3882         btrfs_put_block_group(bg);
3883         btrfs_put_block_group(bg);
3884 }
3885
3886 static int btrfs_wait_nocow_writers_atomic_t(atomic_t *a)
3887 {
3888         schedule();
3889         return 0;
3890 }
3891
3892 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3893 {
3894         wait_on_atomic_t(&bg->nocow_writers,
3895                          btrfs_wait_nocow_writers_atomic_t,
3896                          TASK_UNINTERRUPTIBLE);
3897 }
3898
3899 static const char *alloc_name(u64 flags)
3900 {
3901         switch (flags) {
3902         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3903                 return "mixed";
3904         case BTRFS_BLOCK_GROUP_METADATA:
3905                 return "metadata";
3906         case BTRFS_BLOCK_GROUP_DATA:
3907                 return "data";
3908         case BTRFS_BLOCK_GROUP_SYSTEM:
3909                 return "system";
3910         default:
3911                 WARN_ON(1);
3912                 return "invalid-combination";
3913         };
3914 }
3915
3916 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3917                              u64 total_bytes, u64 bytes_used,
3918                              u64 bytes_readonly,
3919                              struct btrfs_space_info **space_info)
3920 {
3921         struct btrfs_space_info *found;
3922         int i;
3923         int factor;
3924         int ret;
3925
3926         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3927                      BTRFS_BLOCK_GROUP_RAID10))
3928                 factor = 2;
3929         else
3930                 factor = 1;
3931
3932         found = __find_space_info(info, flags);
3933         if (found) {
3934                 spin_lock(&found->lock);
3935                 found->total_bytes += total_bytes;
3936                 found->disk_total += total_bytes * factor;
3937                 found->bytes_used += bytes_used;
3938                 found->disk_used += bytes_used * factor;
3939                 found->bytes_readonly += bytes_readonly;
3940                 if (total_bytes > 0)
3941                         found->full = 0;
3942                 space_info_add_new_bytes(info, found, total_bytes -
3943                                          bytes_used - bytes_readonly);
3944                 spin_unlock(&found->lock);
3945                 *space_info = found;
3946                 return 0;
3947         }
3948         found = kzalloc(sizeof(*found), GFP_NOFS);
3949         if (!found)
3950                 return -ENOMEM;
3951
3952         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3953         if (ret) {
3954                 kfree(found);
3955                 return ret;
3956         }
3957
3958         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3959                 INIT_LIST_HEAD(&found->block_groups[i]);
3960         init_rwsem(&found->groups_sem);
3961         spin_lock_init(&found->lock);
3962         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3963         found->total_bytes = total_bytes;
3964         found->disk_total = total_bytes * factor;
3965         found->bytes_used = bytes_used;
3966         found->disk_used = bytes_used * factor;
3967         found->bytes_pinned = 0;
3968         found->bytes_reserved = 0;
3969         found->bytes_readonly = bytes_readonly;
3970         found->bytes_may_use = 0;
3971         found->full = 0;
3972         found->max_extent_size = 0;
3973         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3974         found->chunk_alloc = 0;
3975         found->flush = 0;
3976         init_waitqueue_head(&found->wait);
3977         INIT_LIST_HEAD(&found->ro_bgs);
3978         INIT_LIST_HEAD(&found->tickets);
3979         INIT_LIST_HEAD(&found->priority_tickets);
3980
3981         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3982                                     info->space_info_kobj, "%s",
3983                                     alloc_name(found->flags));
3984         if (ret) {
3985                 kfree(found);
3986                 return ret;
3987         }
3988
3989         *space_info = found;
3990         list_add_rcu(&found->list, &info->space_info);
3991         if (flags & BTRFS_BLOCK_GROUP_DATA)
3992                 info->data_sinfo = found;
3993
3994         return ret;
3995 }
3996
3997 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3998 {
3999         u64 extra_flags = chunk_to_extended(flags) &
4000                                 BTRFS_EXTENDED_PROFILE_MASK;
4001
4002         write_seqlock(&fs_info->profiles_lock);
4003         if (flags & BTRFS_BLOCK_GROUP_DATA)
4004                 fs_info->avail_data_alloc_bits |= extra_flags;
4005         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4006                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4007         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4008                 fs_info->avail_system_alloc_bits |= extra_flags;
4009         write_sequnlock(&fs_info->profiles_lock);
4010 }
4011
4012 /*
4013  * returns target flags in extended format or 0 if restripe for this
4014  * chunk_type is not in progress
4015  *
4016  * should be called with either volume_mutex or balance_lock held
4017  */
4018 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4019 {
4020         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4021         u64 target = 0;
4022
4023         if (!bctl)
4024                 return 0;
4025
4026         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4027             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4028                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4029         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4030                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4031                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4032         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4033                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4034                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4035         }
4036
4037         return target;
4038 }
4039
4040 /*
4041  * @flags: available profiles in extended format (see ctree.h)
4042  *
4043  * Returns reduced profile in chunk format.  If profile changing is in
4044  * progress (either running or paused) picks the target profile (if it's
4045  * already available), otherwise falls back to plain reducing.
4046  */
4047 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4048 {
4049         u64 num_devices = fs_info->fs_devices->rw_devices;
4050         u64 target;
4051         u64 raid_type;
4052         u64 allowed = 0;
4053
4054         /*
4055          * see if restripe for this chunk_type is in progress, if so
4056          * try to reduce to the target profile
4057          */
4058         spin_lock(&fs_info->balance_lock);
4059         target = get_restripe_target(fs_info, flags);
4060         if (target) {
4061                 /* pick target profile only if it's already available */
4062                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4063                         spin_unlock(&fs_info->balance_lock);
4064                         return extended_to_chunk(target);
4065                 }
4066         }
4067         spin_unlock(&fs_info->balance_lock);
4068
4069         /* First, mask out the RAID levels which aren't possible */
4070         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4071                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4072                         allowed |= btrfs_raid_group[raid_type];
4073         }
4074         allowed &= flags;
4075
4076         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4077                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4078         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4079                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4080         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4081                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4082         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4083                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4084         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4085                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4086
4087         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4088
4089         return extended_to_chunk(flags | allowed);
4090 }
4091
4092 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4093 {
4094         unsigned seq;
4095         u64 flags;
4096
4097         do {
4098                 flags = orig_flags;
4099                 seq = read_seqbegin(&fs_info->profiles_lock);
4100
4101                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4102                         flags |= fs_info->avail_data_alloc_bits;
4103                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4104                         flags |= fs_info->avail_system_alloc_bits;
4105                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4106                         flags |= fs_info->avail_metadata_alloc_bits;
4107         } while (read_seqretry(&fs_info->profiles_lock, seq));
4108
4109         return btrfs_reduce_alloc_profile(fs_info, flags);
4110 }
4111
4112 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
4113 {
4114         struct btrfs_fs_info *fs_info = root->fs_info;
4115         u64 flags;
4116         u64 ret;
4117
4118         if (data)
4119                 flags = BTRFS_BLOCK_GROUP_DATA;
4120         else if (root == fs_info->chunk_root)
4121                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4122         else
4123                 flags = BTRFS_BLOCK_GROUP_METADATA;
4124
4125         ret = get_alloc_profile(fs_info, flags);
4126         return ret;
4127 }
4128
4129 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4130                                  bool may_use_included)
4131 {
4132         ASSERT(s_info);
4133         return s_info->bytes_used + s_info->bytes_reserved +
4134                 s_info->bytes_pinned + s_info->bytes_readonly +
4135                 (may_use_included ? s_info->bytes_may_use : 0);
4136 }
4137
4138 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4139 {
4140         struct btrfs_space_info *data_sinfo;
4141         struct btrfs_root *root = inode->root;
4142         struct btrfs_fs_info *fs_info = root->fs_info;
4143         u64 used;
4144         int ret = 0;
4145         int need_commit = 2;
4146         int have_pinned_space;
4147
4148         /* make sure bytes are sectorsize aligned */
4149         bytes = ALIGN(bytes, fs_info->sectorsize);
4150
4151         if (btrfs_is_free_space_inode(inode)) {
4152                 need_commit = 0;
4153                 ASSERT(current->journal_info);
4154         }
4155
4156         data_sinfo = fs_info->data_sinfo;
4157         if (!data_sinfo)
4158                 goto alloc;
4159
4160 again:
4161         /* make sure we have enough space to handle the data first */
4162         spin_lock(&data_sinfo->lock);
4163         used = btrfs_space_info_used(data_sinfo, true);
4164
4165         if (used + bytes > data_sinfo->total_bytes) {
4166                 struct btrfs_trans_handle *trans;
4167
4168                 /*
4169                  * if we don't have enough free bytes in this space then we need
4170                  * to alloc a new chunk.
4171                  */
4172                 if (!data_sinfo->full) {
4173                         u64 alloc_target;
4174
4175                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4176                         spin_unlock(&data_sinfo->lock);
4177 alloc:
4178                         alloc_target = btrfs_get_alloc_profile(root, 1);
4179                         /*
4180                          * It is ugly that we don't call nolock join
4181                          * transaction for the free space inode case here.
4182                          * But it is safe because we only do the data space
4183                          * reservation for the free space cache in the
4184                          * transaction context, the common join transaction
4185                          * just increase the counter of the current transaction
4186                          * handler, doesn't try to acquire the trans_lock of
4187                          * the fs.
4188                          */
4189                         trans = btrfs_join_transaction(root);
4190                         if (IS_ERR(trans))
4191                                 return PTR_ERR(trans);
4192
4193                         ret = do_chunk_alloc(trans, fs_info, alloc_target,
4194                                              CHUNK_ALLOC_NO_FORCE);
4195                         btrfs_end_transaction(trans);
4196                         if (ret < 0) {
4197                                 if (ret != -ENOSPC)
4198                                         return ret;
4199                                 else {
4200                                         have_pinned_space = 1;
4201                                         goto commit_trans;
4202                                 }
4203                         }
4204
4205                         if (!data_sinfo)
4206                                 data_sinfo = fs_info->data_sinfo;
4207
4208                         goto again;
4209                 }
4210
4211                 /*
4212                  * If we don't have enough pinned space to deal with this
4213                  * allocation, and no removed chunk in current transaction,
4214                  * don't bother committing the transaction.
4215                  */
4216                 have_pinned_space = percpu_counter_compare(
4217                         &data_sinfo->total_bytes_pinned,
4218                         used + bytes - data_sinfo->total_bytes);
4219                 spin_unlock(&data_sinfo->lock);
4220
4221                 /* commit the current transaction and try again */
4222 commit_trans:
4223                 if (need_commit &&
4224                     !atomic_read(&fs_info->open_ioctl_trans)) {
4225                         need_commit--;
4226
4227                         if (need_commit > 0) {
4228                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4229                                 btrfs_wait_ordered_roots(fs_info, -1, 0,
4230                                                          (u64)-1);
4231                         }
4232
4233                         trans = btrfs_join_transaction(root);
4234                         if (IS_ERR(trans))
4235                                 return PTR_ERR(trans);
4236                         if (have_pinned_space >= 0 ||
4237                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4238                                      &trans->transaction->flags) ||
4239                             need_commit > 0) {
4240                                 ret = btrfs_commit_transaction(trans);
4241                                 if (ret)
4242                                         return ret;
4243                                 /*
4244                                  * The cleaner kthread might still be doing iput
4245                                  * operations. Wait for it to finish so that
4246                                  * more space is released.
4247                                  */
4248                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4249                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4250                                 goto again;
4251                         } else {
4252                                 btrfs_end_transaction(trans);
4253                         }
4254                 }
4255
4256                 trace_btrfs_space_reservation(fs_info,
4257                                               "space_info:enospc",
4258                                               data_sinfo->flags, bytes, 1);
4259                 return -ENOSPC;
4260         }
4261         data_sinfo->bytes_may_use += bytes;
4262         trace_btrfs_space_reservation(fs_info, "space_info",
4263                                       data_sinfo->flags, bytes, 1);
4264         spin_unlock(&data_sinfo->lock);
4265
4266         return ret;
4267 }
4268
4269 /*
4270  * New check_data_free_space() with ability for precious data reservation
4271  * Will replace old btrfs_check_data_free_space(), but for patch split,
4272  * add a new function first and then replace it.
4273  */
4274 int btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4275 {
4276         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4277         int ret;
4278
4279         /* align the range */
4280         len = round_up(start + len, fs_info->sectorsize) -
4281               round_down(start, fs_info->sectorsize);
4282         start = round_down(start, fs_info->sectorsize);
4283
4284         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4285         if (ret < 0)
4286                 return ret;
4287
4288         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4289         ret = btrfs_qgroup_reserve_data(inode, start, len);
4290         if (ret)
4291                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4292         return ret;
4293 }
4294
4295 /*
4296  * Called if we need to clear a data reservation for this inode
4297  * Normally in a error case.
4298  *
4299  * This one will *NOT* use accurate qgroup reserved space API, just for case
4300  * which we can't sleep and is sure it won't affect qgroup reserved space.
4301  * Like clear_bit_hook().
4302  */
4303 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4304                                             u64 len)
4305 {
4306         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4307         struct btrfs_space_info *data_sinfo;
4308
4309         /* Make sure the range is aligned to sectorsize */
4310         len = round_up(start + len, fs_info->sectorsize) -
4311               round_down(start, fs_info->sectorsize);
4312         start = round_down(start, fs_info->sectorsize);
4313
4314         data_sinfo = fs_info->data_sinfo;
4315         spin_lock(&data_sinfo->lock);
4316         if (WARN_ON(data_sinfo->bytes_may_use < len))
4317                 data_sinfo->bytes_may_use = 0;
4318         else
4319                 data_sinfo->bytes_may_use -= len;
4320         trace_btrfs_space_reservation(fs_info, "space_info",
4321                                       data_sinfo->flags, len, 0);
4322         spin_unlock(&data_sinfo->lock);
4323 }
4324
4325 /*
4326  * Called if we need to clear a data reservation for this inode
4327  * Normally in a error case.
4328  *
4329  * This one will handle the per-inode data rsv map for accurate reserved
4330  * space framework.
4331  */
4332 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4333 {
4334         struct btrfs_root *root = BTRFS_I(inode)->root;
4335
4336         /* Make sure the range is aligned to sectorsize */
4337         len = round_up(start + len, root->fs_info->sectorsize) -
4338               round_down(start, root->fs_info->sectorsize);
4339         start = round_down(start, root->fs_info->sectorsize);
4340
4341         btrfs_free_reserved_data_space_noquota(inode, start, len);
4342         btrfs_qgroup_free_data(inode, start, len);
4343 }
4344
4345 static void force_metadata_allocation(struct btrfs_fs_info *info)
4346 {
4347         struct list_head *head = &info->space_info;
4348         struct btrfs_space_info *found;
4349
4350         rcu_read_lock();
4351         list_for_each_entry_rcu(found, head, list) {
4352                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4353                         found->force_alloc = CHUNK_ALLOC_FORCE;
4354         }
4355         rcu_read_unlock();
4356 }
4357
4358 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4359 {
4360         return (global->size << 1);
4361 }
4362
4363 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4364                               struct btrfs_space_info *sinfo, int force)
4365 {
4366         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4367         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4368         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4369         u64 thresh;
4370
4371         if (force == CHUNK_ALLOC_FORCE)
4372                 return 1;
4373
4374         /*
4375          * We need to take into account the global rsv because for all intents
4376          * and purposes it's used space.  Don't worry about locking the
4377          * global_rsv, it doesn't change except when the transaction commits.
4378          */
4379         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4380                 num_allocated += calc_global_rsv_need_space(global_rsv);
4381
4382         /*
4383          * in limited mode, we want to have some free space up to
4384          * about 1% of the FS size.
4385          */
4386         if (force == CHUNK_ALLOC_LIMITED) {
4387                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4388                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4389
4390                 if (num_bytes - num_allocated < thresh)
4391                         return 1;
4392         }
4393
4394         if (num_allocated + SZ_2M < div_factor(num_bytes, 8))
4395                 return 0;
4396         return 1;
4397 }
4398
4399 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4400 {
4401         u64 num_dev;
4402
4403         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4404                     BTRFS_BLOCK_GROUP_RAID0 |
4405                     BTRFS_BLOCK_GROUP_RAID5 |
4406                     BTRFS_BLOCK_GROUP_RAID6))
4407                 num_dev = fs_info->fs_devices->rw_devices;
4408         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4409                 num_dev = 2;
4410         else
4411                 num_dev = 1;    /* DUP or single */
4412
4413         return num_dev;
4414 }
4415
4416 /*
4417  * If @is_allocation is true, reserve space in the system space info necessary
4418  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4419  * removing a chunk.
4420  */
4421 void check_system_chunk(struct btrfs_trans_handle *trans,
4422                         struct btrfs_fs_info *fs_info, u64 type)
4423 {
4424         struct btrfs_space_info *info;
4425         u64 left;
4426         u64 thresh;
4427         int ret = 0;
4428         u64 num_devs;
4429
4430         /*
4431          * Needed because we can end up allocating a system chunk and for an
4432          * atomic and race free space reservation in the chunk block reserve.
4433          */
4434         ASSERT(mutex_is_locked(&fs_info->chunk_mutex));
4435
4436         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4437         spin_lock(&info->lock);
4438         left = info->total_bytes - btrfs_space_info_used(info, true);
4439         spin_unlock(&info->lock);
4440
4441         num_devs = get_profile_num_devs(fs_info, type);
4442
4443         /* num_devs device items to update and 1 chunk item to add or remove */
4444         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4445                 btrfs_calc_trans_metadata_size(fs_info, 1);
4446
4447         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4448                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4449                            left, thresh, type);
4450                 dump_space_info(fs_info, info, 0, 0);
4451         }
4452
4453         if (left < thresh) {
4454                 u64 flags;
4455
4456                 flags = btrfs_get_alloc_profile(fs_info->chunk_root, 0);
4457                 /*
4458                  * Ignore failure to create system chunk. We might end up not
4459                  * needing it, as we might not need to COW all nodes/leafs from
4460                  * the paths we visit in the chunk tree (they were already COWed
4461                  * or created in the current transaction for example).
4462                  */
4463                 ret = btrfs_alloc_chunk(trans, fs_info, flags);
4464         }
4465
4466         if (!ret) {
4467                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4468                                           &fs_info->chunk_block_rsv,
4469                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4470                 if (!ret)
4471                         trans->chunk_bytes_reserved += thresh;
4472         }
4473 }
4474
4475 /*
4476  * If force is CHUNK_ALLOC_FORCE:
4477  *    - return 1 if it successfully allocates a chunk,
4478  *    - return errors including -ENOSPC otherwise.
4479  * If force is NOT CHUNK_ALLOC_FORCE:
4480  *    - return 0 if it doesn't need to allocate a new chunk,
4481  *    - return 1 if it successfully allocates a chunk,
4482  *    - return errors including -ENOSPC otherwise.
4483  */
4484 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4485                           struct btrfs_fs_info *fs_info, u64 flags, int force)
4486 {
4487         struct btrfs_space_info *space_info;
4488         int wait_for_alloc = 0;
4489         int ret = 0;
4490
4491         /* Don't re-enter if we're already allocating a chunk */
4492         if (trans->allocating_chunk)
4493                 return -ENOSPC;
4494
4495         space_info = __find_space_info(fs_info, flags);
4496         if (!space_info) {
4497                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
4498                 BUG_ON(ret); /* -ENOMEM */
4499         }
4500         BUG_ON(!space_info); /* Logic error */
4501
4502 again:
4503         spin_lock(&space_info->lock);
4504         if (force < space_info->force_alloc)
4505                 force = space_info->force_alloc;
4506         if (space_info->full) {
4507                 if (should_alloc_chunk(fs_info, space_info, force))
4508                         ret = -ENOSPC;
4509                 else
4510                         ret = 0;
4511                 spin_unlock(&space_info->lock);
4512                 return ret;
4513         }
4514
4515         if (!should_alloc_chunk(fs_info, space_info, force)) {
4516                 spin_unlock(&space_info->lock);
4517                 return 0;
4518         } else if (space_info->chunk_alloc) {
4519                 wait_for_alloc = 1;
4520         } else {
4521                 space_info->chunk_alloc = 1;
4522         }
4523
4524         spin_unlock(&space_info->lock);
4525
4526         mutex_lock(&fs_info->chunk_mutex);
4527
4528         /*
4529          * The chunk_mutex is held throughout the entirety of a chunk
4530          * allocation, so once we've acquired the chunk_mutex we know that the
4531          * other guy is done and we need to recheck and see if we should
4532          * allocate.
4533          */
4534         if (wait_for_alloc) {
4535                 mutex_unlock(&fs_info->chunk_mutex);
4536                 wait_for_alloc = 0;
4537                 goto again;
4538         }
4539
4540         trans->allocating_chunk = true;
4541
4542         /*
4543          * If we have mixed data/metadata chunks we want to make sure we keep
4544          * allocating mixed chunks instead of individual chunks.
4545          */
4546         if (btrfs_mixed_space_info(space_info))
4547                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4548
4549         /*
4550          * if we're doing a data chunk, go ahead and make sure that
4551          * we keep a reasonable number of metadata chunks allocated in the
4552          * FS as well.
4553          */
4554         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4555                 fs_info->data_chunk_allocations++;
4556                 if (!(fs_info->data_chunk_allocations %
4557                       fs_info->metadata_ratio))
4558                         force_metadata_allocation(fs_info);
4559         }
4560
4561         /*
4562          * Check if we have enough space in SYSTEM chunk because we may need
4563          * to update devices.
4564          */
4565         check_system_chunk(trans, fs_info, flags);
4566
4567         ret = btrfs_alloc_chunk(trans, fs_info, flags);
4568         trans->allocating_chunk = false;
4569
4570         spin_lock(&space_info->lock);
4571         if (ret < 0 && ret != -ENOSPC)
4572                 goto out;
4573         if (ret)
4574                 space_info->full = 1;
4575         else
4576                 ret = 1;
4577
4578         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4579 out:
4580         space_info->chunk_alloc = 0;
4581         spin_unlock(&space_info->lock);
4582         mutex_unlock(&fs_info->chunk_mutex);
4583         /*
4584          * When we allocate a new chunk we reserve space in the chunk block
4585          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4586          * add new nodes/leafs to it if we end up needing to do it when
4587          * inserting the chunk item and updating device items as part of the
4588          * second phase of chunk allocation, performed by
4589          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4590          * large number of new block groups to create in our transaction
4591          * handle's new_bgs list to avoid exhausting the chunk block reserve
4592          * in extreme cases - like having a single transaction create many new
4593          * block groups when starting to write out the free space caches of all
4594          * the block groups that were made dirty during the lifetime of the
4595          * transaction.
4596          */
4597         if (trans->can_flush_pending_bgs &&
4598             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4599                 btrfs_create_pending_block_groups(trans, fs_info);
4600                 btrfs_trans_release_chunk_metadata(trans);
4601         }
4602         return ret;
4603 }
4604
4605 static int can_overcommit(struct btrfs_root *root,
4606                           struct btrfs_space_info *space_info, u64 bytes,
4607                           enum btrfs_reserve_flush_enum flush)
4608 {
4609         struct btrfs_fs_info *fs_info = root->fs_info;
4610         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4611         u64 profile;
4612         u64 space_size;
4613         u64 avail;
4614         u64 used;
4615
4616         /* Don't overcommit when in mixed mode. */
4617         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4618                 return 0;
4619
4620         profile = btrfs_get_alloc_profile(root, 0);
4621         used = btrfs_space_info_used(space_info, false);
4622
4623         /*
4624          * We only want to allow over committing if we have lots of actual space
4625          * free, but if we don't have enough space to handle the global reserve
4626          * space then we could end up having a real enospc problem when trying
4627          * to allocate a chunk or some other such important allocation.
4628          */
4629         spin_lock(&global_rsv->lock);
4630         space_size = calc_global_rsv_need_space(global_rsv);
4631         spin_unlock(&global_rsv->lock);
4632         if (used + space_size >= space_info->total_bytes)
4633                 return 0;
4634
4635         used += space_info->bytes_may_use;
4636
4637         spin_lock(&fs_info->free_chunk_lock);
4638         avail = fs_info->free_chunk_space;
4639         spin_unlock(&fs_info->free_chunk_lock);
4640
4641         /*
4642          * If we have dup, raid1 or raid10 then only half of the free
4643          * space is actually useable.  For raid56, the space info used
4644          * doesn't include the parity drive, so we don't have to
4645          * change the math
4646          */
4647         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4648                        BTRFS_BLOCK_GROUP_RAID1 |
4649                        BTRFS_BLOCK_GROUP_RAID10))
4650                 avail >>= 1;
4651
4652         /*
4653          * If we aren't flushing all things, let us overcommit up to
4654          * 1/2th of the space. If we can flush, don't let us overcommit
4655          * too much, let it overcommit up to 1/8 of the space.
4656          */
4657         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4658                 avail >>= 3;
4659         else
4660                 avail >>= 1;
4661
4662         if (used + bytes < space_info->total_bytes + avail)
4663                 return 1;
4664         return 0;
4665 }
4666
4667 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4668                                          unsigned long nr_pages, int nr_items)
4669 {
4670         struct super_block *sb = fs_info->sb;
4671
4672         if (down_read_trylock(&sb->s_umount)) {
4673                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4674                 up_read(&sb->s_umount);
4675         } else {
4676                 /*
4677                  * We needn't worry the filesystem going from r/w to r/o though
4678                  * we don't acquire ->s_umount mutex, because the filesystem
4679                  * should guarantee the delalloc inodes list be empty after
4680                  * the filesystem is readonly(all dirty pages are written to
4681                  * the disk).
4682                  */
4683                 btrfs_start_delalloc_roots(fs_info, 0, nr_items);
4684                 if (!current->journal_info)
4685                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4686         }
4687 }
4688
4689 static inline int calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4690                                         u64 to_reclaim)
4691 {
4692         u64 bytes;
4693         int nr;
4694
4695         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4696         nr = (int)div64_u64(to_reclaim, bytes);
4697         if (!nr)
4698                 nr = 1;
4699         return nr;
4700 }
4701
4702 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4703
4704 /*
4705  * shrink metadata reservation for delalloc
4706  */
4707 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4708                             bool wait_ordered)
4709 {
4710         struct btrfs_fs_info *fs_info = root->fs_info;
4711         struct btrfs_block_rsv *block_rsv;
4712         struct btrfs_space_info *space_info;
4713         struct btrfs_trans_handle *trans;
4714         u64 delalloc_bytes;
4715         u64 max_reclaim;
4716         long time_left;
4717         unsigned long nr_pages;
4718         int loops;
4719         int items;
4720         enum btrfs_reserve_flush_enum flush;
4721
4722         /* Calc the number of the pages we need flush for space reservation */
4723         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4724         to_reclaim = (u64)items * EXTENT_SIZE_PER_ITEM;
4725
4726         trans = (struct btrfs_trans_handle *)current->journal_info;
4727         block_rsv = &fs_info->delalloc_block_rsv;
4728         space_info = block_rsv->space_info;
4729
4730         delalloc_bytes = percpu_counter_sum_positive(
4731                                                 &fs_info->delalloc_bytes);
4732         if (delalloc_bytes == 0) {
4733                 if (trans)
4734                         return;
4735                 if (wait_ordered)
4736                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4737                 return;
4738         }
4739
4740         loops = 0;
4741         while (delalloc_bytes && loops < 3) {
4742                 max_reclaim = min(delalloc_bytes, to_reclaim);
4743                 nr_pages = max_reclaim >> PAGE_SHIFT;
4744                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4745                 /*
4746                  * We need to wait for the async pages to actually start before
4747                  * we do anything.
4748                  */
4749                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4750                 if (!max_reclaim)
4751                         goto skip_async;
4752
4753                 if (max_reclaim <= nr_pages)
4754                         max_reclaim = 0;
4755                 else
4756                         max_reclaim -= nr_pages;
4757
4758                 wait_event(fs_info->async_submit_wait,
4759                            atomic_read(&fs_info->async_delalloc_pages) <=
4760                            (int)max_reclaim);
4761 skip_async:
4762                 if (!trans)
4763                         flush = BTRFS_RESERVE_FLUSH_ALL;
4764                 else
4765                         flush = BTRFS_RESERVE_NO_FLUSH;
4766                 spin_lock(&space_info->lock);
4767                 if (can_overcommit(root, space_info, orig, flush)) {
4768                         spin_unlock(&space_info->lock);
4769                         break;
4770                 }
4771                 if (list_empty(&space_info->tickets) &&
4772                     list_empty(&space_info->priority_tickets)) {
4773                         spin_unlock(&space_info->lock);
4774                         break;
4775                 }
4776                 spin_unlock(&space_info->lock);
4777
4778                 loops++;
4779                 if (wait_ordered && !trans) {
4780                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4781                 } else {
4782                         time_left = schedule_timeout_killable(1);
4783                         if (time_left)
4784                                 break;
4785                 }
4786                 delalloc_bytes = percpu_counter_sum_positive(
4787                                                 &fs_info->delalloc_bytes);
4788         }
4789 }
4790
4791 /**
4792  * maybe_commit_transaction - possibly commit the transaction if its ok to
4793  * @root - the root we're allocating for
4794  * @bytes - the number of bytes we want to reserve
4795  * @force - force the commit
4796  *
4797  * This will check to make sure that committing the transaction will actually
4798  * get us somewhere and then commit the transaction if it does.  Otherwise it
4799  * will return -ENOSPC.
4800  */
4801 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4802                                   struct btrfs_space_info *space_info,
4803                                   u64 bytes, int force)
4804 {
4805         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4806         struct btrfs_trans_handle *trans;
4807
4808         trans = (struct btrfs_trans_handle *)current->journal_info;
4809         if (trans)
4810                 return -EAGAIN;
4811
4812         if (force)
4813                 goto commit;
4814
4815         /* See if there is enough pinned space to make this reservation */
4816         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4817                                    bytes) >= 0)
4818                 goto commit;
4819
4820         /*
4821          * See if there is some space in the delayed insertion reservation for
4822          * this reservation.
4823          */
4824         if (space_info != delayed_rsv->space_info)
4825                 return -ENOSPC;
4826
4827         spin_lock(&delayed_rsv->lock);
4828         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4829                                    bytes - delayed_rsv->size) >= 0) {
4830                 spin_unlock(&delayed_rsv->lock);
4831                 return -ENOSPC;
4832         }
4833         spin_unlock(&delayed_rsv->lock);
4834
4835 commit:
4836         trans = btrfs_join_transaction(fs_info->fs_root);
4837         if (IS_ERR(trans))
4838                 return -ENOSPC;
4839
4840         return btrfs_commit_transaction(trans);
4841 }
4842
4843 struct reserve_ticket {
4844         u64 bytes;
4845         int error;
4846         struct list_head list;
4847         wait_queue_head_t wait;
4848 };
4849
4850 static int flush_space(struct btrfs_fs_info *fs_info,
4851                        struct btrfs_space_info *space_info, u64 num_bytes,
4852                        u64 orig_bytes, int state)
4853 {
4854         struct btrfs_root *root = fs_info->fs_root;
4855         struct btrfs_trans_handle *trans;
4856         int nr;
4857         int ret = 0;
4858
4859         switch (state) {
4860         case FLUSH_DELAYED_ITEMS_NR:
4861         case FLUSH_DELAYED_ITEMS:
4862                 if (state == FLUSH_DELAYED_ITEMS_NR)
4863                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4864                 else
4865                         nr = -1;
4866
4867                 trans = btrfs_join_transaction(root);
4868                 if (IS_ERR(trans)) {
4869                         ret = PTR_ERR(trans);
4870                         break;
4871                 }
4872                 ret = btrfs_run_delayed_items_nr(trans, fs_info, nr);
4873                 btrfs_end_transaction(trans);
4874                 break;
4875         case FLUSH_DELALLOC:
4876         case FLUSH_DELALLOC_WAIT:
4877                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4878                                 state == FLUSH_DELALLOC_WAIT);
4879                 break;
4880         case ALLOC_CHUNK:
4881                 trans = btrfs_join_transaction(root);
4882                 if (IS_ERR(trans)) {
4883                         ret = PTR_ERR(trans);
4884                         break;
4885                 }
4886                 ret = do_chunk_alloc(trans, fs_info,
4887                                      btrfs_get_alloc_profile(root, 0),
4888                                      CHUNK_ALLOC_NO_FORCE);
4889                 btrfs_end_transaction(trans);
4890                 if (ret > 0 || ret == -ENOSPC)
4891                         ret = 0;
4892                 break;
4893         case COMMIT_TRANS:
4894                 ret = may_commit_transaction(fs_info, space_info,
4895                                              orig_bytes, 0);
4896                 break;
4897         default:
4898                 ret = -ENOSPC;
4899                 break;
4900         }
4901
4902         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes,
4903                                 orig_bytes, state, ret);
4904         return ret;
4905 }
4906
4907 static inline u64
4908 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4909                                  struct btrfs_space_info *space_info)
4910 {
4911         struct reserve_ticket *ticket;
4912         u64 used;
4913         u64 expected;
4914         u64 to_reclaim = 0;
4915
4916         list_for_each_entry(ticket, &space_info->tickets, list)
4917                 to_reclaim += ticket->bytes;
4918         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4919                 to_reclaim += ticket->bytes;
4920         if (to_reclaim)
4921                 return to_reclaim;
4922
4923         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4924         if (can_overcommit(root, space_info, to_reclaim,
4925                            BTRFS_RESERVE_FLUSH_ALL))
4926                 return 0;
4927
4928         used = space_info->bytes_used + space_info->bytes_reserved +
4929                space_info->bytes_pinned + space_info->bytes_readonly +
4930                space_info->bytes_may_use;
4931         if (can_overcommit(root, space_info, SZ_1M, BTRFS_RESERVE_FLUSH_ALL))
4932                 expected = div_factor_fine(space_info->total_bytes, 95);
4933         else
4934                 expected = div_factor_fine(space_info->total_bytes, 90);
4935
4936         if (used > expected)
4937                 to_reclaim = used - expected;
4938         else
4939                 to_reclaim = 0;
4940         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4941                                      space_info->bytes_reserved);
4942         return to_reclaim;
4943 }
4944
4945 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4946                                         struct btrfs_root *root, u64 used)
4947 {
4948         struct btrfs_fs_info *fs_info = root->fs_info;
4949         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4950
4951         /* If we're just plain full then async reclaim just slows us down. */
4952         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4953                 return 0;
4954
4955         if (!btrfs_calc_reclaim_metadata_size(root, space_info))
4956                 return 0;
4957
4958         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4959                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4960 }
4961
4962 static void wake_all_tickets(struct list_head *head)
4963 {
4964         struct reserve_ticket *ticket;
4965
4966         while (!list_empty(head)) {
4967                 ticket = list_first_entry(head, struct reserve_ticket, list);
4968                 list_del_init(&ticket->list);
4969                 ticket->error = -ENOSPC;
4970                 wake_up(&ticket->wait);
4971         }
4972 }
4973
4974 /*
4975  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4976  * will loop and continuously try to flush as long as we are making progress.
4977  * We count progress as clearing off tickets each time we have to loop.
4978  */
4979 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4980 {
4981         struct btrfs_fs_info *fs_info;
4982         struct btrfs_space_info *space_info;
4983         u64 to_reclaim;
4984         int flush_state;
4985         int commit_cycles = 0;
4986         u64 last_tickets_id;
4987
4988         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4989         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4990
4991         spin_lock(&space_info->lock);
4992         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4993                                                       space_info);
4994         if (!to_reclaim) {
4995                 space_info->flush = 0;
4996                 spin_unlock(&space_info->lock);
4997                 return;
4998         }
4999         last_tickets_id = space_info->tickets_id;
5000         spin_unlock(&space_info->lock);
5001
5002         flush_state = FLUSH_DELAYED_ITEMS_NR;
5003         do {
5004                 struct reserve_ticket *ticket;
5005                 int ret;
5006
5007                 ret = flush_space(fs_info, space_info, to_reclaim, to_reclaim,
5008                                   flush_state);
5009                 spin_lock(&space_info->lock);
5010                 if (list_empty(&space_info->tickets)) {
5011                         space_info->flush = 0;
5012                         spin_unlock(&space_info->lock);
5013                         return;
5014                 }
5015                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5016                                                               space_info);
5017                 ticket = list_first_entry(&space_info->tickets,
5018                                           struct reserve_ticket, list);
5019                 if (last_tickets_id == space_info->tickets_id) {
5020                         flush_state++;
5021                 } else {
5022                         last_tickets_id = space_info->tickets_id;
5023                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5024                         if (commit_cycles)
5025                                 commit_cycles--;
5026                 }
5027
5028                 if (flush_state > COMMIT_TRANS) {
5029                         commit_cycles++;
5030                         if (commit_cycles > 2) {
5031                                 wake_all_tickets(&space_info->tickets);
5032                                 space_info->flush = 0;
5033                         } else {
5034                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5035                         }
5036                 }
5037                 spin_unlock(&space_info->lock);
5038         } while (flush_state <= COMMIT_TRANS);
5039 }
5040
5041 void btrfs_init_async_reclaim_work(struct work_struct *work)
5042 {
5043         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5044 }
5045
5046 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5047                                             struct btrfs_space_info *space_info,
5048                                             struct reserve_ticket *ticket)
5049 {
5050         u64 to_reclaim;
5051         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5052
5053         spin_lock(&space_info->lock);
5054         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5055                                                       space_info);
5056         if (!to_reclaim) {
5057                 spin_unlock(&space_info->lock);
5058                 return;
5059         }
5060         spin_unlock(&space_info->lock);
5061
5062         do {
5063                 flush_space(fs_info, space_info, to_reclaim, to_reclaim,
5064                             flush_state);
5065                 flush_state++;
5066                 spin_lock(&space_info->lock);
5067                 if (ticket->bytes == 0) {
5068                         spin_unlock(&space_info->lock);
5069                         return;
5070                 }
5071                 spin_unlock(&space_info->lock);
5072
5073                 /*
5074                  * Priority flushers can't wait on delalloc without
5075                  * deadlocking.
5076                  */
5077                 if (flush_state == FLUSH_DELALLOC ||
5078                     flush_state == FLUSH_DELALLOC_WAIT)
5079                         flush_state = ALLOC_CHUNK;
5080         } while (flush_state < COMMIT_TRANS);
5081 }
5082
5083 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5084                                struct btrfs_space_info *space_info,
5085                                struct reserve_ticket *ticket, u64 orig_bytes)
5086
5087 {
5088         DEFINE_WAIT(wait);
5089         int ret = 0;
5090
5091         spin_lock(&space_info->lock);
5092         while (ticket->bytes > 0 && ticket->error == 0) {
5093                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5094                 if (ret) {
5095                         ret = -EINTR;
5096                         break;
5097                 }
5098                 spin_unlock(&space_info->lock);
5099
5100                 schedule();
5101
5102                 finish_wait(&ticket->wait, &wait);
5103                 spin_lock(&space_info->lock);
5104         }
5105         if (!ret)
5106                 ret = ticket->error;
5107         if (!list_empty(&ticket->list))
5108                 list_del_init(&ticket->list);
5109         if (ticket->bytes && ticket->bytes < orig_bytes) {
5110                 u64 num_bytes = orig_bytes - ticket->bytes;
5111                 space_info->bytes_may_use -= num_bytes;
5112                 trace_btrfs_space_reservation(fs_info, "space_info",
5113                                               space_info->flags, num_bytes, 0);
5114         }
5115         spin_unlock(&space_info->lock);
5116
5117         return ret;
5118 }
5119
5120 /**
5121  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5122  * @root - the root we're allocating for
5123  * @space_info - the space info we want to allocate from
5124  * @orig_bytes - the number of bytes we want
5125  * @flush - whether or not we can flush to make our reservation
5126  *
5127  * This will reserve orig_bytes number of bytes from the space info associated
5128  * with the block_rsv.  If there is not enough space it will make an attempt to
5129  * flush out space to make room.  It will do this by flushing delalloc if
5130  * possible or committing the transaction.  If flush is 0 then no attempts to
5131  * regain reservations will be made and this will fail if there is not enough
5132  * space already.
5133  */
5134 static int __reserve_metadata_bytes(struct btrfs_root *root,
5135                                     struct btrfs_space_info *space_info,
5136                                     u64 orig_bytes,
5137                                     enum btrfs_reserve_flush_enum flush)
5138 {
5139         struct btrfs_fs_info *fs_info = root->fs_info;
5140         struct reserve_ticket ticket;
5141         u64 used;
5142         int ret = 0;
5143
5144         ASSERT(orig_bytes);
5145         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5146
5147         spin_lock(&space_info->lock);
5148         ret = -ENOSPC;
5149         used = btrfs_space_info_used(space_info, true);
5150
5151         /*
5152          * If we have enough space then hooray, make our reservation and carry
5153          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5154          * If not things get more complicated.
5155          */
5156         if (used + orig_bytes <= space_info->total_bytes) {
5157                 space_info->bytes_may_use += orig_bytes;
5158                 trace_btrfs_space_reservation(fs_info, "space_info",
5159                                               space_info->flags, orig_bytes, 1);
5160                 ret = 0;
5161         } else if (can_overcommit(root, space_info, orig_bytes, flush)) {
5162                 space_info->bytes_may_use += orig_bytes;
5163                 trace_btrfs_space_reservation(fs_info, "space_info",
5164                                               space_info->flags, orig_bytes, 1);
5165                 ret = 0;
5166         }
5167
5168         /*
5169          * If we couldn't make a reservation then setup our reservation ticket
5170          * and kick the async worker if it's not already running.
5171          *
5172          * If we are a priority flusher then we just need to add our ticket to
5173          * the list and we will do our own flushing further down.
5174          */
5175         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5176                 ticket.bytes = orig_bytes;
5177                 ticket.error = 0;
5178                 init_waitqueue_head(&ticket.wait);
5179                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5180                         list_add_tail(&ticket.list, &space_info->tickets);
5181                         if (!space_info->flush) {
5182                                 space_info->flush = 1;
5183                                 trace_btrfs_trigger_flush(fs_info,
5184                                                           space_info->flags,
5185                                                           orig_bytes, flush,
5186                                                           "enospc");
5187                                 queue_work(system_unbound_wq,
5188                                            &root->fs_info->async_reclaim_work);
5189                         }
5190                 } else {
5191                         list_add_tail(&ticket.list,
5192                                       &space_info->priority_tickets);
5193                 }
5194         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5195                 used += orig_bytes;
5196                 /*
5197                  * We will do the space reservation dance during log replay,
5198                  * which means we won't have fs_info->fs_root set, so don't do
5199                  * the async reclaim as we will panic.
5200                  */
5201                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5202                     need_do_async_reclaim(space_info, root, used) &&
5203                     !work_busy(&fs_info->async_reclaim_work)) {
5204                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5205                                                   orig_bytes, flush, "preempt");
5206                         queue_work(system_unbound_wq,
5207                                    &fs_info->async_reclaim_work);
5208                 }
5209         }
5210         spin_unlock(&space_info->lock);
5211         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5212                 return ret;
5213
5214         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5215                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5216                                            orig_bytes);
5217
5218         ret = 0;
5219         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5220         spin_lock(&space_info->lock);
5221         if (ticket.bytes) {
5222                 if (ticket.bytes < orig_bytes) {
5223                         u64 num_bytes = orig_bytes - ticket.bytes;
5224                         space_info->bytes_may_use -= num_bytes;
5225                         trace_btrfs_space_reservation(fs_info, "space_info",
5226                                                       space_info->flags,
5227                                                       num_bytes, 0);
5228
5229                 }
5230                 list_del_init(&ticket.list);
5231                 ret = -ENOSPC;
5232         }
5233         spin_unlock(&space_info->lock);
5234         ASSERT(list_empty(&ticket.list));
5235         return ret;
5236 }
5237
5238 /**
5239  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5240  * @root - the root we're allocating for
5241  * @block_rsv - the block_rsv we're allocating for
5242  * @orig_bytes - the number of bytes we want
5243  * @flush - whether or not we can flush to make our reservation
5244  *
5245  * This will reserve orgi_bytes number of bytes from the space info associated
5246  * with the block_rsv.  If there is not enough space it will make an attempt to
5247  * flush out space to make room.  It will do this by flushing delalloc if
5248  * possible or committing the transaction.  If flush is 0 then no attempts to
5249  * regain reservations will be made and this will fail if there is not enough
5250  * space already.
5251  */
5252 static int reserve_metadata_bytes(struct btrfs_root *root,
5253                                   struct btrfs_block_rsv *block_rsv,
5254                                   u64 orig_bytes,
5255                                   enum btrfs_reserve_flush_enum flush)
5256 {
5257         struct btrfs_fs_info *fs_info = root->fs_info;
5258         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5259         int ret;
5260
5261         ret = __reserve_metadata_bytes(root, block_rsv->space_info, orig_bytes,
5262                                        flush);
5263         if (ret == -ENOSPC &&
5264             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5265                 if (block_rsv != global_rsv &&
5266                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5267                         ret = 0;
5268         }
5269         if (ret == -ENOSPC)
5270                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5271                                               block_rsv->space_info->flags,
5272                                               orig_bytes, 1);
5273         return ret;
5274 }
5275
5276 static struct btrfs_block_rsv *get_block_rsv(
5277                                         const struct btrfs_trans_handle *trans,
5278                                         const struct btrfs_root *root)
5279 {
5280         struct btrfs_fs_info *fs_info = root->fs_info;
5281         struct btrfs_block_rsv *block_rsv = NULL;
5282
5283         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5284             (root == fs_info->csum_root && trans->adding_csums) ||
5285             (root == fs_info->uuid_root))
5286                 block_rsv = trans->block_rsv;
5287
5288         if (!block_rsv)
5289                 block_rsv = root->block_rsv;
5290
5291         if (!block_rsv)
5292                 block_rsv = &fs_info->empty_block_rsv;
5293
5294         return block_rsv;
5295 }
5296
5297 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5298                                u64 num_bytes)
5299 {
5300         int ret = -ENOSPC;
5301         spin_lock(&block_rsv->lock);
5302         if (block_rsv->reserved >= num_bytes) {
5303                 block_rsv->reserved -= num_bytes;
5304                 if (block_rsv->reserved < block_rsv->size)
5305                         block_rsv->full = 0;
5306                 ret = 0;
5307         }
5308         spin_unlock(&block_rsv->lock);
5309         return ret;
5310 }
5311
5312 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5313                                 u64 num_bytes, int update_size)
5314 {
5315         spin_lock(&block_rsv->lock);
5316         block_rsv->reserved += num_bytes;
5317         if (update_size)
5318                 block_rsv->size += num_bytes;
5319         else if (block_rsv->reserved >= block_rsv->size)
5320                 block_rsv->full = 1;
5321         spin_unlock(&block_rsv->lock);
5322 }
5323
5324 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5325                              struct btrfs_block_rsv *dest, u64 num_bytes,
5326                              int min_factor)
5327 {
5328         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5329         u64 min_bytes;
5330
5331         if (global_rsv->space_info != dest->space_info)
5332                 return -ENOSPC;
5333
5334         spin_lock(&global_rsv->lock);
5335         min_bytes = div_factor(global_rsv->size, min_factor);
5336         if (global_rsv->reserved < min_bytes + num_bytes) {
5337                 spin_unlock(&global_rsv->lock);
5338                 return -ENOSPC;
5339         }
5340         global_rsv->reserved -= num_bytes;
5341         if (global_rsv->reserved < global_rsv->size)
5342                 global_rsv->full = 0;
5343         spin_unlock(&global_rsv->lock);
5344
5345         block_rsv_add_bytes(dest, num_bytes, 1);
5346         return 0;
5347 }
5348
5349 /*
5350  * This is for space we already have accounted in space_info->bytes_may_use, so
5351  * basically when we're returning space from block_rsv's.
5352  */
5353 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5354                                      struct btrfs_space_info *space_info,
5355                                      u64 num_bytes)
5356 {
5357         struct reserve_ticket *ticket;
5358         struct list_head *head;
5359         u64 used;
5360         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5361         bool check_overcommit = false;
5362
5363         spin_lock(&space_info->lock);
5364         head = &space_info->priority_tickets;
5365
5366         /*
5367          * If we are over our limit then we need to check and see if we can
5368          * overcommit, and if we can't then we just need to free up our space
5369          * and not satisfy any requests.
5370          */
5371         used = space_info->bytes_used + space_info->bytes_reserved +
5372                 space_info->bytes_pinned + space_info->bytes_readonly +
5373                 space_info->bytes_may_use;
5374         if (used - num_bytes >= space_info->total_bytes)
5375                 check_overcommit = true;
5376 again:
5377         while (!list_empty(head) && num_bytes) {
5378                 ticket = list_first_entry(head, struct reserve_ticket,
5379                                           list);
5380                 /*
5381                  * We use 0 bytes because this space is already reserved, so
5382                  * adding the ticket space would be a double count.
5383                  */
5384                 if (check_overcommit &&
5385                     !can_overcommit(fs_info->extent_root, space_info, 0,
5386                                     flush))
5387                         break;
5388                 if (num_bytes >= ticket->bytes) {
5389                         list_del_init(&ticket->list);
5390                         num_bytes -= ticket->bytes;
5391                         ticket->bytes = 0;
5392                         space_info->tickets_id++;
5393                         wake_up(&ticket->wait);
5394                 } else {
5395                         ticket->bytes -= num_bytes;
5396                         num_bytes = 0;
5397                 }
5398         }
5399
5400         if (num_bytes && head == &space_info->priority_tickets) {
5401                 head = &space_info->tickets;
5402                 flush = BTRFS_RESERVE_FLUSH_ALL;
5403                 goto again;
5404         }
5405         space_info->bytes_may_use -= num_bytes;
5406         trace_btrfs_space_reservation(fs_info, "space_info",
5407                                       space_info->flags, num_bytes, 0);
5408         spin_unlock(&space_info->lock);
5409 }
5410
5411 /*
5412  * This is for newly allocated space that isn't accounted in
5413  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5414  * we use this helper.
5415  */
5416 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5417                                      struct btrfs_space_info *space_info,
5418                                      u64 num_bytes)
5419 {
5420         struct reserve_ticket *ticket;
5421         struct list_head *head = &space_info->priority_tickets;
5422
5423 again:
5424         while (!list_empty(head) && num_bytes) {
5425                 ticket = list_first_entry(head, struct reserve_ticket,
5426                                           list);
5427                 if (num_bytes >= ticket->bytes) {
5428                         trace_btrfs_space_reservation(fs_info, "space_info",
5429                                                       space_info->flags,
5430                                                       ticket->bytes, 1);
5431                         list_del_init(&ticket->list);
5432                         num_bytes -= ticket->bytes;
5433                         space_info->bytes_may_use += ticket->bytes;
5434                         ticket->bytes = 0;
5435                         space_info->tickets_id++;
5436                         wake_up(&ticket->wait);
5437                 } else {
5438                         trace_btrfs_space_reservation(fs_info, "space_info",
5439                                                       space_info->flags,
5440                                                       num_bytes, 1);
5441                         space_info->bytes_may_use += num_bytes;
5442                         ticket->bytes -= num_bytes;
5443                         num_bytes = 0;
5444                 }
5445         }
5446
5447         if (num_bytes && head == &space_info->priority_tickets) {
5448                 head = &space_info->tickets;
5449                 goto again;
5450         }
5451 }
5452
5453 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5454                                     struct btrfs_block_rsv *block_rsv,
5455                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5456 {
5457         struct btrfs_space_info *space_info = block_rsv->space_info;
5458
5459         spin_lock(&block_rsv->lock);
5460         if (num_bytes == (u64)-1)
5461                 num_bytes = block_rsv->size;
5462         block_rsv->size -= num_bytes;
5463         if (block_rsv->reserved >= block_rsv->size) {
5464                 num_bytes = block_rsv->reserved - block_rsv->size;
5465                 block_rsv->reserved = block_rsv->size;
5466                 block_rsv->full = 1;
5467         } else {
5468                 num_bytes = 0;
5469         }
5470         spin_unlock(&block_rsv->lock);
5471
5472         if (num_bytes > 0) {
5473                 if (dest) {
5474                         spin_lock(&dest->lock);
5475                         if (!dest->full) {
5476                                 u64 bytes_to_add;
5477
5478                                 bytes_to_add = dest->size - dest->reserved;
5479                                 bytes_to_add = min(num_bytes, bytes_to_add);
5480                                 dest->reserved += bytes_to_add;
5481                                 if (dest->reserved >= dest->size)
5482                                         dest->full = 1;
5483                                 num_bytes -= bytes_to_add;
5484                         }
5485                         spin_unlock(&dest->lock);
5486                 }
5487                 if (num_bytes)
5488                         space_info_add_old_bytes(fs_info, space_info,
5489                                                  num_bytes);
5490         }
5491 }
5492
5493 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5494                             struct btrfs_block_rsv *dst, u64 num_bytes,
5495                             int update_size)
5496 {
5497         int ret;
5498
5499         ret = block_rsv_use_bytes(src, num_bytes);
5500         if (ret)
5501                 return ret;
5502
5503         block_rsv_add_bytes(dst, num_bytes, update_size);
5504         return 0;
5505 }
5506
5507 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5508 {
5509         memset(rsv, 0, sizeof(*rsv));
5510         spin_lock_init(&rsv->lock);
5511         rsv->type = type;
5512 }
5513
5514 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5515                                               unsigned short type)
5516 {
5517         struct btrfs_block_rsv *block_rsv;
5518
5519         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5520         if (!block_rsv)
5521                 return NULL;
5522
5523         btrfs_init_block_rsv(block_rsv, type);
5524         block_rsv->space_info = __find_space_info(fs_info,
5525                                                   BTRFS_BLOCK_GROUP_METADATA);
5526         return block_rsv;
5527 }
5528
5529 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5530                           struct btrfs_block_rsv *rsv)
5531 {
5532         if (!rsv)
5533                 return;
5534         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5535         kfree(rsv);
5536 }
5537
5538 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5539 {
5540         kfree(rsv);
5541 }
5542
5543 int btrfs_block_rsv_add(struct btrfs_root *root,
5544                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5545                         enum btrfs_reserve_flush_enum flush)
5546 {
5547         int ret;
5548
5549         if (num_bytes == 0)
5550                 return 0;
5551
5552         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5553         if (!ret) {
5554                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5555                 return 0;
5556         }
5557
5558         return ret;
5559 }
5560
5561 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5562 {
5563         u64 num_bytes = 0;
5564         int ret = -ENOSPC;
5565
5566         if (!block_rsv)
5567                 return 0;
5568
5569         spin_lock(&block_rsv->lock);
5570         num_bytes = div_factor(block_rsv->size, min_factor);
5571         if (block_rsv->reserved >= num_bytes)
5572                 ret = 0;
5573         spin_unlock(&block_rsv->lock);
5574
5575         return ret;
5576 }
5577
5578 int btrfs_block_rsv_refill(struct btrfs_root *root,
5579                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5580                            enum btrfs_reserve_flush_enum flush)
5581 {
5582         u64 num_bytes = 0;
5583         int ret = -ENOSPC;
5584
5585         if (!block_rsv)
5586                 return 0;
5587
5588         spin_lock(&block_rsv->lock);
5589         num_bytes = min_reserved;
5590         if (block_rsv->reserved >= num_bytes)
5591                 ret = 0;
5592         else
5593                 num_bytes -= block_rsv->reserved;
5594         spin_unlock(&block_rsv->lock);
5595
5596         if (!ret)
5597                 return 0;
5598
5599         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5600         if (!ret) {
5601                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5602                 return 0;
5603         }
5604
5605         return ret;
5606 }
5607
5608 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5609                              struct btrfs_block_rsv *block_rsv,
5610                              u64 num_bytes)
5611 {
5612         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5613
5614         if (global_rsv == block_rsv ||
5615             block_rsv->space_info != global_rsv->space_info)
5616                 global_rsv = NULL;
5617         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes);
5618 }
5619
5620 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5621 {
5622         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5623         struct btrfs_space_info *sinfo = block_rsv->space_info;
5624         u64 num_bytes;
5625
5626         /*
5627          * The global block rsv is based on the size of the extent tree, the
5628          * checksum tree and the root tree.  If the fs is empty we want to set
5629          * it to a minimal amount for safety.
5630          */
5631         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5632                 btrfs_root_used(&fs_info->csum_root->root_item) +
5633                 btrfs_root_used(&fs_info->tree_root->root_item);
5634         num_bytes = max_t(u64, num_bytes, SZ_16M);
5635
5636         spin_lock(&sinfo->lock);
5637         spin_lock(&block_rsv->lock);
5638
5639         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5640
5641         if (block_rsv->reserved < block_rsv->size) {
5642                 num_bytes = btrfs_space_info_used(sinfo, true);
5643                 if (sinfo->total_bytes > num_bytes) {
5644                         num_bytes = sinfo->total_bytes - num_bytes;
5645                         num_bytes = min(num_bytes,
5646                                         block_rsv->size - block_rsv->reserved);
5647                         block_rsv->reserved += num_bytes;
5648                         sinfo->bytes_may_use += num_bytes;
5649                         trace_btrfs_space_reservation(fs_info, "space_info",
5650                                                       sinfo->flags, num_bytes,
5651                                                       1);
5652                 }
5653         } else if (block_rsv->reserved > block_rsv->size) {
5654                 num_bytes = block_rsv->reserved - block_rsv->size;
5655                 sinfo->bytes_may_use -= num_bytes;
5656                 trace_btrfs_space_reservation(fs_info, "space_info",
5657                                       sinfo->flags, num_bytes, 0);
5658                 block_rsv->reserved = block_rsv->size;
5659         }
5660
5661         if (block_rsv->reserved == block_rsv->size)
5662                 block_rsv->full = 1;
5663         else
5664                 block_rsv->full = 0;
5665
5666         spin_unlock(&block_rsv->lock);
5667         spin_unlock(&sinfo->lock);
5668 }
5669
5670 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5671 {
5672         struct btrfs_space_info *space_info;
5673
5674         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5675         fs_info->chunk_block_rsv.space_info = space_info;
5676
5677         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5678         fs_info->global_block_rsv.space_info = space_info;
5679         fs_info->delalloc_block_rsv.space_info = space_info;
5680         fs_info->trans_block_rsv.space_info = space_info;
5681         fs_info->empty_block_rsv.space_info = space_info;
5682         fs_info->delayed_block_rsv.space_info = space_info;
5683
5684         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5685         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5686         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5687         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5688         if (fs_info->quota_root)
5689                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5690         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5691
5692         update_global_block_rsv(fs_info);
5693 }
5694
5695 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5696 {
5697         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5698                                 (u64)-1);
5699         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5700         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5701         WARN_ON(fs_info->trans_block_rsv.size > 0);
5702         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5703         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5704         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5705         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5706         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5707 }
5708
5709 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5710                                   struct btrfs_fs_info *fs_info)
5711 {
5712         if (!trans->block_rsv)
5713                 return;
5714
5715         if (!trans->bytes_reserved)
5716                 return;
5717
5718         trace_btrfs_space_reservation(fs_info, "transaction",
5719                                       trans->transid, trans->bytes_reserved, 0);
5720         btrfs_block_rsv_release(fs_info, trans->block_rsv,
5721                                 trans->bytes_reserved);
5722         trans->bytes_reserved = 0;
5723 }
5724
5725 /*
5726  * To be called after all the new block groups attached to the transaction
5727  * handle have been created (btrfs_create_pending_block_groups()).
5728  */
5729 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5730 {
5731         struct btrfs_fs_info *fs_info = trans->fs_info;
5732
5733         if (!trans->chunk_bytes_reserved)
5734                 return;
5735
5736         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5737
5738         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5739                                 trans->chunk_bytes_reserved);
5740         trans->chunk_bytes_reserved = 0;
5741 }
5742
5743 /* Can only return 0 or -ENOSPC */
5744 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5745                                   struct btrfs_inode *inode)
5746 {
5747         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5748         struct btrfs_root *root = inode->root;
5749         /*
5750          * We always use trans->block_rsv here as we will have reserved space
5751          * for our orphan when starting the transaction, using get_block_rsv()
5752          * here will sometimes make us choose the wrong block rsv as we could be
5753          * doing a reloc inode for a non refcounted root.
5754          */
5755         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5756         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5757
5758         /*
5759          * We need to hold space in order to delete our orphan item once we've
5760          * added it, so this takes the reservation so we can release it later
5761          * when we are truly done with the orphan item.
5762          */
5763         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5764
5765         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5766                         num_bytes, 1);
5767         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5768 }
5769
5770 void btrfs_orphan_release_metadata(struct inode *inode)
5771 {
5772         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5773         struct btrfs_root *root = BTRFS_I(inode)->root;
5774         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5775
5776         trace_btrfs_space_reservation(fs_info, "orphan",
5777                                       btrfs_ino(BTRFS_I(inode)), num_bytes, 0);
5778         btrfs_block_rsv_release(fs_info, root->orphan_block_rsv, num_bytes);
5779 }
5780
5781 /*
5782  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5783  * root: the root of the parent directory
5784  * rsv: block reservation
5785  * items: the number of items that we need do reservation
5786  * qgroup_reserved: used to return the reserved size in qgroup
5787  *
5788  * This function is used to reserve the space for snapshot/subvolume
5789  * creation and deletion. Those operations are different with the
5790  * common file/directory operations, they change two fs/file trees
5791  * and root tree, the number of items that the qgroup reserves is
5792  * different with the free space reservation. So we can not use
5793  * the space reservation mechanism in start_transaction().
5794  */
5795 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5796                                      struct btrfs_block_rsv *rsv,
5797                                      int items,
5798                                      u64 *qgroup_reserved,
5799                                      bool use_global_rsv)
5800 {
5801         u64 num_bytes;
5802         int ret;
5803         struct btrfs_fs_info *fs_info = root->fs_info;
5804         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5805
5806         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5807                 /* One for parent inode, two for dir entries */
5808                 num_bytes = 3 * fs_info->nodesize;
5809                 ret = btrfs_qgroup_reserve_meta(root, num_bytes, true);
5810                 if (ret)
5811                         return ret;
5812         } else {
5813                 num_bytes = 0;
5814         }
5815
5816         *qgroup_reserved = num_bytes;
5817
5818         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5819         rsv->space_info = __find_space_info(fs_info,
5820                                             BTRFS_BLOCK_GROUP_METADATA);
5821         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5822                                   BTRFS_RESERVE_FLUSH_ALL);
5823
5824         if (ret == -ENOSPC && use_global_rsv)
5825                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5826
5827         if (ret && *qgroup_reserved)
5828                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5829
5830         return ret;
5831 }
5832
5833 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
5834                                       struct btrfs_block_rsv *rsv)
5835 {
5836         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5837 }
5838
5839 /**
5840  * drop_outstanding_extent - drop an outstanding extent
5841  * @inode: the inode we're dropping the extent for
5842  * @num_bytes: the number of bytes we're releasing.
5843  *
5844  * This is called when we are freeing up an outstanding extent, either called
5845  * after an error or after an extent is written.  This will return the number of
5846  * reserved extents that need to be freed.  This must be called with
5847  * BTRFS_I(inode)->lock held.
5848  */
5849 static unsigned drop_outstanding_extent(struct btrfs_inode *inode,
5850                 u64 num_bytes)
5851 {
5852         unsigned drop_inode_space = 0;
5853         unsigned dropped_extents = 0;
5854         unsigned num_extents;
5855
5856         num_extents = count_max_extents(num_bytes);
5857         ASSERT(num_extents);
5858         ASSERT(inode->outstanding_extents >= num_extents);
5859         inode->outstanding_extents -= num_extents;
5860
5861         if (inode->outstanding_extents == 0 &&
5862             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5863                                &inode->runtime_flags))
5864                 drop_inode_space = 1;
5865
5866         /*
5867          * If we have more or the same amount of outstanding extents than we have
5868          * reserved then we need to leave the reserved extents count alone.
5869          */
5870         if (inode->outstanding_extents >= inode->reserved_extents)
5871                 return drop_inode_space;
5872
5873         dropped_extents = inode->reserved_extents - inode->outstanding_extents;
5874         inode->reserved_extents -= dropped_extents;
5875         return dropped_extents + drop_inode_space;
5876 }
5877
5878 /**
5879  * calc_csum_metadata_size - return the amount of metadata space that must be
5880  *      reserved/freed for the given bytes.
5881  * @inode: the inode we're manipulating
5882  * @num_bytes: the number of bytes in question
5883  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5884  *
5885  * This adjusts the number of csum_bytes in the inode and then returns the
5886  * correct amount of metadata that must either be reserved or freed.  We
5887  * calculate how many checksums we can fit into one leaf and then divide the
5888  * number of bytes that will need to be checksumed by this value to figure out
5889  * how many checksums will be required.  If we are adding bytes then the number
5890  * may go up and we will return the number of additional bytes that must be
5891  * reserved.  If it is going down we will return the number of bytes that must
5892  * be freed.
5893  *
5894  * This must be called with BTRFS_I(inode)->lock held.
5895  */
5896 static u64 calc_csum_metadata_size(struct btrfs_inode *inode, u64 num_bytes,
5897                                    int reserve)
5898 {
5899         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5900         u64 old_csums, num_csums;
5901
5902         if (inode->flags & BTRFS_INODE_NODATASUM && inode->csum_bytes == 0)
5903                 return 0;
5904
5905         old_csums = btrfs_csum_bytes_to_leaves(fs_info, inode->csum_bytes);
5906         if (reserve)
5907                 inode->csum_bytes += num_bytes;
5908         else
5909                 inode->csum_bytes -= num_bytes;
5910         num_csums = btrfs_csum_bytes_to_leaves(fs_info, inode->csum_bytes);
5911
5912         /* No change, no need to reserve more */
5913         if (old_csums == num_csums)
5914                 return 0;
5915
5916         if (reserve)
5917                 return btrfs_calc_trans_metadata_size(fs_info,
5918                                                       num_csums - old_csums);
5919
5920         return btrfs_calc_trans_metadata_size(fs_info, old_csums - num_csums);
5921 }
5922
5923 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5924 {
5925         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5926         struct btrfs_root *root = BTRFS_I(inode)->root;
5927         struct btrfs_block_rsv *block_rsv = &fs_info->delalloc_block_rsv;
5928         u64 to_reserve = 0;
5929         u64 csum_bytes;
5930         unsigned nr_extents;
5931         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5932         int ret = 0;
5933         bool delalloc_lock = true;
5934         u64 to_free = 0;
5935         unsigned dropped;
5936         bool release_extra = false;
5937
5938         /* If we are a free space inode we need to not flush since we will be in
5939          * the middle of a transaction commit.  We also don't need the delalloc
5940          * mutex since we won't race with anybody.  We need this mostly to make
5941          * lockdep shut its filthy mouth.
5942          *
5943          * If we have a transaction open (can happen if we call truncate_block
5944          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5945          */
5946         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
5947                 flush = BTRFS_RESERVE_NO_FLUSH;
5948                 delalloc_lock = false;
5949         } else if (current->journal_info) {
5950                 flush = BTRFS_RESERVE_FLUSH_LIMIT;
5951         }
5952
5953         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5954             btrfs_transaction_in_commit(fs_info))
5955                 schedule_timeout(1);
5956
5957         if (delalloc_lock)
5958                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5959
5960         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
5961
5962         spin_lock(&BTRFS_I(inode)->lock);
5963         nr_extents = count_max_extents(num_bytes);
5964         BTRFS_I(inode)->outstanding_extents += nr_extents;
5965
5966         nr_extents = 0;
5967         if (BTRFS_I(inode)->outstanding_extents >
5968             BTRFS_I(inode)->reserved_extents)
5969                 nr_extents += BTRFS_I(inode)->outstanding_extents -
5970                         BTRFS_I(inode)->reserved_extents;
5971
5972         /* We always want to reserve a slot for updating the inode. */
5973         to_reserve = btrfs_calc_trans_metadata_size(fs_info, nr_extents + 1);
5974         to_reserve += calc_csum_metadata_size(BTRFS_I(inode), num_bytes, 1);
5975         csum_bytes = BTRFS_I(inode)->csum_bytes;
5976         spin_unlock(&BTRFS_I(inode)->lock);
5977
5978         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5979                 ret = btrfs_qgroup_reserve_meta(root,
5980                                 nr_extents * fs_info->nodesize, true);
5981                 if (ret)
5982                         goto out_fail;
5983         }
5984
5985         ret = btrfs_block_rsv_add(root, block_rsv, to_reserve, flush);
5986         if (unlikely(ret)) {
5987                 btrfs_qgroup_free_meta(root,
5988                                        nr_extents * fs_info->nodesize);
5989                 goto out_fail;
5990         }
5991
5992         spin_lock(&BTRFS_I(inode)->lock);
5993         if (test_and_set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5994                              &BTRFS_I(inode)->runtime_flags)) {
5995                 to_reserve -= btrfs_calc_trans_metadata_size(fs_info, 1);
5996                 release_extra = true;
5997         }
5998         BTRFS_I(inode)->reserved_extents += nr_extents;
5999         spin_unlock(&BTRFS_I(inode)->lock);
6000
6001         if (delalloc_lock)
6002                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6003
6004         if (to_reserve)
6005                 trace_btrfs_space_reservation(fs_info, "delalloc",
6006                                       btrfs_ino(BTRFS_I(inode)), to_reserve, 1);
6007         if (release_extra)
6008                 btrfs_block_rsv_release(fs_info, block_rsv,
6009                                 btrfs_calc_trans_metadata_size(fs_info, 1));
6010         return 0;
6011
6012 out_fail:
6013         spin_lock(&BTRFS_I(inode)->lock);
6014         dropped = drop_outstanding_extent(BTRFS_I(inode), num_bytes);
6015         /*
6016          * If the inodes csum_bytes is the same as the original
6017          * csum_bytes then we know we haven't raced with any free()ers
6018          * so we can just reduce our inodes csum bytes and carry on.
6019          */
6020         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
6021                 calc_csum_metadata_size(BTRFS_I(inode), num_bytes, 0);
6022         } else {
6023                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
6024                 u64 bytes;
6025
6026                 /*
6027                  * This is tricky, but first we need to figure out how much we
6028                  * freed from any free-ers that occurred during this
6029                  * reservation, so we reset ->csum_bytes to the csum_bytes
6030                  * before we dropped our lock, and then call the free for the
6031                  * number of bytes that were freed while we were trying our
6032                  * reservation.
6033                  */
6034                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
6035                 BTRFS_I(inode)->csum_bytes = csum_bytes;
6036                 to_free = calc_csum_metadata_size(BTRFS_I(inode), bytes, 0);
6037
6038
6039                 /*
6040                  * Now we need to see how much we would have freed had we not
6041                  * been making this reservation and our ->csum_bytes were not
6042                  * artificially inflated.
6043                  */
6044                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
6045                 bytes = csum_bytes - orig_csum_bytes;
6046                 bytes = calc_csum_metadata_size(BTRFS_I(inode), bytes, 0);
6047
6048                 /*
6049                  * Now reset ->csum_bytes to what it should be.  If bytes is
6050                  * more than to_free then we would have freed more space had we
6051                  * not had an artificially high ->csum_bytes, so we need to free
6052                  * the remainder.  If bytes is the same or less then we don't
6053                  * need to do anything, the other free-ers did the correct
6054                  * thing.
6055                  */
6056                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
6057                 if (bytes > to_free)
6058                         to_free = bytes - to_free;
6059                 else
6060                         to_free = 0;
6061         }
6062         spin_unlock(&BTRFS_I(inode)->lock);
6063         if (dropped)
6064                 to_free += btrfs_calc_trans_metadata_size(fs_info, dropped);
6065
6066         if (to_free) {
6067                 btrfs_block_rsv_release(fs_info, block_rsv, to_free);
6068                 trace_btrfs_space_reservation(fs_info, "delalloc",
6069                                       btrfs_ino(BTRFS_I(inode)), to_free, 0);
6070         }
6071         if (delalloc_lock)
6072                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6073         return ret;
6074 }
6075
6076 /**
6077  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6078  * @inode: the inode to release the reservation for
6079  * @num_bytes: the number of bytes we're releasing
6080  *
6081  * This will release the metadata reservation for an inode.  This can be called
6082  * once we complete IO for a given set of bytes to release their metadata
6083  * reservations.
6084  */
6085 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
6086 {
6087         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6088         u64 to_free = 0;
6089         unsigned dropped;
6090
6091         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6092         spin_lock(&BTRFS_I(inode)->lock);
6093         dropped = drop_outstanding_extent(BTRFS_I(inode), num_bytes);
6094
6095         if (num_bytes)
6096                 to_free = calc_csum_metadata_size(BTRFS_I(inode), num_bytes, 0);
6097         spin_unlock(&BTRFS_I(inode)->lock);
6098         if (dropped > 0)
6099                 to_free += btrfs_calc_trans_metadata_size(fs_info, dropped);
6100
6101         if (btrfs_is_testing(fs_info))
6102                 return;
6103
6104         trace_btrfs_space_reservation(fs_info, "delalloc",
6105                                       btrfs_ino(BTRFS_I(inode)), to_free, 0);
6106
6107         btrfs_block_rsv_release(fs_info, &fs_info->delalloc_block_rsv, to_free);
6108 }
6109
6110 /**
6111  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6112  * delalloc
6113  * @inode: inode we're writing to
6114  * @start: start range we are writing to
6115  * @len: how long the range we are writing to
6116  *
6117  * This will do the following things
6118  *
6119  * o reserve space in data space info for num bytes
6120  *   and reserve precious corresponding qgroup space
6121  *   (Done in check_data_free_space)
6122  *
6123  * o reserve space for metadata space, based on the number of outstanding
6124  *   extents and how much csums will be needed
6125  *   also reserve metadata space in a per root over-reserve method.
6126  * o add to the inodes->delalloc_bytes
6127  * o add it to the fs_info's delalloc inodes list.
6128  *   (Above 3 all done in delalloc_reserve_metadata)
6129  *
6130  * Return 0 for success
6131  * Return <0 for error(-ENOSPC or -EQUOT)
6132  */
6133 int btrfs_delalloc_reserve_space(struct inode *inode, u64 start, u64 len)
6134 {
6135         int ret;
6136
6137         ret = btrfs_check_data_free_space(inode, start, len);
6138         if (ret < 0)
6139                 return ret;
6140         ret = btrfs_delalloc_reserve_metadata(inode, len);
6141         if (ret < 0)
6142                 btrfs_free_reserved_data_space(inode, start, len);
6143         return ret;
6144 }
6145
6146 /**
6147  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6148  * @inode: inode we're releasing space for
6149  * @start: start position of the space already reserved
6150  * @len: the len of the space already reserved
6151  *
6152  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
6153  * called in the case that we don't need the metadata AND data reservations
6154  * anymore.  So if there is an error or we insert an inline extent.
6155  *
6156  * This function will release the metadata space that was not used and will
6157  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6158  * list if there are no delalloc bytes left.
6159  * Also it will handle the qgroup reserved space.
6160  */
6161 void btrfs_delalloc_release_space(struct inode *inode, u64 start, u64 len)
6162 {
6163         btrfs_delalloc_release_metadata(inode, len);
6164         btrfs_free_reserved_data_space(inode, start, len);
6165 }
6166
6167 static int update_block_group(struct btrfs_trans_handle *trans,
6168                               struct btrfs_fs_info *info, u64 bytenr,
6169                               u64 num_bytes, int alloc)
6170 {
6171         struct btrfs_block_group_cache *cache = NULL;
6172         u64 total = num_bytes;
6173         u64 old_val;
6174         u64 byte_in_group;
6175         int factor;
6176
6177         /* block accounting for super block */
6178         spin_lock(&info->delalloc_root_lock);
6179         old_val = btrfs_super_bytes_used(info->super_copy);
6180         if (alloc)
6181                 old_val += num_bytes;
6182         else
6183                 old_val -= num_bytes;
6184         btrfs_set_super_bytes_used(info->super_copy, old_val);
6185         spin_unlock(&info->delalloc_root_lock);
6186
6187         while (total) {
6188                 cache = btrfs_lookup_block_group(info, bytenr);
6189                 if (!cache)
6190                         return -ENOENT;
6191                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6192                                     BTRFS_BLOCK_GROUP_RAID1 |
6193                                     BTRFS_BLOCK_GROUP_RAID10))
6194                         factor = 2;
6195                 else
6196                         factor = 1;
6197                 /*
6198                  * If this block group has free space cache written out, we
6199                  * need to make sure to load it if we are removing space.  This
6200                  * is because we need the unpinning stage to actually add the
6201                  * space back to the block group, otherwise we will leak space.
6202                  */
6203                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6204                         cache_block_group(cache, 1);
6205
6206                 byte_in_group = bytenr - cache->key.objectid;
6207                 WARN_ON(byte_in_group > cache->key.offset);
6208
6209                 spin_lock(&cache->space_info->lock);
6210                 spin_lock(&cache->lock);
6211
6212                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6213                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6214                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6215
6216                 old_val = btrfs_block_group_used(&cache->item);
6217                 num_bytes = min(total, cache->key.offset - byte_in_group);
6218                 if (alloc) {
6219                         old_val += num_bytes;
6220                         btrfs_set_block_group_used(&cache->item, old_val);
6221                         cache->reserved -= num_bytes;
6222                         cache->space_info->bytes_reserved -= num_bytes;
6223                         cache->space_info->bytes_used += num_bytes;
6224                         cache->space_info->disk_used += num_bytes * factor;
6225                         spin_unlock(&cache->lock);
6226                         spin_unlock(&cache->space_info->lock);
6227                 } else {
6228                         old_val -= num_bytes;
6229                         btrfs_set_block_group_used(&cache->item, old_val);
6230                         cache->pinned += num_bytes;
6231                         cache->space_info->bytes_pinned += num_bytes;
6232                         cache->space_info->bytes_used -= num_bytes;
6233                         cache->space_info->disk_used -= num_bytes * factor;
6234                         spin_unlock(&cache->lock);
6235                         spin_unlock(&cache->space_info->lock);
6236
6237                         trace_btrfs_space_reservation(info, "pinned",
6238                                                       cache->space_info->flags,
6239                                                       num_bytes, 1);
6240                         set_extent_dirty(info->pinned_extents,
6241                                          bytenr, bytenr + num_bytes - 1,
6242                                          GFP_NOFS | __GFP_NOFAIL);
6243                 }
6244
6245                 spin_lock(&trans->transaction->dirty_bgs_lock);
6246                 if (list_empty(&cache->dirty_list)) {
6247                         list_add_tail(&cache->dirty_list,
6248                                       &trans->transaction->dirty_bgs);
6249                                 trans->transaction->num_dirty_bgs++;
6250                         btrfs_get_block_group(cache);
6251                 }
6252                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6253
6254                 /*
6255                  * No longer have used bytes in this block group, queue it for
6256                  * deletion. We do this after adding the block group to the
6257                  * dirty list to avoid races between cleaner kthread and space
6258                  * cache writeout.
6259                  */
6260                 if (!alloc && old_val == 0) {
6261                         spin_lock(&info->unused_bgs_lock);
6262                         if (list_empty(&cache->bg_list)) {
6263                                 btrfs_get_block_group(cache);
6264                                 list_add_tail(&cache->bg_list,
6265                                               &info->unused_bgs);
6266                         }
6267                         spin_unlock(&info->unused_bgs_lock);
6268                 }
6269
6270                 btrfs_put_block_group(cache);
6271                 total -= num_bytes;
6272                 bytenr += num_bytes;
6273         }
6274         return 0;
6275 }
6276
6277 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6278 {
6279         struct btrfs_block_group_cache *cache;
6280         u64 bytenr;
6281
6282         spin_lock(&fs_info->block_group_cache_lock);
6283         bytenr = fs_info->first_logical_byte;
6284         spin_unlock(&fs_info->block_group_cache_lock);
6285
6286         if (bytenr < (u64)-1)
6287                 return bytenr;
6288
6289         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6290         if (!cache)
6291                 return 0;
6292
6293         bytenr = cache->key.objectid;
6294         btrfs_put_block_group(cache);
6295
6296         return bytenr;
6297 }
6298
6299 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6300                            struct btrfs_block_group_cache *cache,
6301                            u64 bytenr, u64 num_bytes, int reserved)
6302 {
6303         spin_lock(&cache->space_info->lock);
6304         spin_lock(&cache->lock);
6305         cache->pinned += num_bytes;
6306         cache->space_info->bytes_pinned += num_bytes;
6307         if (reserved) {
6308                 cache->reserved -= num_bytes;
6309                 cache->space_info->bytes_reserved -= num_bytes;
6310         }
6311         spin_unlock(&cache->lock);
6312         spin_unlock(&cache->space_info->lock);
6313
6314         trace_btrfs_space_reservation(fs_info, "pinned",
6315                                       cache->space_info->flags, num_bytes, 1);
6316         set_extent_dirty(fs_info->pinned_extents, bytenr,
6317                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6318         return 0;
6319 }
6320
6321 /*
6322  * this function must be called within transaction
6323  */
6324 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6325                      u64 bytenr, u64 num_bytes, int reserved)
6326 {
6327         struct btrfs_block_group_cache *cache;
6328
6329         cache = btrfs_lookup_block_group(fs_info, bytenr);
6330         BUG_ON(!cache); /* Logic error */
6331
6332         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6333
6334         btrfs_put_block_group(cache);
6335         return 0;
6336 }
6337
6338 /*
6339  * this function must be called within transaction
6340  */
6341 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6342                                     u64 bytenr, u64 num_bytes)
6343 {
6344         struct btrfs_block_group_cache *cache;
6345         int ret;
6346
6347         cache = btrfs_lookup_block_group(fs_info, bytenr);
6348         if (!cache)
6349                 return -EINVAL;
6350
6351         /*
6352          * pull in the free space cache (if any) so that our pin
6353          * removes the free space from the cache.  We have load_only set
6354          * to one because the slow code to read in the free extents does check
6355          * the pinned extents.
6356          */
6357         cache_block_group(cache, 1);
6358
6359         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6360
6361         /* remove us from the free space cache (if we're there at all) */
6362         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6363         btrfs_put_block_group(cache);
6364         return ret;
6365 }
6366
6367 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6368                                    u64 start, u64 num_bytes)
6369 {
6370         int ret;
6371         struct btrfs_block_group_cache *block_group;
6372         struct btrfs_caching_control *caching_ctl;
6373
6374         block_group = btrfs_lookup_block_group(fs_info, start);
6375         if (!block_group)
6376                 return -EINVAL;
6377
6378         cache_block_group(block_group, 0);
6379         caching_ctl = get_caching_control(block_group);
6380
6381         if (!caching_ctl) {
6382                 /* Logic error */
6383                 BUG_ON(!block_group_cache_done(block_group));
6384                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6385         } else {
6386                 mutex_lock(&caching_ctl->mutex);
6387
6388                 if (start >= caching_ctl->progress) {
6389                         ret = add_excluded_extent(fs_info, start, num_bytes);
6390                 } else if (start + num_bytes <= caching_ctl->progress) {
6391                         ret = btrfs_remove_free_space(block_group,
6392                                                       start, num_bytes);
6393                 } else {
6394                         num_bytes = caching_ctl->progress - start;
6395                         ret = btrfs_remove_free_space(block_group,
6396                                                       start, num_bytes);
6397                         if (ret)
6398                                 goto out_lock;
6399
6400                         num_bytes = (start + num_bytes) -
6401                                 caching_ctl->progress;
6402                         start = caching_ctl->progress;
6403                         ret = add_excluded_extent(fs_info, start, num_bytes);
6404                 }
6405 out_lock:
6406                 mutex_unlock(&caching_ctl->mutex);
6407                 put_caching_control(caching_ctl);
6408         }
6409         btrfs_put_block_group(block_group);
6410         return ret;
6411 }
6412
6413 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6414                                  struct extent_buffer *eb)
6415 {
6416         struct btrfs_file_extent_item *item;
6417         struct btrfs_key key;
6418         int found_type;
6419         int i;
6420
6421         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6422                 return 0;
6423
6424         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6425                 btrfs_item_key_to_cpu(eb, &key, i);
6426                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6427                         continue;
6428                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6429                 found_type = btrfs_file_extent_type(eb, item);
6430                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6431                         continue;
6432                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6433                         continue;
6434                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6435                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6436                 __exclude_logged_extent(fs_info, key.objectid, key.offset);
6437         }
6438
6439         return 0;
6440 }
6441
6442 static void
6443 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6444 {
6445         atomic_inc(&bg->reservations);
6446 }
6447
6448 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6449                                         const u64 start)
6450 {
6451         struct btrfs_block_group_cache *bg;
6452
6453         bg = btrfs_lookup_block_group(fs_info, start);
6454         ASSERT(bg);
6455         if (atomic_dec_and_test(&bg->reservations))
6456                 wake_up_atomic_t(&bg->reservations);
6457         btrfs_put_block_group(bg);
6458 }
6459
6460 static int btrfs_wait_bg_reservations_atomic_t(atomic_t *a)
6461 {
6462         schedule();
6463         return 0;
6464 }
6465
6466 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6467 {
6468         struct btrfs_space_info *space_info = bg->space_info;
6469
6470         ASSERT(bg->ro);
6471
6472         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6473                 return;
6474
6475         /*
6476          * Our block group is read only but before we set it to read only,
6477          * some task might have had allocated an extent from it already, but it
6478          * has not yet created a respective ordered extent (and added it to a
6479          * root's list of ordered extents).
6480          * Therefore wait for any task currently allocating extents, since the
6481          * block group's reservations counter is incremented while a read lock
6482          * on the groups' semaphore is held and decremented after releasing
6483          * the read access on that semaphore and creating the ordered extent.
6484          */
6485         down_write(&space_info->groups_sem);
6486         up_write(&space_info->groups_sem);
6487
6488         wait_on_atomic_t(&bg->reservations,
6489                          btrfs_wait_bg_reservations_atomic_t,
6490                          TASK_UNINTERRUPTIBLE);
6491 }
6492
6493 /**
6494  * btrfs_add_reserved_bytes - update the block_group and space info counters
6495  * @cache:      The cache we are manipulating
6496  * @ram_bytes:  The number of bytes of file content, and will be same to
6497  *              @num_bytes except for the compress path.
6498  * @num_bytes:  The number of bytes in question
6499  * @delalloc:   The blocks are allocated for the delalloc write
6500  *
6501  * This is called by the allocator when it reserves space. If this is a
6502  * reservation and the block group has become read only we cannot make the
6503  * reservation and return -EAGAIN, otherwise this function always succeeds.
6504  */
6505 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6506                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6507 {
6508         struct btrfs_space_info *space_info = cache->space_info;
6509         int ret = 0;
6510
6511         spin_lock(&space_info->lock);
6512         spin_lock(&cache->lock);
6513         if (cache->ro) {
6514                 ret = -EAGAIN;
6515         } else {
6516                 cache->reserved += num_bytes;
6517                 space_info->bytes_reserved += num_bytes;
6518
6519                 trace_btrfs_space_reservation(cache->fs_info,
6520                                 "space_info", space_info->flags,
6521                                 ram_bytes, 0);
6522                 space_info->bytes_may_use -= ram_bytes;
6523                 if (delalloc)
6524                         cache->delalloc_bytes += num_bytes;
6525         }
6526         spin_unlock(&cache->lock);
6527         spin_unlock(&space_info->lock);
6528         return ret;
6529 }
6530
6531 /**
6532  * btrfs_free_reserved_bytes - update the block_group and space info counters
6533  * @cache:      The cache we are manipulating
6534  * @num_bytes:  The number of bytes in question
6535  * @delalloc:   The blocks are allocated for the delalloc write
6536  *
6537  * This is called by somebody who is freeing space that was never actually used
6538  * on disk.  For example if you reserve some space for a new leaf in transaction
6539  * A and before transaction A commits you free that leaf, you call this with
6540  * reserve set to 0 in order to clear the reservation.
6541  */
6542
6543 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6544                                      u64 num_bytes, int delalloc)
6545 {
6546         struct btrfs_space_info *space_info = cache->space_info;
6547         int ret = 0;
6548
6549         spin_lock(&space_info->lock);
6550         spin_lock(&cache->lock);
6551         if (cache->ro)
6552                 space_info->bytes_readonly += num_bytes;
6553         cache->reserved -= num_bytes;
6554         space_info->bytes_reserved -= num_bytes;
6555
6556         if (delalloc)
6557                 cache->delalloc_bytes -= num_bytes;
6558         spin_unlock(&cache->lock);
6559         spin_unlock(&space_info->lock);
6560         return ret;
6561 }
6562 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6563 {
6564         struct btrfs_caching_control *next;
6565         struct btrfs_caching_control *caching_ctl;
6566         struct btrfs_block_group_cache *cache;
6567
6568         down_write(&fs_info->commit_root_sem);
6569
6570         list_for_each_entry_safe(caching_ctl, next,
6571                                  &fs_info->caching_block_groups, list) {
6572                 cache = caching_ctl->block_group;
6573                 if (block_group_cache_done(cache)) {
6574                         cache->last_byte_to_unpin = (u64)-1;
6575                         list_del_init(&caching_ctl->list);
6576                         put_caching_control(caching_ctl);
6577                 } else {
6578                         cache->last_byte_to_unpin = caching_ctl->progress;
6579                 }
6580         }
6581
6582         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6583                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6584         else
6585                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6586
6587         up_write(&fs_info->commit_root_sem);
6588
6589         update_global_block_rsv(fs_info);
6590 }
6591
6592 /*
6593  * Returns the free cluster for the given space info and sets empty_cluster to
6594  * what it should be based on the mount options.
6595  */
6596 static struct btrfs_free_cluster *
6597 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6598                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6599 {
6600         struct btrfs_free_cluster *ret = NULL;
6601         bool ssd = btrfs_test_opt(fs_info, SSD);
6602
6603         *empty_cluster = 0;
6604         if (btrfs_mixed_space_info(space_info))
6605                 return ret;
6606
6607         if (ssd)
6608                 *empty_cluster = SZ_2M;
6609         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6610                 ret = &fs_info->meta_alloc_cluster;
6611                 if (!ssd)
6612                         *empty_cluster = SZ_64K;
6613         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) && ssd) {
6614                 ret = &fs_info->data_alloc_cluster;
6615         }
6616
6617         return ret;
6618 }
6619
6620 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6621                               u64 start, u64 end,
6622                               const bool return_free_space)
6623 {
6624         struct btrfs_block_group_cache *cache = NULL;
6625         struct btrfs_space_info *space_info;
6626         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6627         struct btrfs_free_cluster *cluster = NULL;
6628         u64 len;
6629         u64 total_unpinned = 0;
6630         u64 empty_cluster = 0;
6631         bool readonly;
6632
6633         while (start <= end) {
6634                 readonly = false;
6635                 if (!cache ||
6636                     start >= cache->key.objectid + cache->key.offset) {
6637                         if (cache)
6638                                 btrfs_put_block_group(cache);
6639                         total_unpinned = 0;
6640                         cache = btrfs_lookup_block_group(fs_info, start);
6641                         BUG_ON(!cache); /* Logic error */
6642
6643                         cluster = fetch_cluster_info(fs_info,
6644                                                      cache->space_info,
6645                                                      &empty_cluster);
6646                         empty_cluster <<= 1;
6647                 }
6648
6649                 len = cache->key.objectid + cache->key.offset - start;
6650                 len = min(len, end + 1 - start);
6651
6652                 if (start < cache->last_byte_to_unpin) {
6653                         len = min(len, cache->last_byte_to_unpin - start);
6654                         if (return_free_space)
6655                                 btrfs_add_free_space(cache, start, len);
6656                 }
6657
6658                 start += len;
6659                 total_unpinned += len;
6660                 space_info = cache->space_info;
6661
6662                 /*
6663                  * If this space cluster has been marked as fragmented and we've
6664                  * unpinned enough in this block group to potentially allow a
6665                  * cluster to be created inside of it go ahead and clear the
6666                  * fragmented check.
6667                  */
6668                 if (cluster && cluster->fragmented &&
6669                     total_unpinned > empty_cluster) {
6670                         spin_lock(&cluster->lock);
6671                         cluster->fragmented = 0;
6672                         spin_unlock(&cluster->lock);
6673                 }
6674
6675                 spin_lock(&space_info->lock);
6676                 spin_lock(&cache->lock);
6677                 cache->pinned -= len;
6678                 space_info->bytes_pinned -= len;
6679
6680                 trace_btrfs_space_reservation(fs_info, "pinned",
6681                                               space_info->flags, len, 0);
6682                 space_info->max_extent_size = 0;
6683                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6684                 if (cache->ro) {
6685                         space_info->bytes_readonly += len;
6686                         readonly = true;
6687                 }
6688                 spin_unlock(&cache->lock);
6689                 if (!readonly && return_free_space &&
6690                     global_rsv->space_info == space_info) {
6691                         u64 to_add = len;
6692                         WARN_ON(!return_free_space);
6693                         spin_lock(&global_rsv->lock);
6694                         if (!global_rsv->full) {
6695                                 to_add = min(len, global_rsv->size -
6696                                              global_rsv->reserved);
6697                                 global_rsv->reserved += to_add;
6698                                 space_info->bytes_may_use += to_add;
6699                                 if (global_rsv->reserved >= global_rsv->size)
6700                                         global_rsv->full = 1;
6701                                 trace_btrfs_space_reservation(fs_info,
6702                                                               "space_info",
6703                                                               space_info->flags,
6704                                                               to_add, 1);
6705                                 len -= to_add;
6706                         }
6707                         spin_unlock(&global_rsv->lock);
6708                         /* Add to any tickets we may have */
6709                         if (len)
6710                                 space_info_add_new_bytes(fs_info, space_info,
6711                                                          len);
6712                 }
6713                 spin_unlock(&space_info->lock);
6714         }
6715
6716         if (cache)
6717                 btrfs_put_block_group(cache);
6718         return 0;
6719 }
6720
6721 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6722                                struct btrfs_fs_info *fs_info)
6723 {
6724         struct btrfs_block_group_cache *block_group, *tmp;
6725         struct list_head *deleted_bgs;
6726         struct extent_io_tree *unpin;
6727         u64 start;
6728         u64 end;
6729         int ret;
6730
6731         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6732                 unpin = &fs_info->freed_extents[1];
6733         else
6734                 unpin = &fs_info->freed_extents[0];
6735
6736         while (!trans->aborted) {
6737                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6738                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6739                                             EXTENT_DIRTY, NULL);
6740                 if (ret) {
6741                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6742                         break;
6743                 }
6744
6745                 if (btrfs_test_opt(fs_info, DISCARD))
6746                         ret = btrfs_discard_extent(fs_info, start,
6747                                                    end + 1 - start, NULL);
6748
6749                 clear_extent_dirty(unpin, start, end);
6750                 unpin_extent_range(fs_info, start, end, true);
6751                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6752                 cond_resched();
6753         }
6754
6755         /*
6756          * Transaction is finished.  We don't need the lock anymore.  We
6757          * do need to clean up the block groups in case of a transaction
6758          * abort.
6759          */
6760         deleted_bgs = &trans->transaction->deleted_bgs;
6761         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6762                 u64 trimmed = 0;
6763
6764                 ret = -EROFS;
6765                 if (!trans->aborted)
6766                         ret = btrfs_discard_extent(fs_info,
6767                                                    block_group->key.objectid,
6768                                                    block_group->key.offset,
6769                                                    &trimmed);
6770
6771                 list_del_init(&block_group->bg_list);
6772                 btrfs_put_block_group_trimming(block_group);
6773                 btrfs_put_block_group(block_group);
6774
6775                 if (ret) {
6776                         const char *errstr = btrfs_decode_error(ret);
6777                         btrfs_warn(fs_info,
6778                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6779                                    ret, errstr);
6780                 }
6781         }
6782
6783         return 0;
6784 }
6785
6786 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6787                              u64 owner, u64 root_objectid)
6788 {
6789         struct btrfs_space_info *space_info;
6790         u64 flags;
6791
6792         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6793                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6794                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6795                 else
6796                         flags = BTRFS_BLOCK_GROUP_METADATA;
6797         } else {
6798                 flags = BTRFS_BLOCK_GROUP_DATA;
6799         }
6800
6801         space_info = __find_space_info(fs_info, flags);
6802         BUG_ON(!space_info); /* Logic bug */
6803         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6804 }
6805
6806
6807 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6808                                 struct btrfs_fs_info *info,
6809                                 struct btrfs_delayed_ref_node *node, u64 parent,
6810                                 u64 root_objectid, u64 owner_objectid,
6811                                 u64 owner_offset, int refs_to_drop,
6812                                 struct btrfs_delayed_extent_op *extent_op)
6813 {
6814         struct btrfs_key key;
6815         struct btrfs_path *path;
6816         struct btrfs_root *extent_root = info->extent_root;
6817         struct extent_buffer *leaf;
6818         struct btrfs_extent_item *ei;
6819         struct btrfs_extent_inline_ref *iref;
6820         int ret;
6821         int is_data;
6822         int extent_slot = 0;
6823         int found_extent = 0;
6824         int num_to_del = 1;
6825         u32 item_size;
6826         u64 refs;
6827         u64 bytenr = node->bytenr;
6828         u64 num_bytes = node->num_bytes;
6829         int last_ref = 0;
6830         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6831
6832         path = btrfs_alloc_path();
6833         if (!path)
6834                 return -ENOMEM;
6835
6836         path->reada = READA_FORWARD;
6837         path->leave_spinning = 1;
6838
6839         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6840         BUG_ON(!is_data && refs_to_drop != 1);
6841
6842         if (is_data)
6843                 skinny_metadata = 0;
6844
6845         ret = lookup_extent_backref(trans, info, path, &iref,
6846                                     bytenr, num_bytes, parent,
6847                                     root_objectid, owner_objectid,
6848                                     owner_offset);
6849         if (ret == 0) {
6850                 extent_slot = path->slots[0];
6851                 while (extent_slot >= 0) {
6852                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6853                                               extent_slot);
6854                         if (key.objectid != bytenr)
6855                                 break;
6856                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6857                             key.offset == num_bytes) {
6858                                 found_extent = 1;
6859                                 break;
6860                         }
6861                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6862                             key.offset == owner_objectid) {
6863                                 found_extent = 1;
6864                                 break;
6865                         }
6866                         if (path->slots[0] - extent_slot > 5)
6867                                 break;
6868                         extent_slot--;
6869                 }
6870 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6871                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6872                 if (found_extent && item_size < sizeof(*ei))
6873                         found_extent = 0;
6874 #endif
6875                 if (!found_extent) {
6876                         BUG_ON(iref);
6877                         ret = remove_extent_backref(trans, info, path, NULL,
6878                                                     refs_to_drop,
6879                                                     is_data, &last_ref);
6880                         if (ret) {
6881                                 btrfs_abort_transaction(trans, ret);
6882                                 goto out;
6883                         }
6884                         btrfs_release_path(path);
6885                         path->leave_spinning = 1;
6886
6887                         key.objectid = bytenr;
6888                         key.type = BTRFS_EXTENT_ITEM_KEY;
6889                         key.offset = num_bytes;
6890
6891                         if (!is_data && skinny_metadata) {
6892                                 key.type = BTRFS_METADATA_ITEM_KEY;
6893                                 key.offset = owner_objectid;
6894                         }
6895
6896                         ret = btrfs_search_slot(trans, extent_root,
6897                                                 &key, path, -1, 1);
6898                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6899                                 /*
6900                                  * Couldn't find our skinny metadata item,
6901                                  * see if we have ye olde extent item.
6902                                  */
6903                                 path->slots[0]--;
6904                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6905                                                       path->slots[0]);
6906                                 if (key.objectid == bytenr &&
6907                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6908                                     key.offset == num_bytes)
6909                                         ret = 0;
6910                         }
6911
6912                         if (ret > 0 && skinny_metadata) {
6913                                 skinny_metadata = false;
6914                                 key.objectid = bytenr;
6915                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6916                                 key.offset = num_bytes;
6917                                 btrfs_release_path(path);
6918                                 ret = btrfs_search_slot(trans, extent_root,
6919                                                         &key, path, -1, 1);
6920                         }
6921
6922                         if (ret) {
6923                                 btrfs_err(info,
6924                                           "umm, got %d back from search, was looking for %llu",
6925                                           ret, bytenr);
6926                                 if (ret > 0)
6927                                         btrfs_print_leaf(info, path->nodes[0]);
6928                         }
6929                         if (ret < 0) {
6930                                 btrfs_abort_transaction(trans, ret);
6931                                 goto out;
6932                         }
6933                         extent_slot = path->slots[0];
6934                 }
6935         } else if (WARN_ON(ret == -ENOENT)) {
6936                 btrfs_print_leaf(info, path->nodes[0]);
6937                 btrfs_err(info,
6938                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6939                         bytenr, parent, root_objectid, owner_objectid,
6940                         owner_offset);
6941                 btrfs_abort_transaction(trans, ret);
6942                 goto out;
6943         } else {
6944                 btrfs_abort_transaction(trans, ret);
6945                 goto out;
6946         }
6947
6948         leaf = path->nodes[0];
6949         item_size = btrfs_item_size_nr(leaf, extent_slot);
6950 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6951         if (item_size < sizeof(*ei)) {
6952                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6953                 ret = convert_extent_item_v0(trans, info, path, owner_objectid,
6954                                              0);
6955                 if (ret < 0) {
6956                         btrfs_abort_transaction(trans, ret);
6957                         goto out;
6958                 }
6959
6960                 btrfs_release_path(path);
6961                 path->leave_spinning = 1;
6962
6963                 key.objectid = bytenr;
6964                 key.type = BTRFS_EXTENT_ITEM_KEY;
6965                 key.offset = num_bytes;
6966
6967                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6968                                         -1, 1);
6969                 if (ret) {
6970                         btrfs_err(info,
6971                                   "umm, got %d back from search, was looking for %llu",
6972                                 ret, bytenr);
6973                         btrfs_print_leaf(info, path->nodes[0]);
6974                 }
6975                 if (ret < 0) {
6976                         btrfs_abort_transaction(trans, ret);
6977                         goto out;
6978                 }
6979
6980                 extent_slot = path->slots[0];
6981                 leaf = path->nodes[0];
6982                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6983         }
6984 #endif
6985         BUG_ON(item_size < sizeof(*ei));
6986         ei = btrfs_item_ptr(leaf, extent_slot,
6987                             struct btrfs_extent_item);
6988         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6989             key.type == BTRFS_EXTENT_ITEM_KEY) {
6990                 struct btrfs_tree_block_info *bi;
6991                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6992                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6993                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6994         }
6995
6996         refs = btrfs_extent_refs(leaf, ei);
6997         if (refs < refs_to_drop) {
6998                 btrfs_err(info,
6999                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7000                           refs_to_drop, refs, bytenr);
7001                 ret = -EINVAL;
7002                 btrfs_abort_transaction(trans, ret);
7003                 goto out;
7004         }
7005         refs -= refs_to_drop;
7006
7007         if (refs > 0) {
7008                 if (extent_op)
7009                         __run_delayed_extent_op(extent_op, leaf, ei);
7010                 /*
7011                  * In the case of inline back ref, reference count will
7012                  * be updated by remove_extent_backref
7013                  */
7014                 if (iref) {
7015                         BUG_ON(!found_extent);
7016                 } else {
7017                         btrfs_set_extent_refs(leaf, ei, refs);
7018                         btrfs_mark_buffer_dirty(leaf);
7019                 }
7020                 if (found_extent) {
7021                         ret = remove_extent_backref(trans, info, path,
7022                                                     iref, refs_to_drop,
7023                                                     is_data, &last_ref);
7024                         if (ret) {
7025                                 btrfs_abort_transaction(trans, ret);
7026                                 goto out;
7027                         }
7028                 }
7029                 add_pinned_bytes(info, -num_bytes, owner_objectid,
7030                                  root_objectid);
7031         } else {
7032                 if (found_extent) {
7033                         BUG_ON(is_data && refs_to_drop !=
7034                                extent_data_ref_count(path, iref));
7035                         if (iref) {
7036                                 BUG_ON(path->slots[0] != extent_slot);
7037                         } else {
7038                                 BUG_ON(path->slots[0] != extent_slot + 1);
7039                                 path->slots[0] = extent_slot;
7040                                 num_to_del = 2;
7041                         }
7042                 }
7043
7044                 last_ref = 1;
7045                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7046                                       num_to_del);
7047                 if (ret) {
7048                         btrfs_abort_transaction(trans, ret);
7049                         goto out;
7050                 }
7051                 btrfs_release_path(path);
7052
7053                 if (is_data) {
7054                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7055                         if (ret) {
7056                                 btrfs_abort_transaction(trans, ret);
7057                                 goto out;
7058                         }
7059                 }
7060
7061                 ret = add_to_free_space_tree(trans, info, bytenr, num_bytes);
7062                 if (ret) {
7063                         btrfs_abort_transaction(trans, ret);
7064                         goto out;
7065                 }
7066
7067                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7068                 if (ret) {
7069                         btrfs_abort_transaction(trans, ret);
7070                         goto out;
7071                 }
7072         }
7073         btrfs_release_path(path);
7074
7075 out:
7076         btrfs_free_path(path);
7077         return ret;
7078 }
7079
7080 /*
7081  * when we free an block, it is possible (and likely) that we free the last
7082  * delayed ref for that extent as well.  This searches the delayed ref tree for
7083  * a given extent, and if there are no other delayed refs to be processed, it
7084  * removes it from the tree.
7085  */
7086 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7087                                       u64 bytenr)
7088 {
7089         struct btrfs_delayed_ref_head *head;
7090         struct btrfs_delayed_ref_root *delayed_refs;
7091         int ret = 0;
7092
7093         delayed_refs = &trans->transaction->delayed_refs;
7094         spin_lock(&delayed_refs->lock);
7095         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7096         if (!head)
7097                 goto out_delayed_unlock;
7098
7099         spin_lock(&head->lock);
7100         if (!list_empty(&head->ref_list))
7101                 goto out;
7102
7103         if (head->extent_op) {
7104                 if (!head->must_insert_reserved)
7105                         goto out;
7106                 btrfs_free_delayed_extent_op(head->extent_op);
7107                 head->extent_op = NULL;
7108         }
7109
7110         /*
7111          * waiting for the lock here would deadlock.  If someone else has it
7112          * locked they are already in the process of dropping it anyway
7113          */
7114         if (!mutex_trylock(&head->mutex))
7115                 goto out;
7116
7117         /*
7118          * at this point we have a head with no other entries.  Go
7119          * ahead and process it.
7120          */
7121         head->node.in_tree = 0;
7122         rb_erase(&head->href_node, &delayed_refs->href_root);
7123
7124         atomic_dec(&delayed_refs->num_entries);
7125
7126         /*
7127          * we don't take a ref on the node because we're removing it from the
7128          * tree, so we just steal the ref the tree was holding.
7129          */
7130         delayed_refs->num_heads--;
7131         if (head->processing == 0)
7132                 delayed_refs->num_heads_ready--;
7133         head->processing = 0;
7134         spin_unlock(&head->lock);
7135         spin_unlock(&delayed_refs->lock);
7136
7137         BUG_ON(head->extent_op);
7138         if (head->must_insert_reserved)
7139                 ret = 1;
7140
7141         mutex_unlock(&head->mutex);
7142         btrfs_put_delayed_ref(&head->node);
7143         return ret;
7144 out:
7145         spin_unlock(&head->lock);
7146
7147 out_delayed_unlock:
7148         spin_unlock(&delayed_refs->lock);
7149         return 0;
7150 }
7151
7152 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7153                            struct btrfs_root *root,
7154                            struct extent_buffer *buf,
7155                            u64 parent, int last_ref)
7156 {
7157         struct btrfs_fs_info *fs_info = root->fs_info;
7158         int pin = 1;
7159         int ret;
7160
7161         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7162                 ret = btrfs_add_delayed_tree_ref(fs_info, trans,
7163                                                  buf->start, buf->len,
7164                                                  parent,
7165                                                  root->root_key.objectid,
7166                                                  btrfs_header_level(buf),
7167                                                  BTRFS_DROP_DELAYED_REF, NULL);
7168                 BUG_ON(ret); /* -ENOMEM */
7169         }
7170
7171         if (!last_ref)
7172                 return;
7173
7174         if (btrfs_header_generation(buf) == trans->transid) {
7175                 struct btrfs_block_group_cache *cache;
7176
7177                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7178                         ret = check_ref_cleanup(trans, buf->start);
7179                         if (!ret)
7180                                 goto out;
7181                 }
7182
7183                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7184
7185                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7186                         pin_down_extent(fs_info, cache, buf->start,
7187                                         buf->len, 1);
7188                         btrfs_put_block_group(cache);
7189                         goto out;
7190                 }
7191
7192                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7193
7194                 btrfs_add_free_space(cache, buf->start, buf->len);
7195                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7196                 btrfs_put_block_group(cache);
7197                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7198                 pin = 0;
7199         }
7200 out:
7201         if (pin)
7202                 add_pinned_bytes(fs_info, buf->len, btrfs_header_level(buf),
7203                                  root->root_key.objectid);
7204
7205         /*
7206          * Deleting the buffer, clear the corrupt flag since it doesn't matter
7207          * anymore.
7208          */
7209         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7210 }
7211
7212 /* Can return -ENOMEM */
7213 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7214                       struct btrfs_fs_info *fs_info,
7215                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7216                       u64 owner, u64 offset)
7217 {
7218         int ret;
7219
7220         if (btrfs_is_testing(fs_info))
7221                 return 0;
7222
7223         add_pinned_bytes(fs_info, num_bytes, owner, root_objectid);
7224
7225         /*
7226          * tree log blocks never actually go into the extent allocation
7227          * tree, just update pinning info and exit early.
7228          */
7229         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7230                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7231                 /* unlocks the pinned mutex */
7232                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7233                 ret = 0;
7234         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7235                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7236                                         num_bytes,
7237                                         parent, root_objectid, (int)owner,
7238                                         BTRFS_DROP_DELAYED_REF, NULL);
7239         } else {
7240                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7241                                                 num_bytes,
7242                                                 parent, root_objectid, owner,
7243                                                 offset, 0,
7244                                                 BTRFS_DROP_DELAYED_REF);
7245         }
7246         return ret;
7247 }
7248
7249 /*
7250  * when we wait for progress in the block group caching, its because
7251  * our allocation attempt failed at least once.  So, we must sleep
7252  * and let some progress happen before we try again.
7253  *
7254  * This function will sleep at least once waiting for new free space to
7255  * show up, and then it will check the block group free space numbers
7256  * for our min num_bytes.  Another option is to have it go ahead
7257  * and look in the rbtree for a free extent of a given size, but this
7258  * is a good start.
7259  *
7260  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7261  * any of the information in this block group.
7262  */
7263 static noinline void
7264 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7265                                 u64 num_bytes)
7266 {
7267         struct btrfs_caching_control *caching_ctl;
7268
7269         caching_ctl = get_caching_control(cache);
7270         if (!caching_ctl)
7271                 return;
7272
7273         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7274                    (cache->free_space_ctl->free_space >= num_bytes));
7275
7276         put_caching_control(caching_ctl);
7277 }
7278
7279 static noinline int
7280 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7281 {
7282         struct btrfs_caching_control *caching_ctl;
7283         int ret = 0;
7284
7285         caching_ctl = get_caching_control(cache);
7286         if (!caching_ctl)
7287                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7288
7289         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7290         if (cache->cached == BTRFS_CACHE_ERROR)
7291                 ret = -EIO;
7292         put_caching_control(caching_ctl);
7293         return ret;
7294 }
7295
7296 int __get_raid_index(u64 flags)
7297 {
7298         if (flags & BTRFS_BLOCK_GROUP_RAID10)
7299                 return BTRFS_RAID_RAID10;
7300         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
7301                 return BTRFS_RAID_RAID1;
7302         else if (flags & BTRFS_BLOCK_GROUP_DUP)
7303                 return BTRFS_RAID_DUP;
7304         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
7305                 return BTRFS_RAID_RAID0;
7306         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
7307                 return BTRFS_RAID_RAID5;
7308         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
7309                 return BTRFS_RAID_RAID6;
7310
7311         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
7312 }
7313
7314 int get_block_group_index(struct btrfs_block_group_cache *cache)
7315 {
7316         return __get_raid_index(cache->flags);
7317 }
7318
7319 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7320         [BTRFS_RAID_RAID10]     = "raid10",
7321         [BTRFS_RAID_RAID1]      = "raid1",
7322         [BTRFS_RAID_DUP]        = "dup",
7323         [BTRFS_RAID_RAID0]      = "raid0",
7324         [BTRFS_RAID_SINGLE]     = "single",
7325         [BTRFS_RAID_RAID5]      = "raid5",
7326         [BTRFS_RAID_RAID6]      = "raid6",
7327 };
7328
7329 static const char *get_raid_name(enum btrfs_raid_types type)
7330 {
7331         if (type >= BTRFS_NR_RAID_TYPES)
7332                 return NULL;
7333
7334         return btrfs_raid_type_names[type];
7335 }
7336
7337 enum btrfs_loop_type {
7338         LOOP_CACHING_NOWAIT = 0,
7339         LOOP_CACHING_WAIT = 1,
7340         LOOP_ALLOC_CHUNK = 2,
7341         LOOP_NO_EMPTY_SIZE = 3,
7342 };
7343
7344 static inline void
7345 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7346                        int delalloc)
7347 {
7348         if (delalloc)
7349                 down_read(&cache->data_rwsem);
7350 }
7351
7352 static inline void
7353 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7354                        int delalloc)
7355 {
7356         btrfs_get_block_group(cache);
7357         if (delalloc)
7358                 down_read(&cache->data_rwsem);
7359 }
7360
7361 static struct btrfs_block_group_cache *
7362 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7363                    struct btrfs_free_cluster *cluster,
7364                    int delalloc)
7365 {
7366         struct btrfs_block_group_cache *used_bg = NULL;
7367
7368         spin_lock(&cluster->refill_lock);
7369         while (1) {
7370                 used_bg = cluster->block_group;
7371                 if (!used_bg)
7372                         return NULL;
7373
7374                 if (used_bg == block_group)
7375                         return used_bg;
7376
7377                 btrfs_get_block_group(used_bg);
7378
7379                 if (!delalloc)
7380                         return used_bg;
7381
7382                 if (down_read_trylock(&used_bg->data_rwsem))
7383                         return used_bg;
7384
7385                 spin_unlock(&cluster->refill_lock);
7386
7387                 /* We should only have one-level nested. */
7388                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7389
7390                 spin_lock(&cluster->refill_lock);
7391                 if (used_bg == cluster->block_group)
7392                         return used_bg;
7393
7394                 up_read(&used_bg->data_rwsem);
7395                 btrfs_put_block_group(used_bg);
7396         }
7397 }
7398
7399 static inline void
7400 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7401                          int delalloc)
7402 {
7403         if (delalloc)
7404                 up_read(&cache->data_rwsem);
7405         btrfs_put_block_group(cache);
7406 }
7407
7408 /*
7409  * walks the btree of allocated extents and find a hole of a given size.
7410  * The key ins is changed to record the hole:
7411  * ins->objectid == start position
7412  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7413  * ins->offset == the size of the hole.
7414  * Any available blocks before search_start are skipped.
7415  *
7416  * If there is no suitable free space, we will record the max size of
7417  * the free space extent currently.
7418  */
7419 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7420                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7421                                 u64 hint_byte, struct btrfs_key *ins,
7422                                 u64 flags, int delalloc)
7423 {
7424         int ret = 0;
7425         struct btrfs_root *root = fs_info->extent_root;
7426         struct btrfs_free_cluster *last_ptr = NULL;
7427         struct btrfs_block_group_cache *block_group = NULL;
7428         u64 search_start = 0;
7429         u64 max_extent_size = 0;
7430         u64 empty_cluster = 0;
7431         struct btrfs_space_info *space_info;
7432         int loop = 0;
7433         int index = __get_raid_index(flags);
7434         bool failed_cluster_refill = false;
7435         bool failed_alloc = false;
7436         bool use_cluster = true;
7437         bool have_caching_bg = false;
7438         bool orig_have_caching_bg = false;
7439         bool full_search = false;
7440
7441         WARN_ON(num_bytes < fs_info->sectorsize);
7442         ins->type = BTRFS_EXTENT_ITEM_KEY;
7443         ins->objectid = 0;
7444         ins->offset = 0;
7445
7446         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7447
7448         space_info = __find_space_info(fs_info, flags);
7449         if (!space_info) {
7450                 btrfs_err(fs_info, "No space info for %llu", flags);
7451                 return -ENOSPC;
7452         }
7453
7454         /*
7455          * If our free space is heavily fragmented we may not be able to make
7456          * big contiguous allocations, so instead of doing the expensive search
7457          * for free space, simply return ENOSPC with our max_extent_size so we
7458          * can go ahead and search for a more manageable chunk.
7459          *
7460          * If our max_extent_size is large enough for our allocation simply
7461          * disable clustering since we will likely not be able to find enough
7462          * space to create a cluster and induce latency trying.
7463          */
7464         if (unlikely(space_info->max_extent_size)) {
7465                 spin_lock(&space_info->lock);
7466                 if (space_info->max_extent_size &&
7467                     num_bytes > space_info->max_extent_size) {
7468                         ins->offset = space_info->max_extent_size;
7469                         spin_unlock(&space_info->lock);
7470                         return -ENOSPC;
7471                 } else if (space_info->max_extent_size) {
7472                         use_cluster = false;
7473                 }
7474                 spin_unlock(&space_info->lock);
7475         }
7476
7477         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7478         if (last_ptr) {
7479                 spin_lock(&last_ptr->lock);
7480                 if (last_ptr->block_group)
7481                         hint_byte = last_ptr->window_start;
7482                 if (last_ptr->fragmented) {
7483                         /*
7484                          * We still set window_start so we can keep track of the
7485                          * last place we found an allocation to try and save
7486                          * some time.
7487                          */
7488                         hint_byte = last_ptr->window_start;
7489                         use_cluster = false;
7490                 }
7491                 spin_unlock(&last_ptr->lock);
7492         }
7493
7494         search_start = max(search_start, first_logical_byte(fs_info, 0));
7495         search_start = max(search_start, hint_byte);
7496         if (search_start == hint_byte) {
7497                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7498                 /*
7499                  * we don't want to use the block group if it doesn't match our
7500                  * allocation bits, or if its not cached.
7501                  *
7502                  * However if we are re-searching with an ideal block group
7503                  * picked out then we don't care that the block group is cached.
7504                  */
7505                 if (block_group && block_group_bits(block_group, flags) &&
7506                     block_group->cached != BTRFS_CACHE_NO) {
7507                         down_read(&space_info->groups_sem);
7508                         if (list_empty(&block_group->list) ||
7509                             block_group->ro) {
7510                                 /*
7511                                  * someone is removing this block group,
7512                                  * we can't jump into the have_block_group
7513                                  * target because our list pointers are not
7514                                  * valid
7515                                  */
7516                                 btrfs_put_block_group(block_group);
7517                                 up_read(&space_info->groups_sem);
7518                         } else {
7519                                 index = get_block_group_index(block_group);
7520                                 btrfs_lock_block_group(block_group, delalloc);
7521                                 goto have_block_group;
7522                         }
7523                 } else if (block_group) {
7524                         btrfs_put_block_group(block_group);
7525                 }
7526         }
7527 search:
7528         have_caching_bg = false;
7529         if (index == 0 || index == __get_raid_index(flags))
7530                 full_search = true;
7531         down_read(&space_info->groups_sem);
7532         list_for_each_entry(block_group, &space_info->block_groups[index],
7533                             list) {
7534                 u64 offset;
7535                 int cached;
7536
7537                 btrfs_grab_block_group(block_group, delalloc);
7538                 search_start = block_group->key.objectid;
7539
7540                 /*
7541                  * this can happen if we end up cycling through all the
7542                  * raid types, but we want to make sure we only allocate
7543                  * for the proper type.
7544                  */
7545                 if (!block_group_bits(block_group, flags)) {
7546                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7547                                 BTRFS_BLOCK_GROUP_RAID1 |
7548                                 BTRFS_BLOCK_GROUP_RAID5 |
7549                                 BTRFS_BLOCK_GROUP_RAID6 |
7550                                 BTRFS_BLOCK_GROUP_RAID10;
7551
7552                         /*
7553                          * if they asked for extra copies and this block group
7554                          * doesn't provide them, bail.  This does allow us to
7555                          * fill raid0 from raid1.
7556                          */
7557                         if ((flags & extra) && !(block_group->flags & extra))
7558                                 goto loop;
7559                 }
7560
7561 have_block_group:
7562                 cached = block_group_cache_done(block_group);
7563                 if (unlikely(!cached)) {
7564                         have_caching_bg = true;
7565                         ret = cache_block_group(block_group, 0);
7566                         BUG_ON(ret < 0);
7567                         ret = 0;
7568                 }
7569
7570                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7571                         goto loop;
7572                 if (unlikely(block_group->ro))
7573                         goto loop;
7574
7575                 /*
7576                  * Ok we want to try and use the cluster allocator, so
7577                  * lets look there
7578                  */
7579                 if (last_ptr && use_cluster) {
7580                         struct btrfs_block_group_cache *used_block_group;
7581                         unsigned long aligned_cluster;
7582                         /*
7583                          * the refill lock keeps out other
7584                          * people trying to start a new cluster
7585                          */
7586                         used_block_group = btrfs_lock_cluster(block_group,
7587                                                               last_ptr,
7588                                                               delalloc);
7589                         if (!used_block_group)
7590                                 goto refill_cluster;
7591
7592                         if (used_block_group != block_group &&
7593                             (used_block_group->ro ||
7594                              !block_group_bits(used_block_group, flags)))
7595                                 goto release_cluster;
7596
7597                         offset = btrfs_alloc_from_cluster(used_block_group,
7598                                                 last_ptr,
7599                                                 num_bytes,
7600                                                 used_block_group->key.objectid,
7601                                                 &max_extent_size);
7602                         if (offset) {
7603                                 /* we have a block, we're done */
7604                                 spin_unlock(&last_ptr->refill_lock);
7605                                 trace_btrfs_reserve_extent_cluster(fs_info,
7606                                                 used_block_group,
7607                                                 search_start, num_bytes);
7608                                 if (used_block_group != block_group) {
7609                                         btrfs_release_block_group(block_group,
7610                                                                   delalloc);
7611                                         block_group = used_block_group;
7612                                 }
7613                                 goto checks;
7614                         }
7615
7616                         WARN_ON(last_ptr->block_group != used_block_group);
7617 release_cluster:
7618                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7619                          * set up a new clusters, so lets just skip it
7620                          * and let the allocator find whatever block
7621                          * it can find.  If we reach this point, we
7622                          * will have tried the cluster allocator
7623                          * plenty of times and not have found
7624                          * anything, so we are likely way too
7625                          * fragmented for the clustering stuff to find
7626                          * anything.
7627                          *
7628                          * However, if the cluster is taken from the
7629                          * current block group, release the cluster
7630                          * first, so that we stand a better chance of
7631                          * succeeding in the unclustered
7632                          * allocation.  */
7633                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7634                             used_block_group != block_group) {
7635                                 spin_unlock(&last_ptr->refill_lock);
7636                                 btrfs_release_block_group(used_block_group,
7637                                                           delalloc);
7638                                 goto unclustered_alloc;
7639                         }
7640
7641                         /*
7642                          * this cluster didn't work out, free it and
7643                          * start over
7644                          */
7645                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7646
7647                         if (used_block_group != block_group)
7648                                 btrfs_release_block_group(used_block_group,
7649                                                           delalloc);
7650 refill_cluster:
7651                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7652                                 spin_unlock(&last_ptr->refill_lock);
7653                                 goto unclustered_alloc;
7654                         }
7655
7656                         aligned_cluster = max_t(unsigned long,
7657                                                 empty_cluster + empty_size,
7658                                               block_group->full_stripe_len);
7659
7660                         /* allocate a cluster in this block group */
7661                         ret = btrfs_find_space_cluster(fs_info, block_group,
7662                                                        last_ptr, search_start,
7663                                                        num_bytes,
7664                                                        aligned_cluster);
7665                         if (ret == 0) {
7666                                 /*
7667                                  * now pull our allocation out of this
7668                                  * cluster
7669                                  */
7670                                 offset = btrfs_alloc_from_cluster(block_group,
7671                                                         last_ptr,
7672                                                         num_bytes,
7673                                                         search_start,
7674                                                         &max_extent_size);
7675                                 if (offset) {
7676                                         /* we found one, proceed */
7677                                         spin_unlock(&last_ptr->refill_lock);
7678                                         trace_btrfs_reserve_extent_cluster(fs_info,
7679                                                 block_group, search_start,
7680                                                 num_bytes);
7681                                         goto checks;
7682                                 }
7683                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7684                                    && !failed_cluster_refill) {
7685                                 spin_unlock(&last_ptr->refill_lock);
7686
7687                                 failed_cluster_refill = true;
7688                                 wait_block_group_cache_progress(block_group,
7689                                        num_bytes + empty_cluster + empty_size);
7690                                 goto have_block_group;
7691                         }
7692
7693                         /*
7694                          * at this point we either didn't find a cluster
7695                          * or we weren't able to allocate a block from our
7696                          * cluster.  Free the cluster we've been trying
7697                          * to use, and go to the next block group
7698                          */
7699                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7700                         spin_unlock(&last_ptr->refill_lock);
7701                         goto loop;
7702                 }
7703
7704 unclustered_alloc:
7705                 /*
7706                  * We are doing an unclustered alloc, set the fragmented flag so
7707                  * we don't bother trying to setup a cluster again until we get
7708                  * more space.
7709                  */
7710                 if (unlikely(last_ptr)) {
7711                         spin_lock(&last_ptr->lock);
7712                         last_ptr->fragmented = 1;
7713                         spin_unlock(&last_ptr->lock);
7714                 }
7715                 if (cached) {
7716                         struct btrfs_free_space_ctl *ctl =
7717                                 block_group->free_space_ctl;
7718
7719                         spin_lock(&ctl->tree_lock);
7720                         if (ctl->free_space <
7721                             num_bytes + empty_cluster + empty_size) {
7722                                 if (ctl->free_space > max_extent_size)
7723                                         max_extent_size = ctl->free_space;
7724                                 spin_unlock(&ctl->tree_lock);
7725                                 goto loop;
7726                         }
7727                         spin_unlock(&ctl->tree_lock);
7728                 }
7729
7730                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7731                                                     num_bytes, empty_size,
7732                                                     &max_extent_size);
7733                 /*
7734                  * If we didn't find a chunk, and we haven't failed on this
7735                  * block group before, and this block group is in the middle of
7736                  * caching and we are ok with waiting, then go ahead and wait
7737                  * for progress to be made, and set failed_alloc to true.
7738                  *
7739                  * If failed_alloc is true then we've already waited on this
7740                  * block group once and should move on to the next block group.
7741                  */
7742                 if (!offset && !failed_alloc && !cached &&
7743                     loop > LOOP_CACHING_NOWAIT) {
7744                         wait_block_group_cache_progress(block_group,
7745                                                 num_bytes + empty_size);
7746                         failed_alloc = true;
7747                         goto have_block_group;
7748                 } else if (!offset) {
7749                         goto loop;
7750                 }
7751 checks:
7752                 search_start = ALIGN(offset, fs_info->stripesize);
7753
7754                 /* move on to the next group */
7755                 if (search_start + num_bytes >
7756                     block_group->key.objectid + block_group->key.offset) {
7757                         btrfs_add_free_space(block_group, offset, num_bytes);
7758                         goto loop;
7759                 }
7760
7761                 if (offset < search_start)
7762                         btrfs_add_free_space(block_group, offset,
7763                                              search_start - offset);
7764                 BUG_ON(offset > search_start);
7765
7766                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7767                                 num_bytes, delalloc);
7768                 if (ret == -EAGAIN) {
7769                         btrfs_add_free_space(block_group, offset, num_bytes);
7770                         goto loop;
7771                 }
7772                 btrfs_inc_block_group_reservations(block_group);
7773
7774                 /* we are all good, lets return */
7775                 ins->objectid = search_start;
7776                 ins->offset = num_bytes;
7777
7778                 trace_btrfs_reserve_extent(fs_info, block_group,
7779                                            search_start, num_bytes);
7780                 btrfs_release_block_group(block_group, delalloc);
7781                 break;
7782 loop:
7783                 failed_cluster_refill = false;
7784                 failed_alloc = false;
7785                 BUG_ON(index != get_block_group_index(block_group));
7786                 btrfs_release_block_group(block_group, delalloc);
7787         }
7788         up_read(&space_info->groups_sem);
7789
7790         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7791                 && !orig_have_caching_bg)
7792                 orig_have_caching_bg = true;
7793
7794         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7795                 goto search;
7796
7797         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7798                 goto search;
7799
7800         /*
7801          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7802          *                      caching kthreads as we move along
7803          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7804          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7805          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7806          *                      again
7807          */
7808         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7809                 index = 0;
7810                 if (loop == LOOP_CACHING_NOWAIT) {
7811                         /*
7812                          * We want to skip the LOOP_CACHING_WAIT step if we
7813                          * don't have any uncached bgs and we've already done a
7814                          * full search through.
7815                          */
7816                         if (orig_have_caching_bg || !full_search)
7817                                 loop = LOOP_CACHING_WAIT;
7818                         else
7819                                 loop = LOOP_ALLOC_CHUNK;
7820                 } else {
7821                         loop++;
7822                 }
7823
7824                 if (loop == LOOP_ALLOC_CHUNK) {
7825                         struct btrfs_trans_handle *trans;
7826                         int exist = 0;
7827
7828                         trans = current->journal_info;
7829                         if (trans)
7830                                 exist = 1;
7831                         else
7832                                 trans = btrfs_join_transaction(root);
7833
7834                         if (IS_ERR(trans)) {
7835                                 ret = PTR_ERR(trans);
7836                                 goto out;
7837                         }
7838
7839                         ret = do_chunk_alloc(trans, fs_info, flags,
7840                                              CHUNK_ALLOC_FORCE);
7841
7842                         /*
7843                          * If we can't allocate a new chunk we've already looped
7844                          * through at least once, move on to the NO_EMPTY_SIZE
7845                          * case.
7846                          */
7847                         if (ret == -ENOSPC)
7848                                 loop = LOOP_NO_EMPTY_SIZE;
7849
7850                         /*
7851                          * Do not bail out on ENOSPC since we
7852                          * can do more things.
7853                          */
7854                         if (ret < 0 && ret != -ENOSPC)
7855                                 btrfs_abort_transaction(trans, ret);
7856                         else
7857                                 ret = 0;
7858                         if (!exist)
7859                                 btrfs_end_transaction(trans);
7860                         if (ret)
7861                                 goto out;
7862                 }
7863
7864                 if (loop == LOOP_NO_EMPTY_SIZE) {
7865                         /*
7866                          * Don't loop again if we already have no empty_size and
7867                          * no empty_cluster.
7868                          */
7869                         if (empty_size == 0 &&
7870                             empty_cluster == 0) {
7871                                 ret = -ENOSPC;
7872                                 goto out;
7873                         }
7874                         empty_size = 0;
7875                         empty_cluster = 0;
7876                 }
7877
7878                 goto search;
7879         } else if (!ins->objectid) {
7880                 ret = -ENOSPC;
7881         } else if (ins->objectid) {
7882                 if (!use_cluster && last_ptr) {
7883                         spin_lock(&last_ptr->lock);
7884                         last_ptr->window_start = ins->objectid;
7885                         spin_unlock(&last_ptr->lock);
7886                 }
7887                 ret = 0;
7888         }
7889 out:
7890         if (ret == -ENOSPC) {
7891                 spin_lock(&space_info->lock);
7892                 space_info->max_extent_size = max_extent_size;
7893                 spin_unlock(&space_info->lock);
7894                 ins->offset = max_extent_size;
7895         }
7896         return ret;
7897 }
7898
7899 static void dump_space_info(struct btrfs_fs_info *fs_info,
7900                             struct btrfs_space_info *info, u64 bytes,
7901                             int dump_block_groups)
7902 {
7903         struct btrfs_block_group_cache *cache;
7904         int index = 0;
7905
7906         spin_lock(&info->lock);
7907         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7908                    info->flags,
7909                    info->total_bytes - btrfs_space_info_used(info, true),
7910                    info->full ? "" : "not ");
7911         btrfs_info(fs_info,
7912                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7913                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7914                 info->bytes_reserved, info->bytes_may_use,
7915                 info->bytes_readonly);
7916         spin_unlock(&info->lock);
7917
7918         if (!dump_block_groups)
7919                 return;
7920
7921         down_read(&info->groups_sem);
7922 again:
7923         list_for_each_entry(cache, &info->block_groups[index], list) {
7924                 spin_lock(&cache->lock);
7925                 btrfs_info(fs_info,
7926                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7927                         cache->key.objectid, cache->key.offset,
7928                         btrfs_block_group_used(&cache->item), cache->pinned,
7929                         cache->reserved, cache->ro ? "[readonly]" : "");
7930                 btrfs_dump_free_space(cache, bytes);
7931                 spin_unlock(&cache->lock);
7932         }
7933         if (++index < BTRFS_NR_RAID_TYPES)
7934                 goto again;
7935         up_read(&info->groups_sem);
7936 }
7937
7938 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7939                          u64 num_bytes, u64 min_alloc_size,
7940                          u64 empty_size, u64 hint_byte,
7941                          struct btrfs_key *ins, int is_data, int delalloc)
7942 {
7943         struct btrfs_fs_info *fs_info = root->fs_info;
7944         bool final_tried = num_bytes == min_alloc_size;
7945         u64 flags;
7946         int ret;
7947
7948         flags = btrfs_get_alloc_profile(root, is_data);
7949 again:
7950         WARN_ON(num_bytes < fs_info->sectorsize);
7951         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
7952                                hint_byte, ins, flags, delalloc);
7953         if (!ret && !is_data) {
7954                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
7955         } else if (ret == -ENOSPC) {
7956                 if (!final_tried && ins->offset) {
7957                         num_bytes = min(num_bytes >> 1, ins->offset);
7958                         num_bytes = round_down(num_bytes,
7959                                                fs_info->sectorsize);
7960                         num_bytes = max(num_bytes, min_alloc_size);
7961                         ram_bytes = num_bytes;
7962                         if (num_bytes == min_alloc_size)
7963                                 final_tried = true;
7964                         goto again;
7965                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
7966                         struct btrfs_space_info *sinfo;
7967
7968                         sinfo = __find_space_info(fs_info, flags);
7969                         btrfs_err(fs_info,
7970                                   "allocation failed flags %llu, wanted %llu",
7971                                   flags, num_bytes);
7972                         if (sinfo)
7973                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
7974                 }
7975         }
7976
7977         return ret;
7978 }
7979
7980 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7981                                         u64 start, u64 len,
7982                                         int pin, int delalloc)
7983 {
7984         struct btrfs_block_group_cache *cache;
7985         int ret = 0;
7986
7987         cache = btrfs_lookup_block_group(fs_info, start);
7988         if (!cache) {
7989                 btrfs_err(fs_info, "Unable to find block group for %llu",
7990                           start);
7991                 return -ENOSPC;
7992         }
7993
7994         if (pin)
7995                 pin_down_extent(fs_info, cache, start, len, 1);
7996         else {
7997                 if (btrfs_test_opt(fs_info, DISCARD))
7998                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
7999                 btrfs_add_free_space(cache, start, len);
8000                 btrfs_free_reserved_bytes(cache, len, delalloc);
8001                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8002         }
8003
8004         btrfs_put_block_group(cache);
8005         return ret;
8006 }
8007
8008 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8009                                u64 start, u64 len, int delalloc)
8010 {
8011         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8012 }
8013
8014 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8015                                        u64 start, u64 len)
8016 {
8017         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8018 }
8019
8020 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8021                                       struct btrfs_fs_info *fs_info,
8022                                       u64 parent, u64 root_objectid,
8023                                       u64 flags, u64 owner, u64 offset,
8024                                       struct btrfs_key *ins, int ref_mod)
8025 {
8026         int ret;
8027         struct btrfs_extent_item *extent_item;
8028         struct btrfs_extent_inline_ref *iref;
8029         struct btrfs_path *path;
8030         struct extent_buffer *leaf;
8031         int type;
8032         u32 size;
8033
8034         if (parent > 0)
8035                 type = BTRFS_SHARED_DATA_REF_KEY;
8036         else
8037                 type = BTRFS_EXTENT_DATA_REF_KEY;
8038
8039         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8040
8041         path = btrfs_alloc_path();
8042         if (!path)
8043                 return -ENOMEM;
8044
8045         path->leave_spinning = 1;
8046         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8047                                       ins, size);
8048         if (ret) {
8049                 btrfs_free_path(path);
8050                 return ret;
8051         }
8052
8053         leaf = path->nodes[0];
8054         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8055                                      struct btrfs_extent_item);
8056         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8057         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8058         btrfs_set_extent_flags(leaf, extent_item,
8059                                flags | BTRFS_EXTENT_FLAG_DATA);
8060
8061         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8062         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8063         if (parent > 0) {
8064                 struct btrfs_shared_data_ref *ref;
8065                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8066                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8067                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8068         } else {
8069                 struct btrfs_extent_data_ref *ref;
8070                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8071                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8072                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8073                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8074                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8075         }
8076
8077         btrfs_mark_buffer_dirty(path->nodes[0]);
8078         btrfs_free_path(path);
8079
8080         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8081                                           ins->offset);
8082         if (ret)
8083                 return ret;
8084
8085         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8086         if (ret) { /* -ENOENT, logic error */
8087                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8088                         ins->objectid, ins->offset);
8089                 BUG();
8090         }
8091         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8092         return ret;
8093 }
8094
8095 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8096                                      struct btrfs_fs_info *fs_info,
8097                                      u64 parent, u64 root_objectid,
8098                                      u64 flags, struct btrfs_disk_key *key,
8099                                      int level, struct btrfs_key *ins)
8100 {
8101         int ret;
8102         struct btrfs_extent_item *extent_item;
8103         struct btrfs_tree_block_info *block_info;
8104         struct btrfs_extent_inline_ref *iref;
8105         struct btrfs_path *path;
8106         struct extent_buffer *leaf;
8107         u32 size = sizeof(*extent_item) + sizeof(*iref);
8108         u64 num_bytes = ins->offset;
8109         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8110
8111         if (!skinny_metadata)
8112                 size += sizeof(*block_info);
8113
8114         path = btrfs_alloc_path();
8115         if (!path) {
8116                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8117                                                    fs_info->nodesize);
8118                 return -ENOMEM;
8119         }
8120
8121         path->leave_spinning = 1;
8122         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8123                                       ins, size);
8124         if (ret) {
8125                 btrfs_free_path(path);
8126                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8127                                                    fs_info->nodesize);
8128                 return ret;
8129         }
8130
8131         leaf = path->nodes[0];
8132         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8133                                      struct btrfs_extent_item);
8134         btrfs_set_extent_refs(leaf, extent_item, 1);
8135         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8136         btrfs_set_extent_flags(leaf, extent_item,
8137                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8138
8139         if (skinny_metadata) {
8140                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8141                 num_bytes = fs_info->nodesize;
8142         } else {
8143                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8144                 btrfs_set_tree_block_key(leaf, block_info, key);
8145                 btrfs_set_tree_block_level(leaf, block_info, level);
8146                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8147         }
8148
8149         if (parent > 0) {
8150                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8151                 btrfs_set_extent_inline_ref_type(leaf, iref,
8152                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8153                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8154         } else {
8155                 btrfs_set_extent_inline_ref_type(leaf, iref,
8156                                                  BTRFS_TREE_BLOCK_REF_KEY);
8157                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8158         }
8159
8160         btrfs_mark_buffer_dirty(leaf);
8161         btrfs_free_path(path);
8162
8163         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8164                                           num_bytes);
8165         if (ret)
8166                 return ret;
8167
8168         ret = update_block_group(trans, fs_info, ins->objectid,
8169                                  fs_info->nodesize, 1);
8170         if (ret) { /* -ENOENT, logic error */
8171                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8172                         ins->objectid, ins->offset);
8173                 BUG();
8174         }
8175
8176         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid,
8177                                           fs_info->nodesize);
8178         return ret;
8179 }
8180
8181 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8182                                      u64 root_objectid, u64 owner,
8183                                      u64 offset, u64 ram_bytes,
8184                                      struct btrfs_key *ins)
8185 {
8186         struct btrfs_fs_info *fs_info = trans->fs_info;
8187         int ret;
8188
8189         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
8190
8191         ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid,
8192                                          ins->offset, 0,
8193                                          root_objectid, owner, offset,
8194                                          ram_bytes, BTRFS_ADD_DELAYED_EXTENT);
8195         return ret;
8196 }
8197
8198 /*
8199  * this is used by the tree logging recovery code.  It records that
8200  * an extent has been allocated and makes sure to clear the free
8201  * space cache bits as well
8202  */
8203 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8204                                    struct btrfs_fs_info *fs_info,
8205                                    u64 root_objectid, u64 owner, u64 offset,
8206                                    struct btrfs_key *ins)
8207 {
8208         int ret;
8209         struct btrfs_block_group_cache *block_group;
8210         struct btrfs_space_info *space_info;
8211
8212         /*
8213          * Mixed block groups will exclude before processing the log so we only
8214          * need to do the exclude dance if this fs isn't mixed.
8215          */
8216         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8217                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8218                                               ins->offset);
8219                 if (ret)
8220                         return ret;
8221         }
8222
8223         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8224         if (!block_group)
8225                 return -EINVAL;
8226
8227         space_info = block_group->space_info;
8228         spin_lock(&space_info->lock);
8229         spin_lock(&block_group->lock);
8230         space_info->bytes_reserved += ins->offset;
8231         block_group->reserved += ins->offset;
8232         spin_unlock(&block_group->lock);
8233         spin_unlock(&space_info->lock);
8234
8235         ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid,
8236                                          0, owner, offset, ins, 1);
8237         btrfs_put_block_group(block_group);
8238         return ret;
8239 }
8240
8241 static struct extent_buffer *
8242 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8243                       u64 bytenr, int level)
8244 {
8245         struct btrfs_fs_info *fs_info = root->fs_info;
8246         struct extent_buffer *buf;
8247
8248         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8249         if (IS_ERR(buf))
8250                 return buf;
8251
8252         btrfs_set_header_generation(buf, trans->transid);
8253         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8254         btrfs_tree_lock(buf);
8255         clean_tree_block(fs_info, buf);
8256         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8257
8258         btrfs_set_lock_blocking(buf);
8259         set_extent_buffer_uptodate(buf);
8260
8261         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8262                 buf->log_index = root->log_transid % 2;
8263                 /*
8264                  * we allow two log transactions at a time, use different
8265                  * EXENT bit to differentiate dirty pages.
8266                  */
8267                 if (buf->log_index == 0)
8268                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8269                                         buf->start + buf->len - 1, GFP_NOFS);
8270                 else
8271                         set_extent_new(&root->dirty_log_pages, buf->start,
8272                                         buf->start + buf->len - 1);
8273         } else {
8274                 buf->log_index = -1;
8275                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8276                          buf->start + buf->len - 1, GFP_NOFS);
8277         }
8278         trans->dirty = true;
8279         /* this returns a buffer locked for blocking */
8280         return buf;
8281 }
8282
8283 static struct btrfs_block_rsv *
8284 use_block_rsv(struct btrfs_trans_handle *trans,
8285               struct btrfs_root *root, u32 blocksize)
8286 {
8287         struct btrfs_fs_info *fs_info = root->fs_info;
8288         struct btrfs_block_rsv *block_rsv;
8289         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8290         int ret;
8291         bool global_updated = false;
8292
8293         block_rsv = get_block_rsv(trans, root);
8294
8295         if (unlikely(block_rsv->size == 0))
8296                 goto try_reserve;
8297 again:
8298         ret = block_rsv_use_bytes(block_rsv, blocksize);
8299         if (!ret)
8300                 return block_rsv;
8301
8302         if (block_rsv->failfast)
8303                 return ERR_PTR(ret);
8304
8305         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8306                 global_updated = true;
8307                 update_global_block_rsv(fs_info);
8308                 goto again;
8309         }
8310
8311         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8312                 static DEFINE_RATELIMIT_STATE(_rs,
8313                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8314                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8315                 if (__ratelimit(&_rs))
8316                         WARN(1, KERN_DEBUG
8317                                 "BTRFS: block rsv returned %d\n", ret);
8318         }
8319 try_reserve:
8320         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8321                                      BTRFS_RESERVE_NO_FLUSH);
8322         if (!ret)
8323                 return block_rsv;
8324         /*
8325          * If we couldn't reserve metadata bytes try and use some from
8326          * the global reserve if its space type is the same as the global
8327          * reservation.
8328          */
8329         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8330             block_rsv->space_info == global_rsv->space_info) {
8331                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8332                 if (!ret)
8333                         return global_rsv;
8334         }
8335         return ERR_PTR(ret);
8336 }
8337
8338 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8339                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8340 {
8341         block_rsv_add_bytes(block_rsv, blocksize, 0);
8342         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8343 }
8344
8345 /*
8346  * finds a free extent and does all the dirty work required for allocation
8347  * returns the tree buffer or an ERR_PTR on error.
8348  */
8349 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8350                                              struct btrfs_root *root,
8351                                              u64 parent, u64 root_objectid,
8352                                              const struct btrfs_disk_key *key,
8353                                              int level, u64 hint,
8354                                              u64 empty_size)
8355 {
8356         struct btrfs_fs_info *fs_info = root->fs_info;
8357         struct btrfs_key ins;
8358         struct btrfs_block_rsv *block_rsv;
8359         struct extent_buffer *buf;
8360         struct btrfs_delayed_extent_op *extent_op;
8361         u64 flags = 0;
8362         int ret;
8363         u32 blocksize = fs_info->nodesize;
8364         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8365
8366 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8367         if (btrfs_is_testing(fs_info)) {
8368                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8369                                             level);
8370                 if (!IS_ERR(buf))
8371                         root->alloc_bytenr += blocksize;
8372                 return buf;
8373         }
8374 #endif
8375
8376         block_rsv = use_block_rsv(trans, root, blocksize);
8377         if (IS_ERR(block_rsv))
8378                 return ERR_CAST(block_rsv);
8379
8380         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8381                                    empty_size, hint, &ins, 0, 0);
8382         if (ret)
8383                 goto out_unuse;
8384
8385         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8386         if (IS_ERR(buf)) {
8387                 ret = PTR_ERR(buf);
8388                 goto out_free_reserved;
8389         }
8390
8391         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8392                 if (parent == 0)
8393                         parent = ins.objectid;
8394                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8395         } else
8396                 BUG_ON(parent > 0);
8397
8398         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8399                 extent_op = btrfs_alloc_delayed_extent_op();
8400                 if (!extent_op) {
8401                         ret = -ENOMEM;
8402                         goto out_free_buf;
8403                 }
8404                 if (key)
8405                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8406                 else
8407                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8408                 extent_op->flags_to_set = flags;
8409                 extent_op->update_key = skinny_metadata ? false : true;
8410                 extent_op->update_flags = true;
8411                 extent_op->is_data = false;
8412                 extent_op->level = level;
8413
8414                 ret = btrfs_add_delayed_tree_ref(fs_info, trans,
8415                                                  ins.objectid, ins.offset,
8416                                                  parent, root_objectid, level,
8417                                                  BTRFS_ADD_DELAYED_EXTENT,
8418                                                  extent_op);
8419                 if (ret)
8420                         goto out_free_delayed;
8421         }
8422         return buf;
8423
8424 out_free_delayed:
8425         btrfs_free_delayed_extent_op(extent_op);
8426 out_free_buf:
8427         free_extent_buffer(buf);
8428 out_free_reserved:
8429         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8430 out_unuse:
8431         unuse_block_rsv(fs_info, block_rsv, blocksize);
8432         return ERR_PTR(ret);
8433 }
8434
8435 struct walk_control {
8436         u64 refs[BTRFS_MAX_LEVEL];
8437         u64 flags[BTRFS_MAX_LEVEL];
8438         struct btrfs_key update_progress;
8439         int stage;
8440         int level;
8441         int shared_level;
8442         int update_ref;
8443         int keep_locks;
8444         int reada_slot;
8445         int reada_count;
8446         int for_reloc;
8447 };
8448
8449 #define DROP_REFERENCE  1
8450 #define UPDATE_BACKREF  2
8451
8452 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8453                                      struct btrfs_root *root,
8454                                      struct walk_control *wc,
8455                                      struct btrfs_path *path)
8456 {
8457         struct btrfs_fs_info *fs_info = root->fs_info;
8458         u64 bytenr;
8459         u64 generation;
8460         u64 refs;
8461         u64 flags;
8462         u32 nritems;
8463         struct btrfs_key key;
8464         struct extent_buffer *eb;
8465         int ret;
8466         int slot;
8467         int nread = 0;
8468
8469         if (path->slots[wc->level] < wc->reada_slot) {
8470                 wc->reada_count = wc->reada_count * 2 / 3;
8471                 wc->reada_count = max(wc->reada_count, 2);
8472         } else {
8473                 wc->reada_count = wc->reada_count * 3 / 2;
8474                 wc->reada_count = min_t(int, wc->reada_count,
8475                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8476         }
8477
8478         eb = path->nodes[wc->level];
8479         nritems = btrfs_header_nritems(eb);
8480
8481         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8482                 if (nread >= wc->reada_count)
8483                         break;
8484
8485                 cond_resched();
8486                 bytenr = btrfs_node_blockptr(eb, slot);
8487                 generation = btrfs_node_ptr_generation(eb, slot);
8488
8489                 if (slot == path->slots[wc->level])
8490                         goto reada;
8491
8492                 if (wc->stage == UPDATE_BACKREF &&
8493                     generation <= root->root_key.offset)
8494                         continue;
8495
8496                 /* We don't lock the tree block, it's OK to be racy here */
8497                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8498                                                wc->level - 1, 1, &refs,
8499                                                &flags);
8500                 /* We don't care about errors in readahead. */
8501                 if (ret < 0)
8502                         continue;
8503                 BUG_ON(refs == 0);
8504
8505                 if (wc->stage == DROP_REFERENCE) {
8506                         if (refs == 1)
8507                                 goto reada;
8508
8509                         if (wc->level == 1 &&
8510                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8511                                 continue;
8512                         if (!wc->update_ref ||
8513                             generation <= root->root_key.offset)
8514                                 continue;
8515                         btrfs_node_key_to_cpu(eb, &key, slot);
8516                         ret = btrfs_comp_cpu_keys(&key,
8517                                                   &wc->update_progress);
8518                         if (ret < 0)
8519                                 continue;
8520                 } else {
8521                         if (wc->level == 1 &&
8522                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8523                                 continue;
8524                 }
8525 reada:
8526                 readahead_tree_block(fs_info, bytenr);
8527                 nread++;
8528         }
8529         wc->reada_slot = slot;
8530 }
8531
8532 /*
8533  * helper to process tree block while walking down the tree.
8534  *
8535  * when wc->stage == UPDATE_BACKREF, this function updates
8536  * back refs for pointers in the block.
8537  *
8538  * NOTE: return value 1 means we should stop walking down.
8539  */
8540 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8541                                    struct btrfs_root *root,
8542                                    struct btrfs_path *path,
8543                                    struct walk_control *wc, int lookup_info)
8544 {
8545         struct btrfs_fs_info *fs_info = root->fs_info;
8546         int level = wc->level;
8547         struct extent_buffer *eb = path->nodes[level];
8548         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8549         int ret;
8550
8551         if (wc->stage == UPDATE_BACKREF &&
8552             btrfs_header_owner(eb) != root->root_key.objectid)
8553                 return 1;
8554
8555         /*
8556          * when reference count of tree block is 1, it won't increase
8557          * again. once full backref flag is set, we never clear it.
8558          */
8559         if (lookup_info &&
8560             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8561              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8562                 BUG_ON(!path->locks[level]);
8563                 ret = btrfs_lookup_extent_info(trans, fs_info,
8564                                                eb->start, level, 1,
8565                                                &wc->refs[level],
8566                                                &wc->flags[level]);
8567                 BUG_ON(ret == -ENOMEM);
8568                 if (ret)
8569                         return ret;
8570                 BUG_ON(wc->refs[level] == 0);
8571         }
8572
8573         if (wc->stage == DROP_REFERENCE) {
8574                 if (wc->refs[level] > 1)
8575                         return 1;
8576
8577                 if (path->locks[level] && !wc->keep_locks) {
8578                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8579                         path->locks[level] = 0;
8580                 }
8581                 return 0;
8582         }
8583
8584         /* wc->stage == UPDATE_BACKREF */
8585         if (!(wc->flags[level] & flag)) {
8586                 BUG_ON(!path->locks[level]);
8587                 ret = btrfs_inc_ref(trans, root, eb, 1);
8588                 BUG_ON(ret); /* -ENOMEM */
8589                 ret = btrfs_dec_ref(trans, root, eb, 0);
8590                 BUG_ON(ret); /* -ENOMEM */
8591                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8592                                                   eb->len, flag,
8593                                                   btrfs_header_level(eb), 0);
8594                 BUG_ON(ret); /* -ENOMEM */
8595                 wc->flags[level] |= flag;
8596         }
8597
8598         /*
8599          * the block is shared by multiple trees, so it's not good to
8600          * keep the tree lock
8601          */
8602         if (path->locks[level] && level > 0) {
8603                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8604                 path->locks[level] = 0;
8605         }
8606         return 0;
8607 }
8608
8609 /*
8610  * helper to process tree block pointer.
8611  *
8612  * when wc->stage == DROP_REFERENCE, this function checks
8613  * reference count of the block pointed to. if the block
8614  * is shared and we need update back refs for the subtree
8615  * rooted at the block, this function changes wc->stage to
8616  * UPDATE_BACKREF. if the block is shared and there is no
8617  * need to update back, this function drops the reference
8618  * to the block.
8619  *
8620  * NOTE: return value 1 means we should stop walking down.
8621  */
8622 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8623                                  struct btrfs_root *root,
8624                                  struct btrfs_path *path,
8625                                  struct walk_control *wc, int *lookup_info)
8626 {
8627         struct btrfs_fs_info *fs_info = root->fs_info;
8628         u64 bytenr;
8629         u64 generation;
8630         u64 parent;
8631         u32 blocksize;
8632         struct btrfs_key key;
8633         struct extent_buffer *next;
8634         int level = wc->level;
8635         int reada = 0;
8636         int ret = 0;
8637         bool need_account = false;
8638
8639         generation = btrfs_node_ptr_generation(path->nodes[level],
8640                                                path->slots[level]);
8641         /*
8642          * if the lower level block was created before the snapshot
8643          * was created, we know there is no need to update back refs
8644          * for the subtree
8645          */
8646         if (wc->stage == UPDATE_BACKREF &&
8647             generation <= root->root_key.offset) {
8648                 *lookup_info = 1;
8649                 return 1;
8650         }
8651
8652         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8653         blocksize = fs_info->nodesize;
8654
8655         next = find_extent_buffer(fs_info, bytenr);
8656         if (!next) {
8657                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8658                 if (IS_ERR(next))
8659                         return PTR_ERR(next);
8660
8661                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8662                                                level - 1);
8663                 reada = 1;
8664         }
8665         btrfs_tree_lock(next);
8666         btrfs_set_lock_blocking(next);
8667
8668         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8669                                        &wc->refs[level - 1],
8670                                        &wc->flags[level - 1]);
8671         if (ret < 0)
8672                 goto out_unlock;
8673
8674         if (unlikely(wc->refs[level - 1] == 0)) {
8675                 btrfs_err(fs_info, "Missing references.");
8676                 ret = -EIO;
8677                 goto out_unlock;
8678         }
8679         *lookup_info = 0;
8680
8681         if (wc->stage == DROP_REFERENCE) {
8682                 if (wc->refs[level - 1] > 1) {
8683                         need_account = true;
8684                         if (level == 1 &&
8685                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8686                                 goto skip;
8687
8688                         if (!wc->update_ref ||
8689                             generation <= root->root_key.offset)
8690                                 goto skip;
8691
8692                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8693                                               path->slots[level]);
8694                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8695                         if (ret < 0)
8696                                 goto skip;
8697
8698                         wc->stage = UPDATE_BACKREF;
8699                         wc->shared_level = level - 1;
8700                 }
8701         } else {
8702                 if (level == 1 &&
8703                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8704                         goto skip;
8705         }
8706
8707         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8708                 btrfs_tree_unlock(next);
8709                 free_extent_buffer(next);
8710                 next = NULL;
8711                 *lookup_info = 1;
8712         }
8713
8714         if (!next) {
8715                 if (reada && level == 1)
8716                         reada_walk_down(trans, root, wc, path);
8717                 next = read_tree_block(fs_info, bytenr, generation);
8718                 if (IS_ERR(next)) {
8719                         return PTR_ERR(next);
8720                 } else if (!extent_buffer_uptodate(next)) {
8721                         free_extent_buffer(next);
8722                         return -EIO;
8723                 }
8724                 btrfs_tree_lock(next);
8725                 btrfs_set_lock_blocking(next);
8726         }
8727
8728         level--;
8729         ASSERT(level == btrfs_header_level(next));
8730         if (level != btrfs_header_level(next)) {
8731                 btrfs_err(root->fs_info, "mismatched level");
8732                 ret = -EIO;
8733                 goto out_unlock;
8734         }
8735         path->nodes[level] = next;
8736         path->slots[level] = 0;
8737         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8738         wc->level = level;
8739         if (wc->level == 1)
8740                 wc->reada_slot = 0;
8741         return 0;
8742 skip:
8743         wc->refs[level - 1] = 0;
8744         wc->flags[level - 1] = 0;
8745         if (wc->stage == DROP_REFERENCE) {
8746                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8747                         parent = path->nodes[level]->start;
8748                 } else {
8749                         ASSERT(root->root_key.objectid ==
8750                                btrfs_header_owner(path->nodes[level]));
8751                         if (root->root_key.objectid !=
8752                             btrfs_header_owner(path->nodes[level])) {
8753                                 btrfs_err(root->fs_info,
8754                                                 "mismatched block owner");
8755                                 ret = -EIO;
8756                                 goto out_unlock;
8757                         }
8758                         parent = 0;
8759                 }
8760
8761                 if (need_account) {
8762                         ret = btrfs_qgroup_trace_subtree(trans, root, next,
8763                                                          generation, level - 1);
8764                         if (ret) {
8765                                 btrfs_err_rl(fs_info,
8766                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8767                                              ret);
8768                         }
8769                 }
8770                 ret = btrfs_free_extent(trans, fs_info, bytenr, blocksize,
8771                                         parent, root->root_key.objectid,
8772                                         level - 1, 0);
8773                 if (ret)
8774                         goto out_unlock;
8775         }
8776
8777         *lookup_info = 1;
8778         ret = 1;
8779
8780 out_unlock:
8781         btrfs_tree_unlock(next);
8782         free_extent_buffer(next);
8783
8784         return ret;
8785 }
8786
8787 /*
8788  * helper to process tree block while walking up the tree.
8789  *
8790  * when wc->stage == DROP_REFERENCE, this function drops
8791  * reference count on the block.
8792  *
8793  * when wc->stage == UPDATE_BACKREF, this function changes
8794  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8795  * to UPDATE_BACKREF previously while processing the block.
8796  *
8797  * NOTE: return value 1 means we should stop walking up.
8798  */
8799 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8800                                  struct btrfs_root *root,
8801                                  struct btrfs_path *path,
8802                                  struct walk_control *wc)
8803 {
8804         struct btrfs_fs_info *fs_info = root->fs_info;
8805         int ret;
8806         int level = wc->level;
8807         struct extent_buffer *eb = path->nodes[level];
8808         u64 parent = 0;
8809
8810         if (wc->stage == UPDATE_BACKREF) {
8811                 BUG_ON(wc->shared_level < level);
8812                 if (level < wc->shared_level)
8813                         goto out;
8814
8815                 ret = find_next_key(path, level + 1, &wc->update_progress);
8816                 if (ret > 0)
8817                         wc->update_ref = 0;
8818
8819                 wc->stage = DROP_REFERENCE;
8820                 wc->shared_level = -1;
8821                 path->slots[level] = 0;
8822
8823                 /*
8824                  * check reference count again if the block isn't locked.
8825                  * we should start walking down the tree again if reference
8826                  * count is one.
8827                  */
8828                 if (!path->locks[level]) {
8829                         BUG_ON(level == 0);
8830                         btrfs_tree_lock(eb);
8831                         btrfs_set_lock_blocking(eb);
8832                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8833
8834                         ret = btrfs_lookup_extent_info(trans, fs_info,
8835                                                        eb->start, level, 1,
8836                                                        &wc->refs[level],
8837                                                        &wc->flags[level]);
8838                         if (ret < 0) {
8839                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8840                                 path->locks[level] = 0;
8841                                 return ret;
8842                         }
8843                         BUG_ON(wc->refs[level] == 0);
8844                         if (wc->refs[level] == 1) {
8845                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8846                                 path->locks[level] = 0;
8847                                 return 1;
8848                         }
8849                 }
8850         }
8851
8852         /* wc->stage == DROP_REFERENCE */
8853         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8854
8855         if (wc->refs[level] == 1) {
8856                 if (level == 0) {
8857                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8858                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8859                         else
8860                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8861                         BUG_ON(ret); /* -ENOMEM */
8862                         ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb);
8863                         if (ret) {
8864                                 btrfs_err_rl(fs_info,
8865                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8866                                              ret);
8867                         }
8868                 }
8869                 /* make block locked assertion in clean_tree_block happy */
8870                 if (!path->locks[level] &&
8871                     btrfs_header_generation(eb) == trans->transid) {
8872                         btrfs_tree_lock(eb);
8873                         btrfs_set_lock_blocking(eb);
8874                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8875                 }
8876                 clean_tree_block(fs_info, eb);
8877         }
8878
8879         if (eb == root->node) {
8880                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8881                         parent = eb->start;
8882                 else
8883                         BUG_ON(root->root_key.objectid !=
8884                                btrfs_header_owner(eb));
8885         } else {
8886                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8887                         parent = path->nodes[level + 1]->start;
8888                 else
8889                         BUG_ON(root->root_key.objectid !=
8890                                btrfs_header_owner(path->nodes[level + 1]));
8891         }
8892
8893         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8894 out:
8895         wc->refs[level] = 0;
8896         wc->flags[level] = 0;
8897         return 0;
8898 }
8899
8900 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8901                                    struct btrfs_root *root,
8902                                    struct btrfs_path *path,
8903                                    struct walk_control *wc)
8904 {
8905         int level = wc->level;
8906         int lookup_info = 1;
8907         int ret;
8908
8909         while (level >= 0) {
8910                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8911                 if (ret > 0)
8912                         break;
8913
8914                 if (level == 0)
8915                         break;
8916
8917                 if (path->slots[level] >=
8918                     btrfs_header_nritems(path->nodes[level]))
8919                         break;
8920
8921                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8922                 if (ret > 0) {
8923                         path->slots[level]++;
8924                         continue;
8925                 } else if (ret < 0)
8926                         return ret;
8927                 level = wc->level;
8928         }
8929         return 0;
8930 }
8931
8932 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8933                                  struct btrfs_root *root,
8934                                  struct btrfs_path *path,
8935                                  struct walk_control *wc, int max_level)
8936 {
8937         int level = wc->level;
8938         int ret;
8939
8940         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8941         while (level < max_level && path->nodes[level]) {
8942                 wc->level = level;
8943                 if (path->slots[level] + 1 <
8944                     btrfs_header_nritems(path->nodes[level])) {
8945                         path->slots[level]++;
8946                         return 0;
8947                 } else {
8948                         ret = walk_up_proc(trans, root, path, wc);
8949                         if (ret > 0)
8950                                 return 0;
8951
8952                         if (path->locks[level]) {
8953                                 btrfs_tree_unlock_rw(path->nodes[level],
8954                                                      path->locks[level]);
8955                                 path->locks[level] = 0;
8956                         }
8957                         free_extent_buffer(path->nodes[level]);
8958                         path->nodes[level] = NULL;
8959                         level++;
8960                 }
8961         }
8962         return 1;
8963 }
8964
8965 /*
8966  * drop a subvolume tree.
8967  *
8968  * this function traverses the tree freeing any blocks that only
8969  * referenced by the tree.
8970  *
8971  * when a shared tree block is found. this function decreases its
8972  * reference count by one. if update_ref is true, this function
8973  * also make sure backrefs for the shared block and all lower level
8974  * blocks are properly updated.
8975  *
8976  * If called with for_reloc == 0, may exit early with -EAGAIN
8977  */
8978 int btrfs_drop_snapshot(struct btrfs_root *root,
8979                          struct btrfs_block_rsv *block_rsv, int update_ref,
8980                          int for_reloc)
8981 {
8982         struct btrfs_fs_info *fs_info = root->fs_info;
8983         struct btrfs_path *path;
8984         struct btrfs_trans_handle *trans;
8985         struct btrfs_root *tree_root = fs_info->tree_root;
8986         struct btrfs_root_item *root_item = &root->root_item;
8987         struct walk_control *wc;
8988         struct btrfs_key key;
8989         int err = 0;
8990         int ret;
8991         int level;
8992         bool root_dropped = false;
8993
8994         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
8995
8996         path = btrfs_alloc_path();
8997         if (!path) {
8998                 err = -ENOMEM;
8999                 goto out;
9000         }
9001
9002         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9003         if (!wc) {
9004                 btrfs_free_path(path);
9005                 err = -ENOMEM;
9006                 goto out;
9007         }
9008
9009         trans = btrfs_start_transaction(tree_root, 0);
9010         if (IS_ERR(trans)) {
9011                 err = PTR_ERR(trans);
9012                 goto out_free;
9013         }
9014
9015         if (block_rsv)
9016                 trans->block_rsv = block_rsv;
9017
9018         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9019                 level = btrfs_header_level(root->node);
9020                 path->nodes[level] = btrfs_lock_root_node(root);
9021                 btrfs_set_lock_blocking(path->nodes[level]);
9022                 path->slots[level] = 0;
9023                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9024                 memset(&wc->update_progress, 0,
9025                        sizeof(wc->update_progress));
9026         } else {
9027                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9028                 memcpy(&wc->update_progress, &key,
9029                        sizeof(wc->update_progress));
9030
9031                 level = root_item->drop_level;
9032                 BUG_ON(level == 0);
9033                 path->lowest_level = level;
9034                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9035                 path->lowest_level = 0;
9036                 if (ret < 0) {
9037                         err = ret;
9038                         goto out_end_trans;
9039                 }
9040                 WARN_ON(ret > 0);
9041
9042                 /*
9043                  * unlock our path, this is safe because only this
9044                  * function is allowed to delete this snapshot
9045                  */
9046                 btrfs_unlock_up_safe(path, 0);
9047
9048                 level = btrfs_header_level(root->node);
9049                 while (1) {
9050                         btrfs_tree_lock(path->nodes[level]);
9051                         btrfs_set_lock_blocking(path->nodes[level]);
9052                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9053
9054                         ret = btrfs_lookup_extent_info(trans, fs_info,
9055                                                 path->nodes[level]->start,
9056                                                 level, 1, &wc->refs[level],
9057                                                 &wc->flags[level]);
9058                         if (ret < 0) {
9059                                 err = ret;
9060                                 goto out_end_trans;
9061                         }
9062                         BUG_ON(wc->refs[level] == 0);
9063
9064                         if (level == root_item->drop_level)
9065                                 break;
9066
9067                         btrfs_tree_unlock(path->nodes[level]);
9068                         path->locks[level] = 0;
9069                         WARN_ON(wc->refs[level] != 1);
9070                         level--;
9071                 }
9072         }
9073
9074         wc->level = level;
9075         wc->shared_level = -1;
9076         wc->stage = DROP_REFERENCE;
9077         wc->update_ref = update_ref;
9078         wc->keep_locks = 0;
9079         wc->for_reloc = for_reloc;
9080         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9081
9082         while (1) {
9083
9084                 ret = walk_down_tree(trans, root, path, wc);
9085                 if (ret < 0) {
9086                         err = ret;
9087                         break;
9088                 }
9089
9090                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9091                 if (ret < 0) {
9092                         err = ret;
9093                         break;
9094                 }
9095
9096                 if (ret > 0) {
9097                         BUG_ON(wc->stage != DROP_REFERENCE);
9098                         break;
9099                 }
9100
9101                 if (wc->stage == DROP_REFERENCE) {
9102                         level = wc->level;
9103                         btrfs_node_key(path->nodes[level],
9104                                        &root_item->drop_progress,
9105                                        path->slots[level]);
9106                         root_item->drop_level = level;
9107                 }
9108
9109                 BUG_ON(wc->level == 0);
9110                 if (btrfs_should_end_transaction(trans) ||
9111                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9112                         ret = btrfs_update_root(trans, tree_root,
9113                                                 &root->root_key,
9114                                                 root_item);
9115                         if (ret) {
9116                                 btrfs_abort_transaction(trans, ret);
9117                                 err = ret;
9118                                 goto out_end_trans;
9119                         }
9120
9121                         btrfs_end_transaction_throttle(trans);
9122                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9123                                 btrfs_debug(fs_info,
9124                                             "drop snapshot early exit");
9125                                 err = -EAGAIN;
9126                                 goto out_free;
9127                         }
9128
9129                         trans = btrfs_start_transaction(tree_root, 0);
9130                         if (IS_ERR(trans)) {
9131                                 err = PTR_ERR(trans);
9132                                 goto out_free;
9133                         }
9134                         if (block_rsv)
9135                                 trans->block_rsv = block_rsv;
9136                 }
9137         }
9138         btrfs_release_path(path);
9139         if (err)
9140                 goto out_end_trans;
9141
9142         ret = btrfs_del_root(trans, tree_root, &root->root_key);
9143         if (ret) {
9144                 btrfs_abort_transaction(trans, ret);
9145                 goto out_end_trans;
9146         }
9147
9148         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9149                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9150                                       NULL, NULL);
9151                 if (ret < 0) {
9152                         btrfs_abort_transaction(trans, ret);
9153                         err = ret;
9154                         goto out_end_trans;
9155                 } else if (ret > 0) {
9156                         /* if we fail to delete the orphan item this time
9157                          * around, it'll get picked up the next time.
9158                          *
9159                          * The most common failure here is just -ENOENT.
9160                          */
9161                         btrfs_del_orphan_item(trans, tree_root,
9162                                               root->root_key.objectid);
9163                 }
9164         }
9165
9166         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9167                 btrfs_add_dropped_root(trans, root);
9168         } else {
9169                 free_extent_buffer(root->node);
9170                 free_extent_buffer(root->commit_root);
9171                 btrfs_put_fs_root(root);
9172         }
9173         root_dropped = true;
9174 out_end_trans:
9175         btrfs_end_transaction_throttle(trans);
9176 out_free:
9177         kfree(wc);
9178         btrfs_free_path(path);
9179 out:
9180         /*
9181          * So if we need to stop dropping the snapshot for whatever reason we
9182          * need to make sure to add it back to the dead root list so that we
9183          * keep trying to do the work later.  This also cleans up roots if we
9184          * don't have it in the radix (like when we recover after a power fail
9185          * or unmount) so we don't leak memory.
9186          */
9187         if (!for_reloc && root_dropped == false)
9188                 btrfs_add_dead_root(root);
9189         if (err && err != -EAGAIN)
9190                 btrfs_handle_fs_error(fs_info, err, NULL);
9191         return err;
9192 }
9193
9194 /*
9195  * drop subtree rooted at tree block 'node'.
9196  *
9197  * NOTE: this function will unlock and release tree block 'node'
9198  * only used by relocation code
9199  */
9200 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9201                         struct btrfs_root *root,
9202                         struct extent_buffer *node,
9203                         struct extent_buffer *parent)
9204 {
9205         struct btrfs_fs_info *fs_info = root->fs_info;
9206         struct btrfs_path *path;
9207         struct walk_control *wc;
9208         int level;
9209         int parent_level;
9210         int ret = 0;
9211         int wret;
9212
9213         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9214
9215         path = btrfs_alloc_path();
9216         if (!path)
9217                 return -ENOMEM;
9218
9219         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9220         if (!wc) {
9221                 btrfs_free_path(path);
9222                 return -ENOMEM;
9223         }
9224
9225         btrfs_assert_tree_locked(parent);
9226         parent_level = btrfs_header_level(parent);
9227         extent_buffer_get(parent);
9228         path->nodes[parent_level] = parent;
9229         path->slots[parent_level] = btrfs_header_nritems(parent);
9230
9231         btrfs_assert_tree_locked(node);
9232         level = btrfs_header_level(node);
9233         path->nodes[level] = node;
9234         path->slots[level] = 0;
9235         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9236
9237         wc->refs[parent_level] = 1;
9238         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9239         wc->level = level;
9240         wc->shared_level = -1;
9241         wc->stage = DROP_REFERENCE;
9242         wc->update_ref = 0;
9243         wc->keep_locks = 1;
9244         wc->for_reloc = 1;
9245         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9246
9247         while (1) {
9248                 wret = walk_down_tree(trans, root, path, wc);
9249                 if (wret < 0) {
9250                         ret = wret;
9251                         break;
9252                 }
9253
9254                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9255                 if (wret < 0)
9256                         ret = wret;
9257                 if (wret != 0)
9258                         break;
9259         }
9260
9261         kfree(wc);
9262         btrfs_free_path(path);
9263         return ret;
9264 }
9265
9266 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9267 {
9268         u64 num_devices;
9269         u64 stripped;
9270
9271         /*
9272          * if restripe for this chunk_type is on pick target profile and
9273          * return, otherwise do the usual balance
9274          */
9275         stripped = get_restripe_target(fs_info, flags);
9276         if (stripped)
9277                 return extended_to_chunk(stripped);
9278
9279         num_devices = fs_info->fs_devices->rw_devices;
9280
9281         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9282                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9283                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9284
9285         if (num_devices == 1) {
9286                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9287                 stripped = flags & ~stripped;
9288
9289                 /* turn raid0 into single device chunks */
9290                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9291                         return stripped;
9292
9293                 /* turn mirroring into duplication */
9294                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9295                              BTRFS_BLOCK_GROUP_RAID10))
9296                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9297         } else {
9298                 /* they already had raid on here, just return */
9299                 if (flags & stripped)
9300                         return flags;
9301
9302                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9303                 stripped = flags & ~stripped;
9304
9305                 /* switch duplicated blocks with raid1 */
9306                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9307                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9308
9309                 /* this is drive concat, leave it alone */
9310         }
9311
9312         return flags;
9313 }
9314
9315 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9316 {
9317         struct btrfs_space_info *sinfo = cache->space_info;
9318         u64 num_bytes;
9319         u64 min_allocable_bytes;
9320         int ret = -ENOSPC;
9321
9322         /*
9323          * We need some metadata space and system metadata space for
9324          * allocating chunks in some corner cases until we force to set
9325          * it to be readonly.
9326          */
9327         if ((sinfo->flags &
9328              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9329             !force)
9330                 min_allocable_bytes = SZ_1M;
9331         else
9332                 min_allocable_bytes = 0;
9333
9334         spin_lock(&sinfo->lock);
9335         spin_lock(&cache->lock);
9336
9337         if (cache->ro) {
9338                 cache->ro++;
9339                 ret = 0;
9340                 goto out;
9341         }
9342
9343         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9344                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9345
9346         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9347             min_allocable_bytes <= sinfo->total_bytes) {
9348                 sinfo->bytes_readonly += num_bytes;
9349                 cache->ro++;
9350                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9351                 ret = 0;
9352         }
9353 out:
9354         spin_unlock(&cache->lock);
9355         spin_unlock(&sinfo->lock);
9356         return ret;
9357 }
9358
9359 int btrfs_inc_block_group_ro(struct btrfs_fs_info *fs_info,
9360                              struct btrfs_block_group_cache *cache)
9361
9362 {
9363         struct btrfs_trans_handle *trans;
9364         u64 alloc_flags;
9365         int ret;
9366
9367 again:
9368         trans = btrfs_join_transaction(fs_info->extent_root);
9369         if (IS_ERR(trans))
9370                 return PTR_ERR(trans);
9371
9372         /*
9373          * we're not allowed to set block groups readonly after the dirty
9374          * block groups cache has started writing.  If it already started,
9375          * back off and let this transaction commit
9376          */
9377         mutex_lock(&fs_info->ro_block_group_mutex);
9378         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9379                 u64 transid = trans->transid;
9380
9381                 mutex_unlock(&fs_info->ro_block_group_mutex);
9382                 btrfs_end_transaction(trans);
9383
9384                 ret = btrfs_wait_for_commit(fs_info, transid);
9385                 if (ret)
9386                         return ret;
9387                 goto again;
9388         }
9389
9390         /*
9391          * if we are changing raid levels, try to allocate a corresponding
9392          * block group with the new raid level.
9393          */
9394         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9395         if (alloc_flags != cache->flags) {
9396                 ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9397                                      CHUNK_ALLOC_FORCE);
9398                 /*
9399                  * ENOSPC is allowed here, we may have enough space
9400                  * already allocated at the new raid level to
9401                  * carry on
9402                  */
9403                 if (ret == -ENOSPC)
9404                         ret = 0;
9405                 if (ret < 0)
9406                         goto out;
9407         }
9408
9409         ret = inc_block_group_ro(cache, 0);
9410         if (!ret)
9411                 goto out;
9412         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9413         ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9414                              CHUNK_ALLOC_FORCE);
9415         if (ret < 0)
9416                 goto out;
9417         ret = inc_block_group_ro(cache, 0);
9418 out:
9419         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9420                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9421                 mutex_lock(&fs_info->chunk_mutex);
9422                 check_system_chunk(trans, fs_info, alloc_flags);
9423                 mutex_unlock(&fs_info->chunk_mutex);
9424         }
9425         mutex_unlock(&fs_info->ro_block_group_mutex);
9426
9427         btrfs_end_transaction(trans);
9428         return ret;
9429 }
9430
9431 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9432                             struct btrfs_fs_info *fs_info, u64 type)
9433 {
9434         u64 alloc_flags = get_alloc_profile(fs_info, type);
9435
9436         return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE);
9437 }
9438
9439 /*
9440  * helper to account the unused space of all the readonly block group in the
9441  * space_info. takes mirrors into account.
9442  */
9443 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9444 {
9445         struct btrfs_block_group_cache *block_group;
9446         u64 free_bytes = 0;
9447         int factor;
9448
9449         /* It's df, we don't care if it's racy */
9450         if (list_empty(&sinfo->ro_bgs))
9451                 return 0;
9452
9453         spin_lock(&sinfo->lock);
9454         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9455                 spin_lock(&block_group->lock);
9456
9457                 if (!block_group->ro) {
9458                         spin_unlock(&block_group->lock);
9459                         continue;
9460                 }
9461
9462                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9463                                           BTRFS_BLOCK_GROUP_RAID10 |
9464                                           BTRFS_BLOCK_GROUP_DUP))
9465                         factor = 2;
9466                 else
9467                         factor = 1;
9468
9469                 free_bytes += (block_group->key.offset -
9470                                btrfs_block_group_used(&block_group->item)) *
9471                                factor;
9472
9473                 spin_unlock(&block_group->lock);
9474         }
9475         spin_unlock(&sinfo->lock);
9476
9477         return free_bytes;
9478 }
9479
9480 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9481 {
9482         struct btrfs_space_info *sinfo = cache->space_info;
9483         u64 num_bytes;
9484
9485         BUG_ON(!cache->ro);
9486
9487         spin_lock(&sinfo->lock);
9488         spin_lock(&cache->lock);
9489         if (!--cache->ro) {
9490                 num_bytes = cache->key.offset - cache->reserved -
9491                             cache->pinned - cache->bytes_super -
9492                             btrfs_block_group_used(&cache->item);
9493                 sinfo->bytes_readonly -= num_bytes;
9494                 list_del_init(&cache->ro_list);
9495         }
9496         spin_unlock(&cache->lock);
9497         spin_unlock(&sinfo->lock);
9498 }
9499
9500 /*
9501  * checks to see if its even possible to relocate this block group.
9502  *
9503  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9504  * ok to go ahead and try.
9505  */
9506 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9507 {
9508         struct btrfs_root *root = fs_info->extent_root;
9509         struct btrfs_block_group_cache *block_group;
9510         struct btrfs_space_info *space_info;
9511         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9512         struct btrfs_device *device;
9513         struct btrfs_trans_handle *trans;
9514         u64 min_free;
9515         u64 dev_min = 1;
9516         u64 dev_nr = 0;
9517         u64 target;
9518         int debug;
9519         int index;
9520         int full = 0;
9521         int ret = 0;
9522
9523         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9524
9525         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9526
9527         /* odd, couldn't find the block group, leave it alone */
9528         if (!block_group) {
9529                 if (debug)
9530                         btrfs_warn(fs_info,
9531                                    "can't find block group for bytenr %llu",
9532                                    bytenr);
9533                 return -1;
9534         }
9535
9536         min_free = btrfs_block_group_used(&block_group->item);
9537
9538         /* no bytes used, we're good */
9539         if (!min_free)
9540                 goto out;
9541
9542         space_info = block_group->space_info;
9543         spin_lock(&space_info->lock);
9544
9545         full = space_info->full;
9546
9547         /*
9548          * if this is the last block group we have in this space, we can't
9549          * relocate it unless we're able to allocate a new chunk below.
9550          *
9551          * Otherwise, we need to make sure we have room in the space to handle
9552          * all of the extents from this block group.  If we can, we're good
9553          */
9554         if ((space_info->total_bytes != block_group->key.offset) &&
9555             (btrfs_space_info_used(space_info, false) + min_free <
9556              space_info->total_bytes)) {
9557                 spin_unlock(&space_info->lock);
9558                 goto out;
9559         }
9560         spin_unlock(&space_info->lock);
9561
9562         /*
9563          * ok we don't have enough space, but maybe we have free space on our
9564          * devices to allocate new chunks for relocation, so loop through our
9565          * alloc devices and guess if we have enough space.  if this block
9566          * group is going to be restriped, run checks against the target
9567          * profile instead of the current one.
9568          */
9569         ret = -1;
9570
9571         /*
9572          * index:
9573          *      0: raid10
9574          *      1: raid1
9575          *      2: dup
9576          *      3: raid0
9577          *      4: single
9578          */
9579         target = get_restripe_target(fs_info, block_group->flags);
9580         if (target) {
9581                 index = __get_raid_index(extended_to_chunk(target));
9582         } else {
9583                 /*
9584                  * this is just a balance, so if we were marked as full
9585                  * we know there is no space for a new chunk
9586                  */
9587                 if (full) {
9588                         if (debug)
9589                                 btrfs_warn(fs_info,
9590                                            "no space to alloc new chunk for block group %llu",
9591                                            block_group->key.objectid);
9592                         goto out;
9593                 }
9594
9595                 index = get_block_group_index(block_group);
9596         }
9597
9598         if (index == BTRFS_RAID_RAID10) {
9599                 dev_min = 4;
9600                 /* Divide by 2 */
9601                 min_free >>= 1;
9602         } else if (index == BTRFS_RAID_RAID1) {
9603                 dev_min = 2;
9604         } else if (index == BTRFS_RAID_DUP) {
9605                 /* Multiply by 2 */
9606                 min_free <<= 1;
9607         } else if (index == BTRFS_RAID_RAID0) {
9608                 dev_min = fs_devices->rw_devices;
9609                 min_free = div64_u64(min_free, dev_min);
9610         }
9611
9612         /* We need to do this so that we can look at pending chunks */
9613         trans = btrfs_join_transaction(root);
9614         if (IS_ERR(trans)) {
9615                 ret = PTR_ERR(trans);
9616                 goto out;
9617         }
9618
9619         mutex_lock(&fs_info->chunk_mutex);
9620         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9621                 u64 dev_offset;
9622
9623                 /*
9624                  * check to make sure we can actually find a chunk with enough
9625                  * space to fit our block group in.
9626                  */
9627                 if (device->total_bytes > device->bytes_used + min_free &&
9628                     !device->is_tgtdev_for_dev_replace) {
9629                         ret = find_free_dev_extent(trans, device, min_free,
9630                                                    &dev_offset, NULL);
9631                         if (!ret)
9632                                 dev_nr++;
9633
9634                         if (dev_nr >= dev_min)
9635                                 break;
9636
9637                         ret = -1;
9638                 }
9639         }
9640         if (debug && ret == -1)
9641                 btrfs_warn(fs_info,
9642                            "no space to allocate a new chunk for block group %llu",
9643                            block_group->key.objectid);
9644         mutex_unlock(&fs_info->chunk_mutex);
9645         btrfs_end_transaction(trans);
9646 out:
9647         btrfs_put_block_group(block_group);
9648         return ret;
9649 }
9650
9651 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9652                                   struct btrfs_path *path,
9653                                   struct btrfs_key *key)
9654 {
9655         struct btrfs_root *root = fs_info->extent_root;
9656         int ret = 0;
9657         struct btrfs_key found_key;
9658         struct extent_buffer *leaf;
9659         int slot;
9660
9661         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9662         if (ret < 0)
9663                 goto out;
9664
9665         while (1) {
9666                 slot = path->slots[0];
9667                 leaf = path->nodes[0];
9668                 if (slot >= btrfs_header_nritems(leaf)) {
9669                         ret = btrfs_next_leaf(root, path);
9670                         if (ret == 0)
9671                                 continue;
9672                         if (ret < 0)
9673                                 goto out;
9674                         break;
9675                 }
9676                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9677
9678                 if (found_key.objectid >= key->objectid &&
9679                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9680                         struct extent_map_tree *em_tree;
9681                         struct extent_map *em;
9682
9683                         em_tree = &root->fs_info->mapping_tree.map_tree;
9684                         read_lock(&em_tree->lock);
9685                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9686                                                    found_key.offset);
9687                         read_unlock(&em_tree->lock);
9688                         if (!em) {
9689                                 btrfs_err(fs_info,
9690                         "logical %llu len %llu found bg but no related chunk",
9691                                           found_key.objectid, found_key.offset);
9692                                 ret = -ENOENT;
9693                         } else {
9694                                 ret = 0;
9695                         }
9696                         free_extent_map(em);
9697                         goto out;
9698                 }
9699                 path->slots[0]++;
9700         }
9701 out:
9702         return ret;
9703 }
9704
9705 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9706 {
9707         struct btrfs_block_group_cache *block_group;
9708         u64 last = 0;
9709
9710         while (1) {
9711                 struct inode *inode;
9712
9713                 block_group = btrfs_lookup_first_block_group(info, last);
9714                 while (block_group) {
9715                         spin_lock(&block_group->lock);
9716                         if (block_group->iref)
9717                                 break;
9718                         spin_unlock(&block_group->lock);
9719                         block_group = next_block_group(info, block_group);
9720                 }
9721                 if (!block_group) {
9722                         if (last == 0)
9723                                 break;
9724                         last = 0;
9725                         continue;
9726                 }
9727
9728                 inode = block_group->inode;
9729                 block_group->iref = 0;
9730                 block_group->inode = NULL;
9731                 spin_unlock(&block_group->lock);
9732                 ASSERT(block_group->io_ctl.inode == NULL);
9733                 iput(inode);
9734                 last = block_group->key.objectid + block_group->key.offset;
9735                 btrfs_put_block_group(block_group);
9736         }
9737 }
9738
9739 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9740 {
9741         struct btrfs_block_group_cache *block_group;
9742         struct btrfs_space_info *space_info;
9743         struct btrfs_caching_control *caching_ctl;
9744         struct rb_node *n;
9745
9746         down_write(&info->commit_root_sem);
9747         while (!list_empty(&info->caching_block_groups)) {
9748                 caching_ctl = list_entry(info->caching_block_groups.next,
9749                                          struct btrfs_caching_control, list);
9750                 list_del(&caching_ctl->list);
9751                 put_caching_control(caching_ctl);
9752         }
9753         up_write(&info->commit_root_sem);
9754
9755         spin_lock(&info->unused_bgs_lock);
9756         while (!list_empty(&info->unused_bgs)) {
9757                 block_group = list_first_entry(&info->unused_bgs,
9758                                                struct btrfs_block_group_cache,
9759                                                bg_list);
9760                 list_del_init(&block_group->bg_list);
9761                 btrfs_put_block_group(block_group);
9762         }
9763         spin_unlock(&info->unused_bgs_lock);
9764
9765         spin_lock(&info->block_group_cache_lock);
9766         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9767                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9768                                        cache_node);
9769                 rb_erase(&block_group->cache_node,
9770                          &info->block_group_cache_tree);
9771                 RB_CLEAR_NODE(&block_group->cache_node);
9772                 spin_unlock(&info->block_group_cache_lock);
9773
9774                 down_write(&block_group->space_info->groups_sem);
9775                 list_del(&block_group->list);
9776                 up_write(&block_group->space_info->groups_sem);
9777
9778                 if (block_group->cached == BTRFS_CACHE_STARTED)
9779                         wait_block_group_cache_done(block_group);
9780
9781                 /*
9782                  * We haven't cached this block group, which means we could
9783                  * possibly have excluded extents on this block group.
9784                  */
9785                 if (block_group->cached == BTRFS_CACHE_NO ||
9786                     block_group->cached == BTRFS_CACHE_ERROR)
9787                         free_excluded_extents(info, block_group);
9788
9789                 btrfs_remove_free_space_cache(block_group);
9790                 ASSERT(list_empty(&block_group->dirty_list));
9791                 ASSERT(list_empty(&block_group->io_list));
9792                 ASSERT(list_empty(&block_group->bg_list));
9793                 ASSERT(atomic_read(&block_group->count) == 1);
9794                 btrfs_put_block_group(block_group);
9795
9796                 spin_lock(&info->block_group_cache_lock);
9797         }
9798         spin_unlock(&info->block_group_cache_lock);
9799
9800         /* now that all the block groups are freed, go through and
9801          * free all the space_info structs.  This is only called during
9802          * the final stages of unmount, and so we know nobody is
9803          * using them.  We call synchronize_rcu() once before we start,
9804          * just to be on the safe side.
9805          */
9806         synchronize_rcu();
9807
9808         release_global_block_rsv(info);
9809
9810         while (!list_empty(&info->space_info)) {
9811                 int i;
9812
9813                 space_info = list_entry(info->space_info.next,
9814                                         struct btrfs_space_info,
9815                                         list);
9816
9817                 /*
9818                  * Do not hide this behind enospc_debug, this is actually
9819                  * important and indicates a real bug if this happens.
9820                  */
9821                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9822                             space_info->bytes_reserved > 0 ||
9823                             space_info->bytes_may_use > 0))
9824                         dump_space_info(info, space_info, 0, 0);
9825                 list_del(&space_info->list);
9826                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9827                         struct kobject *kobj;
9828                         kobj = space_info->block_group_kobjs[i];
9829                         space_info->block_group_kobjs[i] = NULL;
9830                         if (kobj) {
9831                                 kobject_del(kobj);
9832                                 kobject_put(kobj);
9833                         }
9834                 }
9835                 kobject_del(&space_info->kobj);
9836                 kobject_put(&space_info->kobj);
9837         }
9838         return 0;
9839 }
9840
9841 static void __link_block_group(struct btrfs_space_info *space_info,
9842                                struct btrfs_block_group_cache *cache)
9843 {
9844         int index = get_block_group_index(cache);
9845         bool first = false;
9846
9847         down_write(&space_info->groups_sem);
9848         if (list_empty(&space_info->block_groups[index]))
9849                 first = true;
9850         list_add_tail(&cache->list, &space_info->block_groups[index]);
9851         up_write(&space_info->groups_sem);
9852
9853         if (first) {
9854                 struct raid_kobject *rkobj;
9855                 int ret;
9856
9857                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9858                 if (!rkobj)
9859                         goto out_err;
9860                 rkobj->raid_type = index;
9861                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9862                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9863                                   "%s", get_raid_name(index));
9864                 if (ret) {
9865                         kobject_put(&rkobj->kobj);
9866                         goto out_err;
9867                 }
9868                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9869         }
9870
9871         return;
9872 out_err:
9873         btrfs_warn(cache->fs_info,
9874                    "failed to add kobject for block cache, ignoring");
9875 }
9876
9877 static struct btrfs_block_group_cache *
9878 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9879                                u64 start, u64 size)
9880 {
9881         struct btrfs_block_group_cache *cache;
9882
9883         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9884         if (!cache)
9885                 return NULL;
9886
9887         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9888                                         GFP_NOFS);
9889         if (!cache->free_space_ctl) {
9890                 kfree(cache);
9891                 return NULL;
9892         }
9893
9894         cache->key.objectid = start;
9895         cache->key.offset = size;
9896         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9897
9898         cache->sectorsize = fs_info->sectorsize;
9899         cache->fs_info = fs_info;
9900         cache->full_stripe_len = btrfs_full_stripe_len(fs_info,
9901                                                        &fs_info->mapping_tree,
9902                                                        start);
9903         set_free_space_tree_thresholds(cache);
9904
9905         atomic_set(&cache->count, 1);
9906         spin_lock_init(&cache->lock);
9907         init_rwsem(&cache->data_rwsem);
9908         INIT_LIST_HEAD(&cache->list);
9909         INIT_LIST_HEAD(&cache->cluster_list);
9910         INIT_LIST_HEAD(&cache->bg_list);
9911         INIT_LIST_HEAD(&cache->ro_list);
9912         INIT_LIST_HEAD(&cache->dirty_list);
9913         INIT_LIST_HEAD(&cache->io_list);
9914         btrfs_init_free_space_ctl(cache);
9915         atomic_set(&cache->trimming, 0);
9916         mutex_init(&cache->free_space_lock);
9917
9918         return cache;
9919 }
9920
9921 int btrfs_read_block_groups(struct btrfs_fs_info *info)
9922 {
9923         struct btrfs_path *path;
9924         int ret;
9925         struct btrfs_block_group_cache *cache;
9926         struct btrfs_space_info *space_info;
9927         struct btrfs_key key;
9928         struct btrfs_key found_key;
9929         struct extent_buffer *leaf;
9930         int need_clear = 0;
9931         u64 cache_gen;
9932         u64 feature;
9933         int mixed;
9934
9935         feature = btrfs_super_incompat_flags(info->super_copy);
9936         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
9937
9938         key.objectid = 0;
9939         key.offset = 0;
9940         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9941         path = btrfs_alloc_path();
9942         if (!path)
9943                 return -ENOMEM;
9944         path->reada = READA_FORWARD;
9945
9946         cache_gen = btrfs_super_cache_generation(info->super_copy);
9947         if (btrfs_test_opt(info, SPACE_CACHE) &&
9948             btrfs_super_generation(info->super_copy) != cache_gen)
9949                 need_clear = 1;
9950         if (btrfs_test_opt(info, CLEAR_CACHE))
9951                 need_clear = 1;
9952
9953         while (1) {
9954                 ret = find_first_block_group(info, path, &key);
9955                 if (ret > 0)
9956                         break;
9957                 if (ret != 0)
9958                         goto error;
9959
9960                 leaf = path->nodes[0];
9961                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9962
9963                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
9964                                                        found_key.offset);
9965                 if (!cache) {
9966                         ret = -ENOMEM;
9967                         goto error;
9968                 }
9969
9970                 if (need_clear) {
9971                         /*
9972                          * When we mount with old space cache, we need to
9973                          * set BTRFS_DC_CLEAR and set dirty flag.
9974                          *
9975                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9976                          *    truncate the old free space cache inode and
9977                          *    setup a new one.
9978                          * b) Setting 'dirty flag' makes sure that we flush
9979                          *    the new space cache info onto disk.
9980                          */
9981                         if (btrfs_test_opt(info, SPACE_CACHE))
9982                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9983                 }
9984
9985                 read_extent_buffer(leaf, &cache->item,
9986                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9987                                    sizeof(cache->item));
9988                 cache->flags = btrfs_block_group_flags(&cache->item);
9989                 if (!mixed &&
9990                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
9991                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
9992                         btrfs_err(info,
9993 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
9994                                   cache->key.objectid);
9995                         ret = -EINVAL;
9996                         goto error;
9997                 }
9998
9999                 key.objectid = found_key.objectid + found_key.offset;
10000                 btrfs_release_path(path);
10001
10002                 /*
10003                  * We need to exclude the super stripes now so that the space
10004                  * info has super bytes accounted for, otherwise we'll think
10005                  * we have more space than we actually do.
10006                  */
10007                 ret = exclude_super_stripes(info, cache);
10008                 if (ret) {
10009                         /*
10010                          * We may have excluded something, so call this just in
10011                          * case.
10012                          */
10013                         free_excluded_extents(info, cache);
10014                         btrfs_put_block_group(cache);
10015                         goto error;
10016                 }
10017
10018                 /*
10019                  * check for two cases, either we are full, and therefore
10020                  * don't need to bother with the caching work since we won't
10021                  * find any space, or we are empty, and we can just add all
10022                  * the space in and be done with it.  This saves us _alot_ of
10023                  * time, particularly in the full case.
10024                  */
10025                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10026                         cache->last_byte_to_unpin = (u64)-1;
10027                         cache->cached = BTRFS_CACHE_FINISHED;
10028                         free_excluded_extents(info, cache);
10029                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10030                         cache->last_byte_to_unpin = (u64)-1;
10031                         cache->cached = BTRFS_CACHE_FINISHED;
10032                         add_new_free_space(cache, info,
10033                                            found_key.objectid,
10034                                            found_key.objectid +
10035                                            found_key.offset);
10036                         free_excluded_extents(info, cache);
10037                 }
10038
10039                 ret = btrfs_add_block_group_cache(info, cache);
10040                 if (ret) {
10041                         btrfs_remove_free_space_cache(cache);
10042                         btrfs_put_block_group(cache);
10043                         goto error;
10044                 }
10045
10046                 trace_btrfs_add_block_group(info, cache, 0);
10047                 ret = update_space_info(info, cache->flags, found_key.offset,
10048                                         btrfs_block_group_used(&cache->item),
10049                                         cache->bytes_super, &space_info);
10050                 if (ret) {
10051                         btrfs_remove_free_space_cache(cache);
10052                         spin_lock(&info->block_group_cache_lock);
10053                         rb_erase(&cache->cache_node,
10054                                  &info->block_group_cache_tree);
10055                         RB_CLEAR_NODE(&cache->cache_node);
10056                         spin_unlock(&info->block_group_cache_lock);
10057                         btrfs_put_block_group(cache);
10058                         goto error;
10059                 }
10060
10061                 cache->space_info = space_info;
10062
10063                 __link_block_group(space_info, cache);
10064
10065                 set_avail_alloc_bits(info, cache->flags);
10066                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10067                         inc_block_group_ro(cache, 1);
10068                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10069                         spin_lock(&info->unused_bgs_lock);
10070                         /* Should always be true but just in case. */
10071                         if (list_empty(&cache->bg_list)) {
10072                                 btrfs_get_block_group(cache);
10073                                 list_add_tail(&cache->bg_list,
10074                                               &info->unused_bgs);
10075                         }
10076                         spin_unlock(&info->unused_bgs_lock);
10077                 }
10078         }
10079
10080         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10081                 if (!(get_alloc_profile(info, space_info->flags) &
10082                       (BTRFS_BLOCK_GROUP_RAID10 |
10083                        BTRFS_BLOCK_GROUP_RAID1 |
10084                        BTRFS_BLOCK_GROUP_RAID5 |
10085                        BTRFS_BLOCK_GROUP_RAID6 |
10086                        BTRFS_BLOCK_GROUP_DUP)))
10087                         continue;
10088                 /*
10089                  * avoid allocating from un-mirrored block group if there are
10090                  * mirrored block groups.
10091                  */
10092                 list_for_each_entry(cache,
10093                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10094                                 list)
10095                         inc_block_group_ro(cache, 1);
10096                 list_for_each_entry(cache,
10097                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10098                                 list)
10099                         inc_block_group_ro(cache, 1);
10100         }
10101
10102         init_global_block_rsv(info);
10103         ret = 0;
10104 error:
10105         btrfs_free_path(path);
10106         return ret;
10107 }
10108
10109 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
10110                                        struct btrfs_fs_info *fs_info)
10111 {
10112         struct btrfs_block_group_cache *block_group, *tmp;
10113         struct btrfs_root *extent_root = fs_info->extent_root;
10114         struct btrfs_block_group_item item;
10115         struct btrfs_key key;
10116         int ret = 0;
10117         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10118
10119         trans->can_flush_pending_bgs = false;
10120         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10121                 if (ret)
10122                         goto next;
10123
10124                 spin_lock(&block_group->lock);
10125                 memcpy(&item, &block_group->item, sizeof(item));
10126                 memcpy(&key, &block_group->key, sizeof(key));
10127                 spin_unlock(&block_group->lock);
10128
10129                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10130                                         sizeof(item));
10131                 if (ret)
10132                         btrfs_abort_transaction(trans, ret);
10133                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10134                                                key.offset);
10135                 if (ret)
10136                         btrfs_abort_transaction(trans, ret);
10137                 add_block_group_free_space(trans, fs_info, block_group);
10138                 /* already aborted the transaction if it failed. */
10139 next:
10140                 list_del_init(&block_group->bg_list);
10141         }
10142         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10143 }
10144
10145 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10146                            struct btrfs_fs_info *fs_info, u64 bytes_used,
10147                            u64 type, u64 chunk_objectid, u64 chunk_offset,
10148                            u64 size)
10149 {
10150         struct btrfs_block_group_cache *cache;
10151         int ret;
10152
10153         btrfs_set_log_full_commit(fs_info, trans);
10154
10155         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10156         if (!cache)
10157                 return -ENOMEM;
10158
10159         btrfs_set_block_group_used(&cache->item, bytes_used);
10160         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
10161         btrfs_set_block_group_flags(&cache->item, type);
10162
10163         cache->flags = type;
10164         cache->last_byte_to_unpin = (u64)-1;
10165         cache->cached = BTRFS_CACHE_FINISHED;
10166         cache->needs_free_space = 1;
10167         ret = exclude_super_stripes(fs_info, cache);
10168         if (ret) {
10169                 /*
10170                  * We may have excluded something, so call this just in
10171                  * case.
10172                  */
10173                 free_excluded_extents(fs_info, cache);
10174                 btrfs_put_block_group(cache);
10175                 return ret;
10176         }
10177
10178         add_new_free_space(cache, fs_info, chunk_offset, chunk_offset + size);
10179
10180         free_excluded_extents(fs_info, cache);
10181
10182 #ifdef CONFIG_BTRFS_DEBUG
10183         if (btrfs_should_fragment_free_space(cache)) {
10184                 u64 new_bytes_used = size - bytes_used;
10185
10186                 bytes_used += new_bytes_used >> 1;
10187                 fragment_free_space(cache);
10188         }
10189 #endif
10190         /*
10191          * Call to ensure the corresponding space_info object is created and
10192          * assigned to our block group, but don't update its counters just yet.
10193          * We want our bg to be added to the rbtree with its ->space_info set.
10194          */
10195         ret = update_space_info(fs_info, cache->flags, 0, 0, 0,
10196                                 &cache->space_info);
10197         if (ret) {
10198                 btrfs_remove_free_space_cache(cache);
10199                 btrfs_put_block_group(cache);
10200                 return ret;
10201         }
10202
10203         ret = btrfs_add_block_group_cache(fs_info, cache);
10204         if (ret) {
10205                 btrfs_remove_free_space_cache(cache);
10206                 btrfs_put_block_group(cache);
10207                 return ret;
10208         }
10209
10210         /*
10211          * Now that our block group has its ->space_info set and is inserted in
10212          * the rbtree, update the space info's counters.
10213          */
10214         trace_btrfs_add_block_group(fs_info, cache, 1);
10215         ret = update_space_info(fs_info, cache->flags, size, bytes_used,
10216                                 cache->bytes_super, &cache->space_info);
10217         if (ret) {
10218                 btrfs_remove_free_space_cache(cache);
10219                 spin_lock(&fs_info->block_group_cache_lock);
10220                 rb_erase(&cache->cache_node,
10221                          &fs_info->block_group_cache_tree);
10222                 RB_CLEAR_NODE(&cache->cache_node);
10223                 spin_unlock(&fs_info->block_group_cache_lock);
10224                 btrfs_put_block_group(cache);
10225                 return ret;
10226         }
10227         update_global_block_rsv(fs_info);
10228
10229         __link_block_group(cache->space_info, cache);
10230
10231         list_add_tail(&cache->bg_list, &trans->new_bgs);
10232
10233         set_avail_alloc_bits(fs_info, type);
10234         return 0;
10235 }
10236
10237 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10238 {
10239         u64 extra_flags = chunk_to_extended(flags) &
10240                                 BTRFS_EXTENDED_PROFILE_MASK;
10241
10242         write_seqlock(&fs_info->profiles_lock);
10243         if (flags & BTRFS_BLOCK_GROUP_DATA)
10244                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10245         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10246                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10247         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10248                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10249         write_sequnlock(&fs_info->profiles_lock);
10250 }
10251
10252 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10253                              struct btrfs_fs_info *fs_info, u64 group_start,
10254                              struct extent_map *em)
10255 {
10256         struct btrfs_root *root = fs_info->extent_root;
10257         struct btrfs_path *path;
10258         struct btrfs_block_group_cache *block_group;
10259         struct btrfs_free_cluster *cluster;
10260         struct btrfs_root *tree_root = fs_info->tree_root;
10261         struct btrfs_key key;
10262         struct inode *inode;
10263         struct kobject *kobj = NULL;
10264         int ret;
10265         int index;
10266         int factor;
10267         struct btrfs_caching_control *caching_ctl = NULL;
10268         bool remove_em;
10269
10270         block_group = btrfs_lookup_block_group(fs_info, group_start);
10271         BUG_ON(!block_group);
10272         BUG_ON(!block_group->ro);
10273
10274         /*
10275          * Free the reserved super bytes from this block group before
10276          * remove it.
10277          */
10278         free_excluded_extents(fs_info, block_group);
10279
10280         memcpy(&key, &block_group->key, sizeof(key));
10281         index = get_block_group_index(block_group);
10282         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10283                                   BTRFS_BLOCK_GROUP_RAID1 |
10284                                   BTRFS_BLOCK_GROUP_RAID10))
10285                 factor = 2;
10286         else
10287                 factor = 1;
10288
10289         /* make sure this block group isn't part of an allocation cluster */
10290         cluster = &fs_info->data_alloc_cluster;
10291         spin_lock(&cluster->refill_lock);
10292         btrfs_return_cluster_to_free_space(block_group, cluster);
10293         spin_unlock(&cluster->refill_lock);
10294
10295         /*
10296          * make sure this block group isn't part of a metadata
10297          * allocation cluster
10298          */
10299         cluster = &fs_info->meta_alloc_cluster;
10300         spin_lock(&cluster->refill_lock);
10301         btrfs_return_cluster_to_free_space(block_group, cluster);
10302         spin_unlock(&cluster->refill_lock);
10303
10304         path = btrfs_alloc_path();
10305         if (!path) {
10306                 ret = -ENOMEM;
10307                 goto out;
10308         }
10309
10310         /*
10311          * get the inode first so any iput calls done for the io_list
10312          * aren't the final iput (no unlinks allowed now)
10313          */
10314         inode = lookup_free_space_inode(fs_info, block_group, path);
10315
10316         mutex_lock(&trans->transaction->cache_write_mutex);
10317         /*
10318          * make sure our free spache cache IO is done before remove the
10319          * free space inode
10320          */
10321         spin_lock(&trans->transaction->dirty_bgs_lock);
10322         if (!list_empty(&block_group->io_list)) {
10323                 list_del_init(&block_group->io_list);
10324
10325                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10326
10327                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10328                 btrfs_wait_cache_io(trans, block_group, path);
10329                 btrfs_put_block_group(block_group);
10330                 spin_lock(&trans->transaction->dirty_bgs_lock);
10331         }
10332
10333         if (!list_empty(&block_group->dirty_list)) {
10334                 list_del_init(&block_group->dirty_list);
10335                 btrfs_put_block_group(block_group);
10336         }
10337         spin_unlock(&trans->transaction->dirty_bgs_lock);
10338         mutex_unlock(&trans->transaction->cache_write_mutex);
10339
10340         if (!IS_ERR(inode)) {
10341                 ret = btrfs_orphan_add(trans, inode);
10342                 if (ret) {
10343                         btrfs_add_delayed_iput(inode);
10344                         goto out;
10345                 }
10346                 clear_nlink(inode);
10347                 /* One for the block groups ref */
10348                 spin_lock(&block_group->lock);
10349                 if (block_group->iref) {
10350                         block_group->iref = 0;
10351                         block_group->inode = NULL;
10352                         spin_unlock(&block_group->lock);
10353                         iput(inode);
10354                 } else {
10355                         spin_unlock(&block_group->lock);
10356                 }
10357                 /* One for our lookup ref */
10358                 btrfs_add_delayed_iput(inode);
10359         }
10360
10361         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10362         key.offset = block_group->key.objectid;
10363         key.type = 0;
10364
10365         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10366         if (ret < 0)
10367                 goto out;
10368         if (ret > 0)
10369                 btrfs_release_path(path);
10370         if (ret == 0) {
10371                 ret = btrfs_del_item(trans, tree_root, path);
10372                 if (ret)
10373                         goto out;
10374                 btrfs_release_path(path);
10375         }
10376
10377         spin_lock(&fs_info->block_group_cache_lock);
10378         rb_erase(&block_group->cache_node,
10379                  &fs_info->block_group_cache_tree);
10380         RB_CLEAR_NODE(&block_group->cache_node);
10381
10382         if (fs_info->first_logical_byte == block_group->key.objectid)
10383                 fs_info->first_logical_byte = (u64)-1;
10384         spin_unlock(&fs_info->block_group_cache_lock);
10385
10386         down_write(&block_group->space_info->groups_sem);
10387         /*
10388          * we must use list_del_init so people can check to see if they
10389          * are still on the list after taking the semaphore
10390          */
10391         list_del_init(&block_group->list);
10392         if (list_empty(&block_group->space_info->block_groups[index])) {
10393                 kobj = block_group->space_info->block_group_kobjs[index];
10394                 block_group->space_info->block_group_kobjs[index] = NULL;
10395                 clear_avail_alloc_bits(fs_info, block_group->flags);
10396         }
10397         up_write(&block_group->space_info->groups_sem);
10398         if (kobj) {
10399                 kobject_del(kobj);
10400                 kobject_put(kobj);
10401         }
10402
10403         if (block_group->has_caching_ctl)
10404                 caching_ctl = get_caching_control(block_group);
10405         if (block_group->cached == BTRFS_CACHE_STARTED)
10406                 wait_block_group_cache_done(block_group);
10407         if (block_group->has_caching_ctl) {
10408                 down_write(&fs_info->commit_root_sem);
10409                 if (!caching_ctl) {
10410                         struct btrfs_caching_control *ctl;
10411
10412                         list_for_each_entry(ctl,
10413                                     &fs_info->caching_block_groups, list)
10414                                 if (ctl->block_group == block_group) {
10415                                         caching_ctl = ctl;
10416                                         atomic_inc(&caching_ctl->count);
10417                                         break;
10418                                 }
10419                 }
10420                 if (caching_ctl)
10421                         list_del_init(&caching_ctl->list);
10422                 up_write(&fs_info->commit_root_sem);
10423                 if (caching_ctl) {
10424                         /* Once for the caching bgs list and once for us. */
10425                         put_caching_control(caching_ctl);
10426                         put_caching_control(caching_ctl);
10427                 }
10428         }
10429
10430         spin_lock(&trans->transaction->dirty_bgs_lock);
10431         if (!list_empty(&block_group->dirty_list)) {
10432                 WARN_ON(1);
10433         }
10434         if (!list_empty(&block_group->io_list)) {
10435                 WARN_ON(1);
10436         }
10437         spin_unlock(&trans->transaction->dirty_bgs_lock);
10438         btrfs_remove_free_space_cache(block_group);
10439
10440         spin_lock(&block_group->space_info->lock);
10441         list_del_init(&block_group->ro_list);
10442
10443         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10444                 WARN_ON(block_group->space_info->total_bytes
10445                         < block_group->key.offset);
10446                 WARN_ON(block_group->space_info->bytes_readonly
10447                         < block_group->key.offset);
10448                 WARN_ON(block_group->space_info->disk_total
10449                         < block_group->key.offset * factor);
10450         }
10451         block_group->space_info->total_bytes -= block_group->key.offset;
10452         block_group->space_info->bytes_readonly -= block_group->key.offset;
10453         block_group->space_info->disk_total -= block_group->key.offset * factor;
10454
10455         spin_unlock(&block_group->space_info->lock);
10456
10457         memcpy(&key, &block_group->key, sizeof(key));
10458
10459         mutex_lock(&fs_info->chunk_mutex);
10460         if (!list_empty(&em->list)) {
10461                 /* We're in the transaction->pending_chunks list. */
10462                 free_extent_map(em);
10463         }
10464         spin_lock(&block_group->lock);
10465         block_group->removed = 1;
10466         /*
10467          * At this point trimming can't start on this block group, because we
10468          * removed the block group from the tree fs_info->block_group_cache_tree
10469          * so no one can't find it anymore and even if someone already got this
10470          * block group before we removed it from the rbtree, they have already
10471          * incremented block_group->trimming - if they didn't, they won't find
10472          * any free space entries because we already removed them all when we
10473          * called btrfs_remove_free_space_cache().
10474          *
10475          * And we must not remove the extent map from the fs_info->mapping_tree
10476          * to prevent the same logical address range and physical device space
10477          * ranges from being reused for a new block group. This is because our
10478          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10479          * completely transactionless, so while it is trimming a range the
10480          * currently running transaction might finish and a new one start,
10481          * allowing for new block groups to be created that can reuse the same
10482          * physical device locations unless we take this special care.
10483          *
10484          * There may also be an implicit trim operation if the file system
10485          * is mounted with -odiscard. The same protections must remain
10486          * in place until the extents have been discarded completely when
10487          * the transaction commit has completed.
10488          */
10489         remove_em = (atomic_read(&block_group->trimming) == 0);
10490         /*
10491          * Make sure a trimmer task always sees the em in the pinned_chunks list
10492          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10493          * before checking block_group->removed).
10494          */
10495         if (!remove_em) {
10496                 /*
10497                  * Our em might be in trans->transaction->pending_chunks which
10498                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10499                  * and so is the fs_info->pinned_chunks list.
10500                  *
10501                  * So at this point we must be holding the chunk_mutex to avoid
10502                  * any races with chunk allocation (more specifically at
10503                  * volumes.c:contains_pending_extent()), to ensure it always
10504                  * sees the em, either in the pending_chunks list or in the
10505                  * pinned_chunks list.
10506                  */
10507                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10508         }
10509         spin_unlock(&block_group->lock);
10510
10511         if (remove_em) {
10512                 struct extent_map_tree *em_tree;
10513
10514                 em_tree = &fs_info->mapping_tree.map_tree;
10515                 write_lock(&em_tree->lock);
10516                 /*
10517                  * The em might be in the pending_chunks list, so make sure the
10518                  * chunk mutex is locked, since remove_extent_mapping() will
10519                  * delete us from that list.
10520                  */
10521                 remove_extent_mapping(em_tree, em);
10522                 write_unlock(&em_tree->lock);
10523                 /* once for the tree */
10524                 free_extent_map(em);
10525         }
10526
10527         mutex_unlock(&fs_info->chunk_mutex);
10528
10529         ret = remove_block_group_free_space(trans, fs_info, block_group);
10530         if (ret)
10531                 goto out;
10532
10533         btrfs_put_block_group(block_group);
10534         btrfs_put_block_group(block_group);
10535
10536         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10537         if (ret > 0)
10538                 ret = -EIO;
10539         if (ret < 0)
10540                 goto out;
10541
10542         ret = btrfs_del_item(trans, root, path);
10543 out:
10544         btrfs_free_path(path);
10545         return ret;
10546 }
10547
10548 struct btrfs_trans_handle *
10549 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10550                                      const u64 chunk_offset)
10551 {
10552         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10553         struct extent_map *em;
10554         struct map_lookup *map;
10555         unsigned int num_items;
10556
10557         read_lock(&em_tree->lock);
10558         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10559         read_unlock(&em_tree->lock);
10560         ASSERT(em && em->start == chunk_offset);
10561
10562         /*
10563          * We need to reserve 3 + N units from the metadata space info in order
10564          * to remove a block group (done at btrfs_remove_chunk() and at
10565          * btrfs_remove_block_group()), which are used for:
10566          *
10567          * 1 unit for adding the free space inode's orphan (located in the tree
10568          * of tree roots).
10569          * 1 unit for deleting the block group item (located in the extent
10570          * tree).
10571          * 1 unit for deleting the free space item (located in tree of tree
10572          * roots).
10573          * N units for deleting N device extent items corresponding to each
10574          * stripe (located in the device tree).
10575          *
10576          * In order to remove a block group we also need to reserve units in the
10577          * system space info in order to update the chunk tree (update one or
10578          * more device items and remove one chunk item), but this is done at
10579          * btrfs_remove_chunk() through a call to check_system_chunk().
10580          */
10581         map = em->map_lookup;
10582         num_items = 3 + map->num_stripes;
10583         free_extent_map(em);
10584
10585         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10586                                                            num_items, 1);
10587 }
10588
10589 /*
10590  * Process the unused_bgs list and remove any that don't have any allocated
10591  * space inside of them.
10592  */
10593 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10594 {
10595         struct btrfs_block_group_cache *block_group;
10596         struct btrfs_space_info *space_info;
10597         struct btrfs_trans_handle *trans;
10598         int ret = 0;
10599
10600         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10601                 return;
10602
10603         spin_lock(&fs_info->unused_bgs_lock);
10604         while (!list_empty(&fs_info->unused_bgs)) {
10605                 u64 start, end;
10606                 int trimming;
10607
10608                 block_group = list_first_entry(&fs_info->unused_bgs,
10609                                                struct btrfs_block_group_cache,
10610                                                bg_list);
10611                 list_del_init(&block_group->bg_list);
10612
10613                 space_info = block_group->space_info;
10614
10615                 if (ret || btrfs_mixed_space_info(space_info)) {
10616                         btrfs_put_block_group(block_group);
10617                         continue;
10618                 }
10619                 spin_unlock(&fs_info->unused_bgs_lock);
10620
10621                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10622
10623                 /* Don't want to race with allocators so take the groups_sem */
10624                 down_write(&space_info->groups_sem);
10625                 spin_lock(&block_group->lock);
10626                 if (block_group->reserved ||
10627                     btrfs_block_group_used(&block_group->item) ||
10628                     block_group->ro ||
10629                     list_is_singular(&block_group->list)) {
10630                         /*
10631                          * We want to bail if we made new allocations or have
10632                          * outstanding allocations in this block group.  We do
10633                          * the ro check in case balance is currently acting on
10634                          * this block group.
10635                          */
10636                         spin_unlock(&block_group->lock);
10637                         up_write(&space_info->groups_sem);
10638                         goto next;
10639                 }
10640                 spin_unlock(&block_group->lock);
10641
10642                 /* We don't want to force the issue, only flip if it's ok. */
10643                 ret = inc_block_group_ro(block_group, 0);
10644                 up_write(&space_info->groups_sem);
10645                 if (ret < 0) {
10646                         ret = 0;
10647                         goto next;
10648                 }
10649
10650                 /*
10651                  * Want to do this before we do anything else so we can recover
10652                  * properly if we fail to join the transaction.
10653                  */
10654                 trans = btrfs_start_trans_remove_block_group(fs_info,
10655                                                      block_group->key.objectid);
10656                 if (IS_ERR(trans)) {
10657                         btrfs_dec_block_group_ro(block_group);
10658                         ret = PTR_ERR(trans);
10659                         goto next;
10660                 }
10661
10662                 /*
10663                  * We could have pending pinned extents for this block group,
10664                  * just delete them, we don't care about them anymore.
10665                  */
10666                 start = block_group->key.objectid;
10667                 end = start + block_group->key.offset - 1;
10668                 /*
10669                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10670                  * btrfs_finish_extent_commit(). If we are at transaction N,
10671                  * another task might be running finish_extent_commit() for the
10672                  * previous transaction N - 1, and have seen a range belonging
10673                  * to the block group in freed_extents[] before we were able to
10674                  * clear the whole block group range from freed_extents[]. This
10675                  * means that task can lookup for the block group after we
10676                  * unpinned it from freed_extents[] and removed it, leading to
10677                  * a BUG_ON() at btrfs_unpin_extent_range().
10678                  */
10679                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10680                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10681                                   EXTENT_DIRTY);
10682                 if (ret) {
10683                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10684                         btrfs_dec_block_group_ro(block_group);
10685                         goto end_trans;
10686                 }
10687                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10688                                   EXTENT_DIRTY);
10689                 if (ret) {
10690                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10691                         btrfs_dec_block_group_ro(block_group);
10692                         goto end_trans;
10693                 }
10694                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10695
10696                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10697                 spin_lock(&space_info->lock);
10698                 spin_lock(&block_group->lock);
10699
10700                 space_info->bytes_pinned -= block_group->pinned;
10701                 space_info->bytes_readonly += block_group->pinned;
10702                 percpu_counter_add(&space_info->total_bytes_pinned,
10703                                    -block_group->pinned);
10704                 block_group->pinned = 0;
10705
10706                 spin_unlock(&block_group->lock);
10707                 spin_unlock(&space_info->lock);
10708
10709                 /* DISCARD can flip during remount */
10710                 trimming = btrfs_test_opt(fs_info, DISCARD);
10711
10712                 /* Implicit trim during transaction commit. */
10713                 if (trimming)
10714                         btrfs_get_block_group_trimming(block_group);
10715
10716                 /*
10717                  * Btrfs_remove_chunk will abort the transaction if things go
10718                  * horribly wrong.
10719                  */
10720                 ret = btrfs_remove_chunk(trans, fs_info,
10721                                          block_group->key.objectid);
10722
10723                 if (ret) {
10724                         if (trimming)
10725                                 btrfs_put_block_group_trimming(block_group);
10726                         goto end_trans;
10727                 }
10728
10729                 /*
10730                  * If we're not mounted with -odiscard, we can just forget
10731                  * about this block group. Otherwise we'll need to wait
10732                  * until transaction commit to do the actual discard.
10733                  */
10734                 if (trimming) {
10735                         spin_lock(&fs_info->unused_bgs_lock);
10736                         /*
10737                          * A concurrent scrub might have added us to the list
10738                          * fs_info->unused_bgs, so use a list_move operation
10739                          * to add the block group to the deleted_bgs list.
10740                          */
10741                         list_move(&block_group->bg_list,
10742                                   &trans->transaction->deleted_bgs);
10743                         spin_unlock(&fs_info->unused_bgs_lock);
10744                         btrfs_get_block_group(block_group);
10745                 }
10746 end_trans:
10747                 btrfs_end_transaction(trans);
10748 next:
10749                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10750                 btrfs_put_block_group(block_group);
10751                 spin_lock(&fs_info->unused_bgs_lock);
10752         }
10753         spin_unlock(&fs_info->unused_bgs_lock);
10754 }
10755
10756 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10757 {
10758         struct btrfs_space_info *space_info;
10759         struct btrfs_super_block *disk_super;
10760         u64 features;
10761         u64 flags;
10762         int mixed = 0;
10763         int ret;
10764
10765         disk_super = fs_info->super_copy;
10766         if (!btrfs_super_root(disk_super))
10767                 return -EINVAL;
10768
10769         features = btrfs_super_incompat_flags(disk_super);
10770         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10771                 mixed = 1;
10772
10773         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10774         ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10775         if (ret)
10776                 goto out;
10777
10778         if (mixed) {
10779                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10780                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10781         } else {
10782                 flags = BTRFS_BLOCK_GROUP_METADATA;
10783                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10784                 if (ret)
10785                         goto out;
10786
10787                 flags = BTRFS_BLOCK_GROUP_DATA;
10788                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10789         }
10790 out:
10791         return ret;
10792 }
10793
10794 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10795                                    u64 start, u64 end)
10796 {
10797         return unpin_extent_range(fs_info, start, end, false);
10798 }
10799
10800 /*
10801  * It used to be that old block groups would be left around forever.
10802  * Iterating over them would be enough to trim unused space.  Since we
10803  * now automatically remove them, we also need to iterate over unallocated
10804  * space.
10805  *
10806  * We don't want a transaction for this since the discard may take a
10807  * substantial amount of time.  We don't require that a transaction be
10808  * running, but we do need to take a running transaction into account
10809  * to ensure that we're not discarding chunks that were released in
10810  * the current transaction.
10811  *
10812  * Holding the chunks lock will prevent other threads from allocating
10813  * or releasing chunks, but it won't prevent a running transaction
10814  * from committing and releasing the memory that the pending chunks
10815  * list head uses.  For that, we need to take a reference to the
10816  * transaction.
10817  */
10818 static int btrfs_trim_free_extents(struct btrfs_device *device,
10819                                    u64 minlen, u64 *trimmed)
10820 {
10821         u64 start = 0, len = 0;
10822         int ret;
10823
10824         *trimmed = 0;
10825
10826         /* Not writeable = nothing to do. */
10827         if (!device->writeable)
10828                 return 0;
10829
10830         /* No free space = nothing to do. */
10831         if (device->total_bytes <= device->bytes_used)
10832                 return 0;
10833
10834         ret = 0;
10835
10836         while (1) {
10837                 struct btrfs_fs_info *fs_info = device->fs_info;
10838                 struct btrfs_transaction *trans;
10839                 u64 bytes;
10840
10841                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10842                 if (ret)
10843                         return ret;
10844
10845                 down_read(&fs_info->commit_root_sem);
10846
10847                 spin_lock(&fs_info->trans_lock);
10848                 trans = fs_info->running_transaction;
10849                 if (trans)
10850                         atomic_inc(&trans->use_count);
10851                 spin_unlock(&fs_info->trans_lock);
10852
10853                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10854                                                  &start, &len);
10855                 if (trans)
10856                         btrfs_put_transaction(trans);
10857
10858                 if (ret) {
10859                         up_read(&fs_info->commit_root_sem);
10860                         mutex_unlock(&fs_info->chunk_mutex);
10861                         if (ret == -ENOSPC)
10862                                 ret = 0;
10863                         break;
10864                 }
10865
10866                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10867                 up_read(&fs_info->commit_root_sem);
10868                 mutex_unlock(&fs_info->chunk_mutex);
10869
10870                 if (ret)
10871                         break;
10872
10873                 start += len;
10874                 *trimmed += bytes;
10875
10876                 if (fatal_signal_pending(current)) {
10877                         ret = -ERESTARTSYS;
10878                         break;
10879                 }
10880
10881                 cond_resched();
10882         }
10883
10884         return ret;
10885 }
10886
10887 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10888 {
10889         struct btrfs_block_group_cache *cache = NULL;
10890         struct btrfs_device *device;
10891         struct list_head *devices;
10892         u64 group_trimmed;
10893         u64 start;
10894         u64 end;
10895         u64 trimmed = 0;
10896         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10897         int ret = 0;
10898
10899         /*
10900          * try to trim all FS space, our block group may start from non-zero.
10901          */
10902         if (range->len == total_bytes)
10903                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10904         else
10905                 cache = btrfs_lookup_block_group(fs_info, range->start);
10906
10907         while (cache) {
10908                 if (cache->key.objectid >= (range->start + range->len)) {
10909                         btrfs_put_block_group(cache);
10910                         break;
10911                 }
10912
10913                 start = max(range->start, cache->key.objectid);
10914                 end = min(range->start + range->len,
10915                                 cache->key.objectid + cache->key.offset);
10916
10917                 if (end - start >= range->minlen) {
10918                         if (!block_group_cache_done(cache)) {
10919                                 ret = cache_block_group(cache, 0);
10920                                 if (ret) {
10921                                         btrfs_put_block_group(cache);
10922                                         break;
10923                                 }
10924                                 ret = wait_block_group_cache_done(cache);
10925                                 if (ret) {
10926                                         btrfs_put_block_group(cache);
10927                                         break;
10928                                 }
10929                         }
10930                         ret = btrfs_trim_block_group(cache,
10931                                                      &group_trimmed,
10932                                                      start,
10933                                                      end,
10934                                                      range->minlen);
10935
10936                         trimmed += group_trimmed;
10937                         if (ret) {
10938                                 btrfs_put_block_group(cache);
10939                                 break;
10940                         }
10941                 }
10942
10943                 cache = next_block_group(fs_info, cache);
10944         }
10945
10946         mutex_lock(&fs_info->fs_devices->device_list_mutex);
10947         devices = &fs_info->fs_devices->alloc_list;
10948         list_for_each_entry(device, devices, dev_alloc_list) {
10949                 ret = btrfs_trim_free_extents(device, range->minlen,
10950                                               &group_trimmed);
10951                 if (ret)
10952                         break;
10953
10954                 trimmed += group_trimmed;
10955         }
10956         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
10957
10958         range->len = trimmed;
10959         return ret;
10960 }
10961
10962 /*
10963  * btrfs_{start,end}_write_no_snapshoting() are similar to
10964  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10965  * data into the page cache through nocow before the subvolume is snapshoted,
10966  * but flush the data into disk after the snapshot creation, or to prevent
10967  * operations while snapshoting is ongoing and that cause the snapshot to be
10968  * inconsistent (writes followed by expanding truncates for example).
10969  */
10970 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10971 {
10972         percpu_counter_dec(&root->subv_writers->counter);
10973         /*
10974          * Make sure counter is updated before we wake up waiters.
10975          */
10976         smp_mb();
10977         if (waitqueue_active(&root->subv_writers->wait))
10978                 wake_up(&root->subv_writers->wait);
10979 }
10980
10981 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10982 {
10983         if (atomic_read(&root->will_be_snapshoted))
10984                 return 0;
10985
10986         percpu_counter_inc(&root->subv_writers->counter);
10987         /*
10988          * Make sure counter is updated before we check for snapshot creation.
10989          */
10990         smp_mb();
10991         if (atomic_read(&root->will_be_snapshoted)) {
10992                 btrfs_end_write_no_snapshoting(root);
10993                 return 0;
10994         }
10995         return 1;
10996 }
10997
10998 static int wait_snapshoting_atomic_t(atomic_t *a)
10999 {
11000         schedule();
11001         return 0;
11002 }
11003
11004 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11005 {
11006         while (true) {
11007                 int ret;
11008
11009                 ret = btrfs_start_write_no_snapshoting(root);
11010                 if (ret)
11011                         break;
11012                 wait_on_atomic_t(&root->will_be_snapshoted,
11013                                  wait_snapshoting_atomic_t,
11014                                  TASK_UNINTERRUPTIBLE);
11015         }
11016 }