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/sched/signal.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/sort.h>
  24 #include <linux/rcupdate.h>
  25 #include <linux/kthread.h>
  26 #include <linux/slab.h>
  27 #include <linux/ratelimit.h>
  28 #include <linux/percpu_counter.h>
  29 #include <linux/lockdep.h>
  30 #include <linux/crc32c.h>
  31 #include "tree-log.h"
  32 #include "disk-io.h"
  33 #include "print-tree.h"
  34 #include "volumes.h"
  35 #include "raid56.h"
  36 #include "locking.h"
  37 #include "free-space-cache.h"
  38 #include "free-space-tree.h"
  39 #include "math.h"
  40 #include "sysfs.h"
  41 #include "qgroup.h"
  42 #include "ref-verify.h"
  43
  44 #undef SCRAMBLE_DELAYED_REFS
  45
  46 /*
  47  * control flags for do_chunk_alloc's force field
  48  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  49  * if we really need one.
  50  *
  51  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  52  * if we have very few chunks already allocated.  This is
  53  * used as part of the clustering code to help make sure
  54  * we have a good pool of storage to cluster in, without
  55  * filling the FS with empty chunks
  56  *
  57  * CHUNK_ALLOC_FORCE means it must try to allocate one
  58  *
  59  */
  60 enum {
  61         CHUNK_ALLOC_NO_FORCE = 0,
  62         CHUNK_ALLOC_LIMITED = 1,
  63         CHUNK_ALLOC_FORCE = 2,
  64 };
  65
  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 block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  94                                u64 num_bytes);
  95 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
  96                                      struct btrfs_space_info *space_info,
  97                                      u64 num_bytes);
  98 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
  99                                      struct btrfs_space_info *space_info,
 100                                      u64 num_bytes);
 101
 102 static noinline int
 103 block_group_cache_done(struct btrfs_block_group_cache *cache)
 104 {
 105         smp_mb();
 106         return cache->cached == BTRFS_CACHE_FINISHED ||
 107                 cache->cached == BTRFS_CACHE_ERROR;
 108 }
 109
 110 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 111 {
 112         return (cache->flags & bits) == bits;
 113 }
 114
 115 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 116 {
 117         atomic_inc(&cache->count);
 118 }
 119
 120 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 121 {
 122         if (atomic_dec_and_test(&cache->count)) {
 123                 WARN_ON(cache->pinned > 0);
 124                 WARN_ON(cache->reserved > 0);
 125
 126                 /*
 127                  * If not empty, someone is still holding mutex of
 128                  * full_stripe_lock, which can only be released by caller.
 129                  * And it will definitely cause use-after-free when caller
 130                  * tries to release full stripe lock.
 131                  *
 132                  * No better way to resolve, but only to warn.
 133                  */
 134                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 135                 kfree(cache->free_space_ctl);
 136                 kfree(cache);
 137         }
 138 }
 139
 140 /*
 141  * this adds the block group to the fs_info rb tree for the block group
 142  * cache
 143  */
 144 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 145                                 struct btrfs_block_group_cache *block_group)
 146 {
 147         struct rb_node **p;
 148         struct rb_node *parent = NULL;
 149         struct btrfs_block_group_cache *cache;
 150
 151         spin_lock(&info->block_group_cache_lock);
 152         p = &info->block_group_cache_tree.rb_node;
 153
 154         while (*p) {
 155                 parent = *p;
 156                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 157                                  cache_node);
 158                 if (block_group->key.objectid < cache->key.objectid) {
 159                         p = &(*p)->rb_left;
 160                 } else if (block_group->key.objectid > cache->key.objectid) {
 161                         p = &(*p)->rb_right;
 162                 } else {
 163                         spin_unlock(&info->block_group_cache_lock);
 164                         return -EEXIST;
 165                 }
 166         }
 167
 168         rb_link_node(&block_group->cache_node, parent, p);
 169         rb_insert_color(&block_group->cache_node,
 170                         &info->block_group_cache_tree);
 171
 172         if (info->first_logical_byte > block_group->key.objectid)
 173                 info->first_logical_byte = block_group->key.objectid;
 174
 175         spin_unlock(&info->block_group_cache_lock);
 176
 177         return 0;
 178 }
 179
 180 /*
 181  * This will return the block group at or after bytenr if contains is 0, else
 182  * it will return the block group that contains the bytenr
 183  */
 184 static struct btrfs_block_group_cache *
 185 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 186                               int contains)
 187 {
 188         struct btrfs_block_group_cache *cache, *ret = NULL;
 189         struct rb_node *n;
 190         u64 end, start;
 191
 192         spin_lock(&info->block_group_cache_lock);
 193         n = info->block_group_cache_tree.rb_node;
 194
 195         while (n) {
 196                 cache = rb_entry(n, struct btrfs_block_group_cache,
 197                                  cache_node);
 198                 end = cache->key.objectid + cache->key.offset - 1;
 199                 start = cache->key.objectid;
 200
 201                 if (bytenr < start) {
 202                         if (!contains && (!ret || start < ret->key.objectid))
 203                                 ret = cache;
 204                         n = n->rb_left;
 205                 } else if (bytenr > start) {
 206                         if (contains && bytenr <= end) {
 207                                 ret = cache;
 208                                 break;
 209                         }
 210                         n = n->rb_right;
 211                 } else {
 212                         ret = cache;
 213                         break;
 214                 }
 215         }
 216         if (ret) {
 217                 btrfs_get_block_group(ret);
 218                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 219                         info->first_logical_byte = ret->key.objectid;
 220         }
 221         spin_unlock(&info->block_group_cache_lock);
 222
 223         return ret;
 224 }
 225
 226 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
 227                                u64 start, u64 num_bytes)
 228 {
 229         u64 end = start + num_bytes - 1;
 230         set_extent_bits(&fs_info->freed_extents[0],
 231                         start, end, EXTENT_UPTODATE);
 232         set_extent_bits(&fs_info->freed_extents[1],
 233                         start, end, EXTENT_UPTODATE);
 234         return 0;
 235 }
 236
 237 static void free_excluded_extents(struct btrfs_fs_info *fs_info,
 238                                   struct btrfs_block_group_cache *cache)
 239 {
 240         u64 start, end;
 241
 242         start = cache->key.objectid;
 243         end = start + cache->key.offset - 1;
 244
 245         clear_extent_bits(&fs_info->freed_extents[0],
 246                           start, end, EXTENT_UPTODATE);
 247         clear_extent_bits(&fs_info->freed_extents[1],
 248                           start, end, EXTENT_UPTODATE);
 249 }
 250
 251 static int exclude_super_stripes(struct btrfs_fs_info *fs_info,
 252                                  struct btrfs_block_group_cache *cache)
 253 {
 254         u64 bytenr;
 255         u64 *logical;
 256         int stripe_len;
 257         int i, nr, ret;
 258
 259         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 260                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 261                 cache->bytes_super += stripe_len;
 262                 ret = add_excluded_extent(fs_info, cache->key.objectid,
 263                                           stripe_len);
 264                 if (ret)
 265                         return ret;
 266         }
 267
 268         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 269                 bytenr = btrfs_sb_offset(i);
 270                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
 271                                        bytenr, 0, &logical, &nr, &stripe_len);
 272                 if (ret)
 273                         return ret;
 274
 275                 while (nr--) {
 276                         u64 start, len;
 277
 278                         if (logical[nr] > cache->key.objectid +
 279                             cache->key.offset)
 280                                 continue;
 281
 282                         if (logical[nr] + stripe_len <= cache->key.objectid)
 283                                 continue;
 284
 285                         start = logical[nr];
 286                         if (start < cache->key.objectid) {
 287                                 start = cache->key.objectid;
 288                                 len = (logical[nr] + stripe_len) - start;
 289                         } else {
 290                                 len = min_t(u64, stripe_len,
 291                                             cache->key.objectid +
 292                                             cache->key.offset - start);
 293                         }
 294
 295                         cache->bytes_super += len;
 296                         ret = add_excluded_extent(fs_info, start, len);
 297                         if (ret) {
 298                                 kfree(logical);
 299                                 return ret;
 300                         }
 301                 }
 302
 303                 kfree(logical);
 304         }
 305         return 0;
 306 }
 307
 308 static struct btrfs_caching_control *
 309 get_caching_control(struct btrfs_block_group_cache *cache)
 310 {
 311         struct btrfs_caching_control *ctl;
 312
 313         spin_lock(&cache->lock);
 314         if (!cache->caching_ctl) {
 315                 spin_unlock(&cache->lock);
 316                 return NULL;
 317         }
 318
 319         ctl = cache->caching_ctl;
 320         refcount_inc(&ctl->count);
 321         spin_unlock(&cache->lock);
 322         return ctl;
 323 }
 324
 325 static void put_caching_control(struct btrfs_caching_control *ctl)
 326 {
 327         if (refcount_dec_and_test(&ctl->count))
 328                 kfree(ctl);
 329 }
 330
 331 #ifdef CONFIG_BTRFS_DEBUG
 332 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
 333 {
 334         struct btrfs_fs_info *fs_info = block_group->fs_info;
 335         u64 start = block_group->key.objectid;
 336         u64 len = block_group->key.offset;
 337         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 338                 fs_info->nodesize : fs_info->sectorsize;
 339         u64 step = chunk << 1;
 340
 341         while (len > chunk) {
 342                 btrfs_remove_free_space(block_group, start, chunk);
 343                 start += step;
 344                 if (len < step)
 345                         len = 0;
 346                 else
 347                         len -= step;
 348         }
 349 }
 350 #endif
 351
 352 /*
 353  * this is only called by cache_block_group, since we could have freed extents
 354  * we need to check the pinned_extents for any extents that can't be used yet
 355  * since their free space will be released as soon as the transaction commits.
 356  */
 357 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 358                        struct btrfs_fs_info *info, u64 start, u64 end)
 359 {
 360         u64 extent_start, extent_end, size, total_added = 0;
 361         int ret;
 362
 363         while (start < end) {
 364                 ret = find_first_extent_bit(info->pinned_extents, start,
 365                                             &extent_start, &extent_end,
 366                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 367                                             NULL);
 368                 if (ret)
 369                         break;
 370
 371                 if (extent_start <= start) {
 372                         start = extent_end + 1;
 373                 } else if (extent_start > start && extent_start < end) {
 374                         size = extent_start - start;
 375                         total_added += size;
 376                         ret = btrfs_add_free_space(block_group, start,
 377                                                    size);
 378                         BUG_ON(ret); /* -ENOMEM or logic error */
 379                         start = extent_end + 1;
 380                 } else {
 381                         break;
 382                 }
 383         }
 384
 385         if (start < end) {
 386                 size = end - start;
 387                 total_added += size;
 388                 ret = btrfs_add_free_space(block_group, start, size);
 389                 BUG_ON(ret); /* -ENOMEM or logic error */
 390         }
 391
 392         return total_added;
 393 }
 394
 395 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 396 {
 397         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
 398         struct btrfs_fs_info *fs_info = block_group->fs_info;
 399         struct btrfs_root *extent_root = fs_info->extent_root;
 400         struct btrfs_path *path;
 401         struct extent_buffer *leaf;
 402         struct btrfs_key key;
 403         u64 total_found = 0;
 404         u64 last = 0;
 405         u32 nritems;
 406         int ret;
 407         bool wakeup = true;
 408
 409         path = btrfs_alloc_path();
 410         if (!path)
 411                 return -ENOMEM;
 412
 413         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 414
 415 #ifdef CONFIG_BTRFS_DEBUG
 416         /*
 417          * If we're fragmenting we don't want to make anybody think we can
 418          * allocate from this block group until we've had a chance to fragment
 419          * the free space.
 420          */
 421         if (btrfs_should_fragment_free_space(block_group))
 422                 wakeup = false;
 423 #endif
 424         /*
 425          * We don't want to deadlock with somebody trying to allocate a new
 426          * extent for the extent root while also trying to search the extent
 427          * root to add free space.  So we skip locking and search the commit
 428          * root, since its read-only
 429          */
 430         path->skip_locking = 1;
 431         path->search_commit_root = 1;
 432         path->reada = READA_FORWARD;
 433
 434         key.objectid = last;
 435         key.offset = 0;
 436         key.type = BTRFS_EXTENT_ITEM_KEY;
 437
 438 next:
 439         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 440         if (ret < 0)
 441                 goto out;
 442
 443         leaf = path->nodes[0];
 444         nritems = btrfs_header_nritems(leaf);
 445
 446         while (1) {
 447                 if (btrfs_fs_closing(fs_info) > 1) {
 448                         last = (u64)-1;
 449                         break;
 450                 }
 451
 452                 if (path->slots[0] < nritems) {
 453                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 454                 } else {
 455                         ret = find_next_key(path, 0, &key);
 456                         if (ret)
 457                                 break;
 458
 459                         if (need_resched() ||
 460                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 461                                 if (wakeup)
 462                                         caching_ctl->progress = last;
 463                                 btrfs_release_path(path);
 464                                 up_read(&fs_info->commit_root_sem);
 465                                 mutex_unlock(&caching_ctl->mutex);
 466                                 cond_resched();
 467                                 mutex_lock(&caching_ctl->mutex);
 468                                 down_read(&fs_info->commit_root_sem);
 469                                 goto next;
 470                         }
 471
 472                         ret = btrfs_next_leaf(extent_root, path);
 473                         if (ret < 0)
 474                                 goto out;
 475                         if (ret)
 476                                 break;
 477                         leaf = path->nodes[0];
 478                         nritems = btrfs_header_nritems(leaf);
 479                         continue;
 480                 }
 481
 482                 if (key.objectid < last) {
 483                         key.objectid = last;
 484                         key.offset = 0;
 485                         key.type = BTRFS_EXTENT_ITEM_KEY;
 486
 487                         if (wakeup)
 488                                 caching_ctl->progress = last;
 489                         btrfs_release_path(path);
 490                         goto next;
 491                 }
 492
 493                 if (key.objectid < block_group->key.objectid) {
 494                         path->slots[0]++;
 495                         continue;
 496                 }
 497
 498                 if (key.objectid >= block_group->key.objectid +
 499                     block_group->key.offset)
 500                         break;
 501
 502                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 503                     key.type == BTRFS_METADATA_ITEM_KEY) {
 504                         total_found += add_new_free_space(block_group,
 505                                                           fs_info, last,
 506                                                           key.objectid);
 507                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 508                                 last = key.objectid +
 509                                         fs_info->nodesize;
 510                         else
 511                                 last = key.objectid + key.offset;
 512
 513                         if (total_found > CACHING_CTL_WAKE_UP) {
 514                                 total_found = 0;
 515                                 if (wakeup)
 516                                         wake_up(&caching_ctl->wait);
 517                         }
 518                 }
 519                 path->slots[0]++;
 520         }
 521         ret = 0;
 522
 523         total_found += add_new_free_space(block_group, fs_info, last,
 524                                           block_group->key.objectid +
 525                                           block_group->key.offset);
 526         caching_ctl->progress = (u64)-1;
 527
 528 out:
 529         btrfs_free_path(path);
 530         return ret;
 531 }
 532
 533 static noinline void caching_thread(struct btrfs_work *work)
 534 {
 535         struct btrfs_block_group_cache *block_group;
 536         struct btrfs_fs_info *fs_info;
 537         struct btrfs_caching_control *caching_ctl;
 538         int ret;
 539
 540         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 541         block_group = caching_ctl->block_group;
 542         fs_info = block_group->fs_info;
 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         refcount_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                 refcount_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 (btrfs_test_opt(fs_info, 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         refcount_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 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
 759                              u64 owner, u64 root_objectid)
 760 {
 761         struct btrfs_space_info *space_info;
 762         u64 flags;
 763
 764         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 765                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
 766                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 767                 else
 768                         flags = BTRFS_BLOCK_GROUP_METADATA;
 769         } else {
 770                 flags = BTRFS_BLOCK_GROUP_DATA;
 771         }
 772
 773         space_info = __find_space_info(fs_info, flags);
 774         ASSERT(space_info);
 775         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
 776 }
 777
 778 /*
 779  * after adding space to the filesystem, we need to clear the full flags
 780  * on all the space infos.
 781  */
 782 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 783 {
 784         struct list_head *head = &info->space_info;
 785         struct btrfs_space_info *found;
 786
 787         rcu_read_lock();
 788         list_for_each_entry_rcu(found, head, list)
 789                 found->full = 0;
 790         rcu_read_unlock();
 791 }
 792
 793 /* simple helper to search for an existing data extent at a given offset */
 794 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
 795 {
 796         int ret;
 797         struct btrfs_key key;
 798         struct btrfs_path *path;
 799
 800         path = btrfs_alloc_path();
 801         if (!path)
 802                 return -ENOMEM;
 803
 804         key.objectid = start;
 805         key.offset = len;
 806         key.type = BTRFS_EXTENT_ITEM_KEY;
 807         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
 808         btrfs_free_path(path);
 809         return ret;
 810 }
 811
 812 /*
 813  * helper function to lookup reference count and flags of a tree block.
 814  *
 815  * the head node for delayed ref is used to store the sum of all the
 816  * reference count modifications queued up in the rbtree. the head
 817  * node may also store the extent flags to set. This way you can check
 818  * to see what the reference count and extent flags would be if all of
 819  * the delayed refs are not processed.
 820  */
 821 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 822                              struct btrfs_fs_info *fs_info, u64 bytenr,
 823                              u64 offset, int metadata, u64 *refs, u64 *flags)
 824 {
 825         struct btrfs_delayed_ref_head *head;
 826         struct btrfs_delayed_ref_root *delayed_refs;
 827         struct btrfs_path *path;
 828         struct btrfs_extent_item *ei;
 829         struct extent_buffer *leaf;
 830         struct btrfs_key key;
 831         u32 item_size;
 832         u64 num_refs;
 833         u64 extent_flags;
 834         int ret;
 835
 836         /*
 837          * If we don't have skinny metadata, don't bother doing anything
 838          * different
 839          */
 840         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
 841                 offset = fs_info->nodesize;
 842                 metadata = 0;
 843         }
 844
 845         path = btrfs_alloc_path();
 846         if (!path)
 847                 return -ENOMEM;
 848
 849         if (!trans) {
 850                 path->skip_locking = 1;
 851                 path->search_commit_root = 1;
 852         }
 853
 854 search_again:
 855         key.objectid = bytenr;
 856         key.offset = offset;
 857         if (metadata)
 858                 key.type = BTRFS_METADATA_ITEM_KEY;
 859         else
 860                 key.type = BTRFS_EXTENT_ITEM_KEY;
 861
 862         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 863         if (ret < 0)
 864                 goto out_free;
 865
 866         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 867                 if (path->slots[0]) {
 868                         path->slots[0]--;
 869                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 870                                               path->slots[0]);
 871                         if (key.objectid == bytenr &&
 872                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 873                             key.offset == fs_info->nodesize)
 874                                 ret = 0;
 875                 }
 876         }
 877
 878         if (ret == 0) {
 879                 leaf = path->nodes[0];
 880                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 881                 if (item_size >= sizeof(*ei)) {
 882                         ei = btrfs_item_ptr(leaf, path->slots[0],
 883                                             struct btrfs_extent_item);
 884                         num_refs = btrfs_extent_refs(leaf, ei);
 885                         extent_flags = btrfs_extent_flags(leaf, ei);
 886                 } else {
 887 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 888                         struct btrfs_extent_item_v0 *ei0;
 889                         BUG_ON(item_size != sizeof(*ei0));
 890                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 891                                              struct btrfs_extent_item_v0);
 892                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 893                         /* FIXME: this isn't correct for data */
 894                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 895 #else
 896                         BUG();
 897 #endif
 898                 }
 899                 BUG_ON(num_refs == 0);
 900         } else {
 901                 num_refs = 0;
 902                 extent_flags = 0;
 903                 ret = 0;
 904         }
 905
 906         if (!trans)
 907                 goto out;
 908
 909         delayed_refs = &trans->transaction->delayed_refs;
 910         spin_lock(&delayed_refs->lock);
 911         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
 912         if (head) {
 913                 if (!mutex_trylock(&head->mutex)) {
 914                         refcount_inc(&head->refs);
 915                         spin_unlock(&delayed_refs->lock);
 916
 917                         btrfs_release_path(path);
 918
 919                         /*
 920                          * Mutex was contended, block until it's released and try
 921                          * again
 922                          */
 923                         mutex_lock(&head->mutex);
 924                         mutex_unlock(&head->mutex);
 925                         btrfs_put_delayed_ref_head(head);
 926                         goto search_again;
 927                 }
 928                 spin_lock(&head->lock);
 929                 if (head->extent_op && head->extent_op->update_flags)
 930                         extent_flags |= head->extent_op->flags_to_set;
 931                 else
 932                         BUG_ON(num_refs == 0);
 933
 934                 num_refs += head->ref_mod;
 935                 spin_unlock(&head->lock);
 936                 mutex_unlock(&head->mutex);
 937         }
 938         spin_unlock(&delayed_refs->lock);
 939 out:
 940         WARN_ON(num_refs == 0);
 941         if (refs)
 942                 *refs = num_refs;
 943         if (flags)
 944                 *flags = extent_flags;
 945 out_free:
 946         btrfs_free_path(path);
 947         return ret;
 948 }
 949
 950 /*
 951  * Back reference rules.  Back refs have three main goals:
 952  *
 953  * 1) differentiate between all holders of references to an extent so that
 954  *    when a reference is dropped we can make sure it was a valid reference
 955  *    before freeing the extent.
 956  *
 957  * 2) Provide enough information to quickly find the holders of an extent
 958  *    if we notice a given block is corrupted or bad.
 959  *
 960  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 961  *    maintenance.  This is actually the same as #2, but with a slightly
 962  *    different use case.
 963  *
 964  * There are two kinds of back refs. The implicit back refs is optimized
 965  * for pointers in non-shared tree blocks. For a given pointer in a block,
 966  * back refs of this kind provide information about the block's owner tree
 967  * and the pointer's key. These information allow us to find the block by
 968  * b-tree searching. The full back refs is for pointers in tree blocks not
 969  * referenced by their owner trees. The location of tree block is recorded
 970  * in the back refs. Actually the full back refs is generic, and can be
 971  * used in all cases the implicit back refs is used. The major shortcoming
 972  * of the full back refs is its overhead. Every time a tree block gets
 973  * COWed, we have to update back refs entry for all pointers in it.
 974  *
 975  * For a newly allocated tree block, we use implicit back refs for
 976  * pointers in it. This means most tree related operations only involve
 977  * implicit back refs. For a tree block created in old transaction, the
 978  * only way to drop a reference to it is COW it. So we can detect the
 979  * event that tree block loses its owner tree's reference and do the
 980  * back refs conversion.
 981  *
 982  * When a tree block is COWed through a tree, there are four cases:
 983  *
 984  * The reference count of the block is one and the tree is the block's
 985  * owner tree. Nothing to do in this case.
 986  *
 987  * The reference count of the block is one and the tree is not the
 988  * block's owner tree. In this case, full back refs is used for pointers
 989  * in the block. Remove these full back refs, add implicit back refs for
 990  * every pointers in the new block.
 991  *
 992  * The reference count of the block is greater than one and the tree is
 993  * the block's owner tree. In this case, implicit back refs is used for
 994  * pointers in the block. Add full back refs for every pointers in the
 995  * block, increase lower level extents' reference counts. The original
 996  * implicit back refs are entailed to the new block.
 997  *
 998  * The reference count of the block is greater than one and the tree is
 999  * not the block's owner tree. Add implicit back refs for every pointer in
1000  * the new block, increase lower level extents' reference count.
1001  *
1002  * Back Reference Key composing:
1003  *
1004  * The key objectid corresponds to the first byte in the extent,
1005  * The key type is used to differentiate between types of back refs.
1006  * There are different meanings of the key offset for different types
1007  * of back refs.
1008  *
1009  * File extents can be referenced by:
1010  *
1011  * - multiple snapshots, subvolumes, or different generations in one subvol
1012  * - different files inside a single subvolume
1013  * - different offsets inside a file (bookend extents in file.c)
1014  *
1015  * The extent ref structure for the implicit back refs has fields for:
1016  *
1017  * - Objectid of the subvolume root
1018  * - objectid of the file holding the reference
1019  * - original offset in the file
1020  * - how many bookend extents
1021  *
1022  * The key offset for the implicit back refs is hash of the first
1023  * three fields.
1024  *
1025  * The extent ref structure for the full back refs has field for:
1026  *
1027  * - number of pointers in the tree leaf
1028  *
1029  * The key offset for the implicit back refs is the first byte of
1030  * the tree leaf
1031  *
1032  * When a file extent is allocated, The implicit back refs is used.
1033  * the fields are filled in:
1034  *
1035  *     (root_key.objectid, inode objectid, offset in file, 1)
1036  *
1037  * When a file extent is removed file truncation, we find the
1038  * corresponding implicit back refs and check the following fields:
1039  *
1040  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1041  *
1042  * Btree extents can be referenced by:
1043  *
1044  * - Different subvolumes
1045  *
1046  * Both the implicit back refs and the full back refs for tree blocks
1047  * only consist of key. The key offset for the implicit back refs is
1048  * objectid of block's owner tree. The key offset for the full back refs
1049  * is the first byte of parent block.
1050  *
1051  * When implicit back refs is used, information about the lowest key and
1052  * level of the tree block are required. These information are stored in
1053  * tree block info structure.
1054  */
1055
1056 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1057 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1058                                   struct btrfs_fs_info *fs_info,
1059                                   struct btrfs_path *path,
1060                                   u64 owner, u32 extra_size)
1061 {
1062         struct btrfs_root *root = fs_info->extent_root;
1063         struct btrfs_extent_item *item;
1064         struct btrfs_extent_item_v0 *ei0;
1065         struct btrfs_extent_ref_v0 *ref0;
1066         struct btrfs_tree_block_info *bi;
1067         struct extent_buffer *leaf;
1068         struct btrfs_key key;
1069         struct btrfs_key found_key;
1070         u32 new_size = sizeof(*item);
1071         u64 refs;
1072         int ret;
1073
1074         leaf = path->nodes[0];
1075         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1076
1077         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1078         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1079                              struct btrfs_extent_item_v0);
1080         refs = btrfs_extent_refs_v0(leaf, ei0);
1081
1082         if (owner == (u64)-1) {
1083                 while (1) {
1084                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1085                                 ret = btrfs_next_leaf(root, path);
1086                                 if (ret < 0)
1087                                         return ret;
1088                                 BUG_ON(ret > 0); /* Corruption */
1089                                 leaf = path->nodes[0];
1090                         }
1091                         btrfs_item_key_to_cpu(leaf, &found_key,
1092                                               path->slots[0]);
1093                         BUG_ON(key.objectid != found_key.objectid);
1094                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1095                                 path->slots[0]++;
1096                                 continue;
1097                         }
1098                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1099                                               struct btrfs_extent_ref_v0);
1100                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1101                         break;
1102                 }
1103         }
1104         btrfs_release_path(path);
1105
1106         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1107                 new_size += sizeof(*bi);
1108
1109         new_size -= sizeof(*ei0);
1110         ret = btrfs_search_slot(trans, root, &key, path,
1111                                 new_size + extra_size, 1);
1112         if (ret < 0)
1113                 return ret;
1114         BUG_ON(ret); /* Corruption */
1115
1116         btrfs_extend_item(fs_info, path, new_size);
1117
1118         leaf = path->nodes[0];
1119         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1120         btrfs_set_extent_refs(leaf, item, refs);
1121         /* FIXME: get real generation */
1122         btrfs_set_extent_generation(leaf, item, 0);
1123         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1124                 btrfs_set_extent_flags(leaf, item,
1125                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1126                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1127                 bi = (struct btrfs_tree_block_info *)(item + 1);
1128                 /* FIXME: get first key of the block */
1129                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1130                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1131         } else {
1132                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1133         }
1134         btrfs_mark_buffer_dirty(leaf);
1135         return 0;
1136 }
1137 #endif
1138
1139 /*
1140  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1141  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1142  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1143  */
1144 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1145                                      struct btrfs_extent_inline_ref *iref,
1146                                      enum btrfs_inline_ref_type is_data)
1147 {
1148         int type = btrfs_extent_inline_ref_type(eb, iref);
1149         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1150
1151         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1152             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1153             type == BTRFS_SHARED_DATA_REF_KEY ||
1154             type == BTRFS_EXTENT_DATA_REF_KEY) {
1155                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1156                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1157                                 return type;
1158                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1159                                 ASSERT(eb->fs_info);
1160                                 /*
1161                                  * Every shared one has parent tree
1162                                  * block, which must be aligned to
1163                                  * nodesize.
1164                                  */
1165                                 if (offset &&
1166                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1167                                         return type;
1168                         }
1169                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1170                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1171                                 return type;
1172                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1173                                 ASSERT(eb->fs_info);
1174                                 /*
1175                                  * Every shared one has parent tree
1176                                  * block, which must be aligned to
1177                                  * nodesize.
1178                                  */
1179                                 if (offset &&
1180                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1181                                         return type;
1182                         }
1183                 } else {
1184                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1185                         return type;
1186                 }
1187         }
1188
1189         btrfs_print_leaf((struct extent_buffer *)eb);
1190         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1191                   eb->start, type);
1192         WARN_ON(1);
1193
1194         return BTRFS_REF_TYPE_INVALID;
1195 }
1196
1197 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1198 {
1199         u32 high_crc = ~(u32)0;
1200         u32 low_crc = ~(u32)0;
1201         __le64 lenum;
1202
1203         lenum = cpu_to_le64(root_objectid);
1204         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1205         lenum = cpu_to_le64(owner);
1206         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1207         lenum = cpu_to_le64(offset);
1208         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1209
1210         return ((u64)high_crc << 31) ^ (u64)low_crc;
1211 }
1212
1213 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1214                                      struct btrfs_extent_data_ref *ref)
1215 {
1216         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1217                                     btrfs_extent_data_ref_objectid(leaf, ref),
1218                                     btrfs_extent_data_ref_offset(leaf, ref));
1219 }
1220
1221 static int match_extent_data_ref(struct extent_buffer *leaf,
1222                                  struct btrfs_extent_data_ref *ref,
1223                                  u64 root_objectid, u64 owner, u64 offset)
1224 {
1225         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1226             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1227             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1228                 return 0;
1229         return 1;
1230 }
1231
1232 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1233                                            struct btrfs_fs_info *fs_info,
1234                                            struct btrfs_path *path,
1235                                            u64 bytenr, u64 parent,
1236                                            u64 root_objectid,
1237                                            u64 owner, u64 offset)
1238 {
1239         struct btrfs_root *root = fs_info->extent_root;
1240         struct btrfs_key key;
1241         struct btrfs_extent_data_ref *ref;
1242         struct extent_buffer *leaf;
1243         u32 nritems;
1244         int ret;
1245         int recow;
1246         int err = -ENOENT;
1247
1248         key.objectid = bytenr;
1249         if (parent) {
1250                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1251                 key.offset = parent;
1252         } else {
1253                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1254                 key.offset = hash_extent_data_ref(root_objectid,
1255                                                   owner, offset);
1256         }
1257 again:
1258         recow = 0;
1259         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1260         if (ret < 0) {
1261                 err = ret;
1262                 goto fail;
1263         }
1264
1265         if (parent) {
1266                 if (!ret)
1267                         return 0;
1268 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1269                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1270                 btrfs_release_path(path);
1271                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1272                 if (ret < 0) {
1273                         err = ret;
1274                         goto fail;
1275                 }
1276                 if (!ret)
1277                         return 0;
1278 #endif
1279                 goto fail;
1280         }
1281
1282         leaf = path->nodes[0];
1283         nritems = btrfs_header_nritems(leaf);
1284         while (1) {
1285                 if (path->slots[0] >= nritems) {
1286                         ret = btrfs_next_leaf(root, path);
1287                         if (ret < 0)
1288                                 err = ret;
1289                         if (ret)
1290                                 goto fail;
1291
1292                         leaf = path->nodes[0];
1293                         nritems = btrfs_header_nritems(leaf);
1294                         recow = 1;
1295                 }
1296
1297                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1298                 if (key.objectid != bytenr ||
1299                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1300                         goto fail;
1301
1302                 ref = btrfs_item_ptr(leaf, path->slots[0],
1303                                      struct btrfs_extent_data_ref);
1304
1305                 if (match_extent_data_ref(leaf, ref, root_objectid,
1306                                           owner, offset)) {
1307                         if (recow) {
1308                                 btrfs_release_path(path);
1309                                 goto again;
1310                         }
1311                         err = 0;
1312                         break;
1313                 }
1314                 path->slots[0]++;
1315         }
1316 fail:
1317         return err;
1318 }
1319
1320 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1321                                            struct btrfs_fs_info *fs_info,
1322                                            struct btrfs_path *path,
1323                                            u64 bytenr, u64 parent,
1324                                            u64 root_objectid, u64 owner,
1325                                            u64 offset, int refs_to_add)
1326 {
1327         struct btrfs_root *root = fs_info->extent_root;
1328         struct btrfs_key key;
1329         struct extent_buffer *leaf;
1330         u32 size;
1331         u32 num_refs;
1332         int ret;
1333
1334         key.objectid = bytenr;
1335         if (parent) {
1336                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1337                 key.offset = parent;
1338                 size = sizeof(struct btrfs_shared_data_ref);
1339         } else {
1340                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1341                 key.offset = hash_extent_data_ref(root_objectid,
1342                                                   owner, offset);
1343                 size = sizeof(struct btrfs_extent_data_ref);
1344         }
1345
1346         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1347         if (ret && ret != -EEXIST)
1348                 goto fail;
1349
1350         leaf = path->nodes[0];
1351         if (parent) {
1352                 struct btrfs_shared_data_ref *ref;
1353                 ref = btrfs_item_ptr(leaf, path->slots[0],
1354                                      struct btrfs_shared_data_ref);
1355                 if (ret == 0) {
1356                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1357                 } else {
1358                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1359                         num_refs += refs_to_add;
1360                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1361                 }
1362         } else {
1363                 struct btrfs_extent_data_ref *ref;
1364                 while (ret == -EEXIST) {
1365                         ref = btrfs_item_ptr(leaf, path->slots[0],
1366                                              struct btrfs_extent_data_ref);
1367                         if (match_extent_data_ref(leaf, ref, root_objectid,
1368                                                   owner, offset))
1369                                 break;
1370                         btrfs_release_path(path);
1371                         key.offset++;
1372                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1373                                                       size);
1374                         if (ret && ret != -EEXIST)
1375                                 goto fail;
1376
1377                         leaf = path->nodes[0];
1378                 }
1379                 ref = btrfs_item_ptr(leaf, path->slots[0],
1380                                      struct btrfs_extent_data_ref);
1381                 if (ret == 0) {
1382                         btrfs_set_extent_data_ref_root(leaf, ref,
1383                                                        root_objectid);
1384                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1385                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1386                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1387                 } else {
1388                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1389                         num_refs += refs_to_add;
1390                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1391                 }
1392         }
1393         btrfs_mark_buffer_dirty(leaf);
1394         ret = 0;
1395 fail:
1396         btrfs_release_path(path);
1397         return ret;
1398 }
1399
1400 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1401                                            struct btrfs_fs_info *fs_info,
1402                                            struct btrfs_path *path,
1403                                            int refs_to_drop, int *last_ref)
1404 {
1405         struct btrfs_key key;
1406         struct btrfs_extent_data_ref *ref1 = NULL;
1407         struct btrfs_shared_data_ref *ref2 = NULL;
1408         struct extent_buffer *leaf;
1409         u32 num_refs = 0;
1410         int ret = 0;
1411
1412         leaf = path->nodes[0];
1413         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1414
1415         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1416                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1417                                       struct btrfs_extent_data_ref);
1418                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1419         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1420                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1421                                       struct btrfs_shared_data_ref);
1422                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1423 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1424         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1425                 struct btrfs_extent_ref_v0 *ref0;
1426                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1427                                       struct btrfs_extent_ref_v0);
1428                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1429 #endif
1430         } else {
1431                 BUG();
1432         }
1433
1434         BUG_ON(num_refs < refs_to_drop);
1435         num_refs -= refs_to_drop;
1436
1437         if (num_refs == 0) {
1438                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1439                 *last_ref = 1;
1440         } else {
1441                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1442                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1443                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1444                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1445 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1446                 else {
1447                         struct btrfs_extent_ref_v0 *ref0;
1448                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1449                                         struct btrfs_extent_ref_v0);
1450                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1451                 }
1452 #endif
1453                 btrfs_mark_buffer_dirty(leaf);
1454         }
1455         return ret;
1456 }
1457
1458 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1459                                           struct btrfs_extent_inline_ref *iref)
1460 {
1461         struct btrfs_key key;
1462         struct extent_buffer *leaf;
1463         struct btrfs_extent_data_ref *ref1;
1464         struct btrfs_shared_data_ref *ref2;
1465         u32 num_refs = 0;
1466         int type;
1467
1468         leaf = path->nodes[0];
1469         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1470         if (iref) {
1471                 /*
1472                  * If type is invalid, we should have bailed out earlier than
1473                  * this call.
1474                  */
1475                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1476                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1477                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1478                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1479                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1480                 } else {
1481                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1482                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1483                 }
1484         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1485                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1486                                       struct btrfs_extent_data_ref);
1487                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1488         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1489                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1490                                       struct btrfs_shared_data_ref);
1491                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1492 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1493         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1494                 struct btrfs_extent_ref_v0 *ref0;
1495                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1496                                       struct btrfs_extent_ref_v0);
1497                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1498 #endif
1499         } else {
1500                 WARN_ON(1);
1501         }
1502         return num_refs;
1503 }
1504
1505 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1506                                           struct btrfs_fs_info *fs_info,
1507                                           struct btrfs_path *path,
1508                                           u64 bytenr, u64 parent,
1509                                           u64 root_objectid)
1510 {
1511         struct btrfs_root *root = fs_info->extent_root;
1512         struct btrfs_key key;
1513         int ret;
1514
1515         key.objectid = bytenr;
1516         if (parent) {
1517                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1518                 key.offset = parent;
1519         } else {
1520                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1521                 key.offset = root_objectid;
1522         }
1523
1524         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1525         if (ret > 0)
1526                 ret = -ENOENT;
1527 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1528         if (ret == -ENOENT && parent) {
1529                 btrfs_release_path(path);
1530                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1531                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1532                 if (ret > 0)
1533                         ret = -ENOENT;
1534         }
1535 #endif
1536         return ret;
1537 }
1538
1539 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1540                                           struct btrfs_fs_info *fs_info,
1541                                           struct btrfs_path *path,
1542                                           u64 bytenr, u64 parent,
1543                                           u64 root_objectid)
1544 {
1545         struct btrfs_key key;
1546         int ret;
1547
1548         key.objectid = bytenr;
1549         if (parent) {
1550                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1551                 key.offset = parent;
1552         } else {
1553                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1554                 key.offset = root_objectid;
1555         }
1556
1557         ret = btrfs_insert_empty_item(trans, fs_info->extent_root,
1558                                       path, &key, 0);
1559         btrfs_release_path(path);
1560         return ret;
1561 }
1562
1563 static inline int extent_ref_type(u64 parent, u64 owner)
1564 {
1565         int type;
1566         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1567                 if (parent > 0)
1568                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1569                 else
1570                         type = BTRFS_TREE_BLOCK_REF_KEY;
1571         } else {
1572                 if (parent > 0)
1573                         type = BTRFS_SHARED_DATA_REF_KEY;
1574                 else
1575                         type = BTRFS_EXTENT_DATA_REF_KEY;
1576         }
1577         return type;
1578 }
1579
1580 static int find_next_key(struct btrfs_path *path, int level,
1581                          struct btrfs_key *key)
1582
1583 {
1584         for (; level < BTRFS_MAX_LEVEL; level++) {
1585                 if (!path->nodes[level])
1586                         break;
1587                 if (path->slots[level] + 1 >=
1588                     btrfs_header_nritems(path->nodes[level]))
1589                         continue;
1590                 if (level == 0)
1591                         btrfs_item_key_to_cpu(path->nodes[level], key,
1592                                               path->slots[level] + 1);
1593                 else
1594                         btrfs_node_key_to_cpu(path->nodes[level], key,
1595                                               path->slots[level] + 1);
1596                 return 0;
1597         }
1598         return 1;
1599 }
1600
1601 /*
1602  * look for inline back ref. if back ref is found, *ref_ret is set
1603  * to the address of inline back ref, and 0 is returned.
1604  *
1605  * if back ref isn't found, *ref_ret is set to the address where it
1606  * should be inserted, and -ENOENT is returned.
1607  *
1608  * if insert is true and there are too many inline back refs, the path
1609  * points to the extent item, and -EAGAIN is returned.
1610  *
1611  * NOTE: inline back refs are ordered in the same way that back ref
1612  *       items in the tree are ordered.
1613  */
1614 static noinline_for_stack
1615 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1616                                  struct btrfs_fs_info *fs_info,
1617                                  struct btrfs_path *path,
1618                                  struct btrfs_extent_inline_ref **ref_ret,
1619                                  u64 bytenr, u64 num_bytes,
1620                                  u64 parent, u64 root_objectid,
1621                                  u64 owner, u64 offset, int insert)
1622 {
1623         struct btrfs_root *root = fs_info->extent_root;
1624         struct btrfs_key key;
1625         struct extent_buffer *leaf;
1626         struct btrfs_extent_item *ei;
1627         struct btrfs_extent_inline_ref *iref;
1628         u64 flags;
1629         u64 item_size;
1630         unsigned long ptr;
1631         unsigned long end;
1632         int extra_size;
1633         int type;
1634         int want;
1635         int ret;
1636         int err = 0;
1637         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1638         int needed;
1639
1640         key.objectid = bytenr;
1641         key.type = BTRFS_EXTENT_ITEM_KEY;
1642         key.offset = num_bytes;
1643
1644         want = extent_ref_type(parent, owner);
1645         if (insert) {
1646                 extra_size = btrfs_extent_inline_ref_size(want);
1647                 path->keep_locks = 1;
1648         } else
1649                 extra_size = -1;
1650
1651         /*
1652          * Owner is our parent level, so we can just add one to get the level
1653          * for the block we are interested in.
1654          */
1655         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1656                 key.type = BTRFS_METADATA_ITEM_KEY;
1657                 key.offset = owner;
1658         }
1659
1660 again:
1661         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1662         if (ret < 0) {
1663                 err = ret;
1664                 goto out;
1665         }
1666
1667         /*
1668          * We may be a newly converted file system which still has the old fat
1669          * extent entries for metadata, so try and see if we have one of those.
1670          */
1671         if (ret > 0 && skinny_metadata) {
1672                 skinny_metadata = false;
1673                 if (path->slots[0]) {
1674                         path->slots[0]--;
1675                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1676                                               path->slots[0]);
1677                         if (key.objectid == bytenr &&
1678                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1679                             key.offset == num_bytes)
1680                                 ret = 0;
1681                 }
1682                 if (ret) {
1683                         key.objectid = bytenr;
1684                         key.type = BTRFS_EXTENT_ITEM_KEY;
1685                         key.offset = num_bytes;
1686                         btrfs_release_path(path);
1687                         goto again;
1688                 }
1689         }
1690
1691         if (ret && !insert) {
1692                 err = -ENOENT;
1693                 goto out;
1694         } else if (WARN_ON(ret)) {
1695                 err = -EIO;
1696                 goto out;
1697         }
1698
1699         leaf = path->nodes[0];
1700         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1701 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1702         if (item_size < sizeof(*ei)) {
1703                 if (!insert) {
1704                         err = -ENOENT;
1705                         goto out;
1706                 }
1707                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1708                                              extra_size);
1709                 if (ret < 0) {
1710                         err = ret;
1711                         goto out;
1712                 }
1713                 leaf = path->nodes[0];
1714                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1715         }
1716 #endif
1717         BUG_ON(item_size < sizeof(*ei));
1718
1719         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1720         flags = btrfs_extent_flags(leaf, ei);
1721
1722         ptr = (unsigned long)(ei + 1);
1723         end = (unsigned long)ei + item_size;
1724
1725         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1726                 ptr += sizeof(struct btrfs_tree_block_info);
1727                 BUG_ON(ptr > end);
1728         }
1729
1730         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1731                 needed = BTRFS_REF_TYPE_DATA;
1732         else
1733                 needed = BTRFS_REF_TYPE_BLOCK;
1734
1735         err = -ENOENT;
1736         while (1) {
1737                 if (ptr >= end) {
1738                         WARN_ON(ptr > end);
1739                         break;
1740                 }
1741                 iref = (struct btrfs_extent_inline_ref *)ptr;
1742                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1743                 if (type == BTRFS_REF_TYPE_INVALID) {
1744                         err = -EINVAL;
1745                         goto out;
1746                 }
1747
1748                 if (want < type)
1749                         break;
1750                 if (want > type) {
1751                         ptr += btrfs_extent_inline_ref_size(type);
1752                         continue;
1753                 }
1754
1755                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1756                         struct btrfs_extent_data_ref *dref;
1757                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1758                         if (match_extent_data_ref(leaf, dref, root_objectid,
1759                                                   owner, offset)) {
1760                                 err = 0;
1761                                 break;
1762                         }
1763                         if (hash_extent_data_ref_item(leaf, dref) <
1764                             hash_extent_data_ref(root_objectid, owner, offset))
1765                                 break;
1766                 } else {
1767                         u64 ref_offset;
1768                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1769                         if (parent > 0) {
1770                                 if (parent == ref_offset) {
1771                                         err = 0;
1772                                         break;
1773                                 }
1774                                 if (ref_offset < parent)
1775                                         break;
1776                         } else {
1777                                 if (root_objectid == ref_offset) {
1778                                         err = 0;
1779                                         break;
1780                                 }
1781                                 if (ref_offset < root_objectid)
1782                                         break;
1783                         }
1784                 }
1785                 ptr += btrfs_extent_inline_ref_size(type);
1786         }
1787         if (err == -ENOENT && insert) {
1788                 if (item_size + extra_size >=
1789                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1790                         err = -EAGAIN;
1791                         goto out;
1792                 }
1793                 /*
1794                  * To add new inline back ref, we have to make sure
1795                  * there is no corresponding back ref item.
1796                  * For simplicity, we just do not add new inline back
1797                  * ref if there is any kind of item for this block
1798                  */
1799                 if (find_next_key(path, 0, &key) == 0 &&
1800                     key.objectid == bytenr &&
1801                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1802                         err = -EAGAIN;
1803                         goto out;
1804                 }
1805         }
1806         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1807 out:
1808         if (insert) {
1809                 path->keep_locks = 0;
1810                 btrfs_unlock_up_safe(path, 1);
1811         }
1812         return err;
1813 }
1814
1815 /*
1816  * helper to add new inline back ref
1817  */
1818 static noinline_for_stack
1819 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1820                                  struct btrfs_path *path,
1821                                  struct btrfs_extent_inline_ref *iref,
1822                                  u64 parent, u64 root_objectid,
1823                                  u64 owner, u64 offset, int refs_to_add,
1824                                  struct btrfs_delayed_extent_op *extent_op)
1825 {
1826         struct extent_buffer *leaf;
1827         struct btrfs_extent_item *ei;
1828         unsigned long ptr;
1829         unsigned long end;
1830         unsigned long item_offset;
1831         u64 refs;
1832         int size;
1833         int type;
1834
1835         leaf = path->nodes[0];
1836         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1837         item_offset = (unsigned long)iref - (unsigned long)ei;
1838
1839         type = extent_ref_type(parent, owner);
1840         size = btrfs_extent_inline_ref_size(type);
1841
1842         btrfs_extend_item(fs_info, path, size);
1843
1844         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1845         refs = btrfs_extent_refs(leaf, ei);
1846         refs += refs_to_add;
1847         btrfs_set_extent_refs(leaf, ei, refs);
1848         if (extent_op)
1849                 __run_delayed_extent_op(extent_op, leaf, ei);
1850
1851         ptr = (unsigned long)ei + item_offset;
1852         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1853         if (ptr < end - size)
1854                 memmove_extent_buffer(leaf, ptr + size, ptr,
1855                                       end - size - ptr);
1856
1857         iref = (struct btrfs_extent_inline_ref *)ptr;
1858         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1859         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1860                 struct btrfs_extent_data_ref *dref;
1861                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1862                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1863                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1864                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1865                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1866         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1867                 struct btrfs_shared_data_ref *sref;
1868                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1869                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1870                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1871         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1872                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1873         } else {
1874                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1875         }
1876         btrfs_mark_buffer_dirty(leaf);
1877 }
1878
1879 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1880                                  struct btrfs_fs_info *fs_info,
1881                                  struct btrfs_path *path,
1882                                  struct btrfs_extent_inline_ref **ref_ret,
1883                                  u64 bytenr, u64 num_bytes, u64 parent,
1884                                  u64 root_objectid, u64 owner, u64 offset)
1885 {
1886         int ret;
1887
1888         ret = lookup_inline_extent_backref(trans, fs_info, path, ref_ret,
1889                                            bytenr, num_bytes, parent,
1890                                            root_objectid, owner, offset, 0);
1891         if (ret != -ENOENT)
1892                 return ret;
1893
1894         btrfs_release_path(path);
1895         *ref_ret = NULL;
1896
1897         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1898                 ret = lookup_tree_block_ref(trans, fs_info, path, bytenr,
1899                                             parent, root_objectid);
1900         } else {
1901                 ret = lookup_extent_data_ref(trans, fs_info, path, bytenr,
1902                                              parent, root_objectid, owner,
1903                                              offset);
1904         }
1905         return ret;
1906 }
1907
1908 /*
1909  * helper to update/remove inline back ref
1910  */
1911 static noinline_for_stack
1912 void update_inline_extent_backref(struct btrfs_fs_info *fs_info,
1913                                   struct btrfs_path *path,
1914                                   struct btrfs_extent_inline_ref *iref,
1915                                   int refs_to_mod,
1916                                   struct btrfs_delayed_extent_op *extent_op,
1917                                   int *last_ref)
1918 {
1919         struct extent_buffer *leaf;
1920         struct btrfs_extent_item *ei;
1921         struct btrfs_extent_data_ref *dref = NULL;
1922         struct btrfs_shared_data_ref *sref = NULL;
1923         unsigned long ptr;
1924         unsigned long end;
1925         u32 item_size;
1926         int size;
1927         int type;
1928         u64 refs;
1929
1930         leaf = path->nodes[0];
1931         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1932         refs = btrfs_extent_refs(leaf, ei);
1933         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1934         refs += refs_to_mod;
1935         btrfs_set_extent_refs(leaf, ei, refs);
1936         if (extent_op)
1937                 __run_delayed_extent_op(extent_op, leaf, ei);
1938
1939         /*
1940          * If type is invalid, we should have bailed out after
1941          * lookup_inline_extent_backref().
1942          */
1943         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1944         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1945
1946         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1947                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1948                 refs = btrfs_extent_data_ref_count(leaf, dref);
1949         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1950                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1951                 refs = btrfs_shared_data_ref_count(leaf, sref);
1952         } else {
1953                 refs = 1;
1954                 BUG_ON(refs_to_mod != -1);
1955         }
1956
1957         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1958         refs += refs_to_mod;
1959
1960         if (refs > 0) {
1961                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1962                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1963                 else
1964                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1965         } else {
1966                 *last_ref = 1;
1967                 size =  btrfs_extent_inline_ref_size(type);
1968                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1969                 ptr = (unsigned long)iref;
1970                 end = (unsigned long)ei + item_size;
1971                 if (ptr + size < end)
1972                         memmove_extent_buffer(leaf, ptr, ptr + size,
1973                                               end - ptr - size);
1974                 item_size -= size;
1975                 btrfs_truncate_item(fs_info, path, item_size, 1);
1976         }
1977         btrfs_mark_buffer_dirty(leaf);
1978 }
1979
1980 static noinline_for_stack
1981 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1982                                  struct btrfs_fs_info *fs_info,
1983                                  struct btrfs_path *path,
1984                                  u64 bytenr, u64 num_bytes, u64 parent,
1985                                  u64 root_objectid, u64 owner,
1986                                  u64 offset, int refs_to_add,
1987                                  struct btrfs_delayed_extent_op *extent_op)
1988 {
1989         struct btrfs_extent_inline_ref *iref;
1990         int ret;
1991
1992         ret = lookup_inline_extent_backref(trans, fs_info, path, &iref,
1993                                            bytenr, num_bytes, parent,
1994                                            root_objectid, owner, offset, 1);
1995         if (ret == 0) {
1996                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1997                 update_inline_extent_backref(fs_info, path, iref,
1998                                              refs_to_add, extent_op, NULL);
1999         } else if (ret == -ENOENT) {
2000                 setup_inline_extent_backref(fs_info, path, iref, parent,
2001                                             root_objectid, owner, offset,
2002                                             refs_to_add, extent_op);
2003                 ret = 0;
2004         }
2005         return ret;
2006 }
2007
2008 static int insert_extent_backref(struct btrfs_trans_handle *trans,
2009                                  struct btrfs_fs_info *fs_info,
2010                                  struct btrfs_path *path,
2011                                  u64 bytenr, u64 parent, u64 root_objectid,
2012                                  u64 owner, u64 offset, int refs_to_add)
2013 {
2014         int ret;
2015         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2016                 BUG_ON(refs_to_add != 1);
2017                 ret = insert_tree_block_ref(trans, fs_info, path, bytenr,
2018                                             parent, root_objectid);
2019         } else {
2020                 ret = insert_extent_data_ref(trans, fs_info, path, bytenr,
2021                                              parent, root_objectid,
2022                                              owner, offset, refs_to_add);
2023         }
2024         return ret;
2025 }
2026
2027 static int remove_extent_backref(struct btrfs_trans_handle *trans,
2028                                  struct btrfs_fs_info *fs_info,
2029                                  struct btrfs_path *path,
2030                                  struct btrfs_extent_inline_ref *iref,
2031                                  int refs_to_drop, int is_data, int *last_ref)
2032 {
2033         int ret = 0;
2034
2035         BUG_ON(!is_data && refs_to_drop != 1);
2036         if (iref) {
2037                 update_inline_extent_backref(fs_info, path, iref,
2038                                              -refs_to_drop, NULL, last_ref);
2039         } else if (is_data) {
2040                 ret = remove_extent_data_ref(trans, fs_info, path, refs_to_drop,
2041                                              last_ref);
2042         } else {
2043                 *last_ref = 1;
2044                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
2045         }
2046         return ret;
2047 }
2048
2049 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
2050 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
2051                                u64 *discarded_bytes)
2052 {
2053         int j, ret = 0;
2054         u64 bytes_left, end;
2055         u64 aligned_start = ALIGN(start, 1 << 9);
2056
2057         if (WARN_ON(start != aligned_start)) {
2058                 len -= aligned_start - start;
2059                 len = round_down(len, 1 << 9);
2060                 start = aligned_start;
2061         }
2062
2063         *discarded_bytes = 0;
2064
2065         if (!len)
2066                 return 0;
2067
2068         end = start + len;
2069         bytes_left = len;
2070
2071         /* Skip any superblocks on this device. */
2072         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
2073                 u64 sb_start = btrfs_sb_offset(j);
2074                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2075                 u64 size = sb_start - start;
2076
2077                 if (!in_range(sb_start, start, bytes_left) &&
2078                     !in_range(sb_end, start, bytes_left) &&
2079                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2080                         continue;
2081
2082                 /*
2083                  * Superblock spans beginning of range.  Adjust start and
2084                  * try again.
2085                  */
2086                 if (sb_start <= start) {
2087                         start += sb_end - start;
2088                         if (start > end) {
2089                                 bytes_left = 0;
2090                                 break;
2091                         }
2092                         bytes_left = end - start;
2093                         continue;
2094                 }
2095
2096                 if (size) {
2097                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2098                                                    GFP_NOFS, 0);
2099                         if (!ret)
2100                                 *discarded_bytes += size;
2101                         else if (ret != -EOPNOTSUPP)
2102                                 return ret;
2103                 }
2104
2105                 start = sb_end;
2106                 if (start > end) {
2107                         bytes_left = 0;
2108                         break;
2109                 }
2110                 bytes_left = end - start;
2111         }
2112
2113         if (bytes_left) {
2114                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2115                                            GFP_NOFS, 0);
2116                 if (!ret)
2117                         *discarded_bytes += bytes_left;
2118         }
2119         return ret;
2120 }
2121
2122 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2123                          u64 num_bytes, u64 *actual_bytes)
2124 {
2125         int ret;
2126         u64 discarded_bytes = 0;
2127         struct btrfs_bio *bbio = NULL;
2128
2129
2130         /*
2131          * Avoid races with device replace and make sure our bbio has devices
2132          * associated to its stripes that don't go away while we are discarding.
2133          */
2134         btrfs_bio_counter_inc_blocked(fs_info);
2135         /* Tell the block device(s) that the sectors can be discarded */
2136         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2137                               &bbio, 0);
2138         /* Error condition is -ENOMEM */
2139         if (!ret) {
2140                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2141                 int i;
2142
2143
2144                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2145                         u64 bytes;
2146                         struct request_queue *req_q;
2147
2148                         if (!stripe->dev->bdev) {
2149                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2150                                 continue;
2151                         }
2152                         req_q = bdev_get_queue(stripe->dev->bdev);
2153                         if (!blk_queue_discard(req_q))
2154                                 continue;
2155
2156                         ret = btrfs_issue_discard(stripe->dev->bdev,
2157                                                   stripe->physical,
2158                                                   stripe->length,
2159                                                   &bytes);
2160                         if (!ret)
2161                                 discarded_bytes += bytes;
2162                         else if (ret != -EOPNOTSUPP)
2163                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2164
2165                         /*
2166                          * Just in case we get back EOPNOTSUPP for some reason,
2167                          * just ignore the return value so we don't screw up
2168                          * people calling discard_extent.
2169                          */
2170                         ret = 0;
2171                 }
2172                 btrfs_put_bbio(bbio);
2173         }
2174         btrfs_bio_counter_dec(fs_info);
2175
2176         if (actual_bytes)
2177                 *actual_bytes = discarded_bytes;
2178
2179
2180         if (ret == -EOPNOTSUPP)
2181                 ret = 0;
2182         return ret;
2183 }
2184
2185 /* Can return -ENOMEM */
2186 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2187                          struct btrfs_root *root,
2188                          u64 bytenr, u64 num_bytes, u64 parent,
2189                          u64 root_objectid, u64 owner, u64 offset)
2190 {
2191         struct btrfs_fs_info *fs_info = root->fs_info;
2192         int old_ref_mod, new_ref_mod;
2193         int ret;
2194
2195         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2196                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2197
2198         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2199                            owner, offset, BTRFS_ADD_DELAYED_REF);
2200
2201         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2202                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2203                                                  num_bytes, parent,
2204                                                  root_objectid, (int)owner,
2205                                                  BTRFS_ADD_DELAYED_REF, NULL,
2206                                                  &old_ref_mod, &new_ref_mod);
2207         } else {
2208                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2209                                                  num_bytes, parent,
2210                                                  root_objectid, owner, offset,
2211                                                  0, BTRFS_ADD_DELAYED_REF,
2212                                                  &old_ref_mod, &new_ref_mod);
2213         }
2214
2215         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
2216                 add_pinned_bytes(fs_info, -num_bytes, owner, root_objectid);
2217
2218         return ret;
2219 }
2220
2221 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2222                                   struct btrfs_fs_info *fs_info,
2223                                   struct btrfs_delayed_ref_node *node,
2224                                   u64 parent, u64 root_objectid,
2225                                   u64 owner, u64 offset, int refs_to_add,
2226                                   struct btrfs_delayed_extent_op *extent_op)
2227 {
2228         struct btrfs_path *path;
2229         struct extent_buffer *leaf;
2230         struct btrfs_extent_item *item;
2231         struct btrfs_key key;
2232         u64 bytenr = node->bytenr;
2233         u64 num_bytes = node->num_bytes;
2234         u64 refs;
2235         int ret;
2236
2237         path = btrfs_alloc_path();
2238         if (!path)
2239                 return -ENOMEM;
2240
2241         path->reada = READA_FORWARD;
2242         path->leave_spinning = 1;
2243         /* this will setup the path even if it fails to insert the back ref */
2244         ret = insert_inline_extent_backref(trans, fs_info, path, bytenr,
2245                                            num_bytes, parent, root_objectid,
2246                                            owner, offset,
2247                                            refs_to_add, extent_op);
2248         if ((ret < 0 && ret != -EAGAIN) || !ret)
2249                 goto out;
2250
2251         /*
2252          * Ok we had -EAGAIN which means we didn't have space to insert and
2253          * inline extent ref, so just update the reference count and add a
2254          * normal backref.
2255          */
2256         leaf = path->nodes[0];
2257         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2258         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2259         refs = btrfs_extent_refs(leaf, item);
2260         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2261         if (extent_op)
2262                 __run_delayed_extent_op(extent_op, leaf, item);
2263
2264         btrfs_mark_buffer_dirty(leaf);
2265         btrfs_release_path(path);
2266
2267         path->reada = READA_FORWARD;
2268         path->leave_spinning = 1;
2269         /* now insert the actual backref */
2270         ret = insert_extent_backref(trans, fs_info, path, bytenr, parent,
2271                                     root_objectid, owner, offset, refs_to_add);
2272         if (ret)
2273                 btrfs_abort_transaction(trans, ret);
2274 out:
2275         btrfs_free_path(path);
2276         return ret;
2277 }
2278
2279 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2280                                 struct btrfs_fs_info *fs_info,
2281                                 struct btrfs_delayed_ref_node *node,
2282                                 struct btrfs_delayed_extent_op *extent_op,
2283                                 int insert_reserved)
2284 {
2285         int ret = 0;
2286         struct btrfs_delayed_data_ref *ref;
2287         struct btrfs_key ins;
2288         u64 parent = 0;
2289         u64 ref_root = 0;
2290         u64 flags = 0;
2291
2292         ins.objectid = node->bytenr;
2293         ins.offset = node->num_bytes;
2294         ins.type = BTRFS_EXTENT_ITEM_KEY;
2295
2296         ref = btrfs_delayed_node_to_data_ref(node);
2297         trace_run_delayed_data_ref(fs_info, node, ref, node->action);
2298
2299         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2300                 parent = ref->parent;
2301         ref_root = ref->root;
2302
2303         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2304                 if (extent_op)
2305                         flags |= extent_op->flags_to_set;
2306                 ret = alloc_reserved_file_extent(trans, fs_info,
2307                                                  parent, ref_root, flags,
2308                                                  ref->objectid, ref->offset,
2309                                                  &ins, node->ref_mod);
2310         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2311                 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent,
2312                                              ref_root, ref->objectid,
2313                                              ref->offset, node->ref_mod,
2314                                              extent_op);
2315         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2316                 ret = __btrfs_free_extent(trans, fs_info, node, parent,
2317                                           ref_root, ref->objectid,
2318                                           ref->offset, node->ref_mod,
2319                                           extent_op);
2320         } else {
2321                 BUG();
2322         }
2323         return ret;
2324 }
2325
2326 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2327                                     struct extent_buffer *leaf,
2328                                     struct btrfs_extent_item *ei)
2329 {
2330         u64 flags = btrfs_extent_flags(leaf, ei);
2331         if (extent_op->update_flags) {
2332                 flags |= extent_op->flags_to_set;
2333                 btrfs_set_extent_flags(leaf, ei, flags);
2334         }
2335
2336         if (extent_op->update_key) {
2337                 struct btrfs_tree_block_info *bi;
2338                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2339                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2340                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2341         }
2342 }
2343
2344 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2345                                  struct btrfs_fs_info *fs_info,
2346                                  struct btrfs_delayed_ref_head *head,
2347                                  struct btrfs_delayed_extent_op *extent_op)
2348 {
2349         struct btrfs_key key;
2350         struct btrfs_path *path;
2351         struct btrfs_extent_item *ei;
2352         struct extent_buffer *leaf;
2353         u32 item_size;
2354         int ret;
2355         int err = 0;
2356         int metadata = !extent_op->is_data;
2357
2358         if (trans->aborted)
2359                 return 0;
2360
2361         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2362                 metadata = 0;
2363
2364         path = btrfs_alloc_path();
2365         if (!path)
2366                 return -ENOMEM;
2367
2368         key.objectid = head->bytenr;
2369
2370         if (metadata) {
2371                 key.type = BTRFS_METADATA_ITEM_KEY;
2372                 key.offset = extent_op->level;
2373         } else {
2374                 key.type = BTRFS_EXTENT_ITEM_KEY;
2375                 key.offset = head->num_bytes;
2376         }
2377
2378 again:
2379         path->reada = READA_FORWARD;
2380         path->leave_spinning = 1;
2381         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2382         if (ret < 0) {
2383                 err = ret;
2384                 goto out;
2385         }
2386         if (ret > 0) {
2387                 if (metadata) {
2388                         if (path->slots[0] > 0) {
2389                                 path->slots[0]--;
2390                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2391                                                       path->slots[0]);
2392                                 if (key.objectid == head->bytenr &&
2393                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2394                                     key.offset == head->num_bytes)
2395                                         ret = 0;
2396                         }
2397                         if (ret > 0) {
2398                                 btrfs_release_path(path);
2399                                 metadata = 0;
2400
2401                                 key.objectid = head->bytenr;
2402                                 key.offset = head->num_bytes;
2403                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2404                                 goto again;
2405                         }
2406                 } else {
2407                         err = -EIO;
2408                         goto out;
2409                 }
2410         }
2411
2412         leaf = path->nodes[0];
2413         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2414 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2415         if (item_size < sizeof(*ei)) {
2416                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2417                 if (ret < 0) {
2418                         err = ret;
2419                         goto out;
2420                 }
2421                 leaf = path->nodes[0];
2422                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2423         }
2424 #endif
2425         BUG_ON(item_size < sizeof(*ei));
2426         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2427         __run_delayed_extent_op(extent_op, leaf, ei);
2428
2429         btrfs_mark_buffer_dirty(leaf);
2430 out:
2431         btrfs_free_path(path);
2432         return err;
2433 }
2434
2435 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2436                                 struct btrfs_fs_info *fs_info,
2437                                 struct btrfs_delayed_ref_node *node,
2438                                 struct btrfs_delayed_extent_op *extent_op,
2439                                 int insert_reserved)
2440 {
2441         int ret = 0;
2442         struct btrfs_delayed_tree_ref *ref;
2443         struct btrfs_key ins;
2444         u64 parent = 0;
2445         u64 ref_root = 0;
2446         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2447
2448         ref = btrfs_delayed_node_to_tree_ref(node);
2449         trace_run_delayed_tree_ref(fs_info, node, ref, node->action);
2450
2451         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2452                 parent = ref->parent;
2453         ref_root = ref->root;
2454
2455         ins.objectid = node->bytenr;
2456         if (skinny_metadata) {
2457                 ins.offset = ref->level;
2458                 ins.type = BTRFS_METADATA_ITEM_KEY;
2459         } else {
2460                 ins.offset = node->num_bytes;
2461                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2462         }
2463
2464         if (node->ref_mod != 1) {
2465                 btrfs_err(fs_info,
2466         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2467                           node->bytenr, node->ref_mod, node->action, ref_root,
2468                           parent);
2469                 return -EIO;
2470         }
2471         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2472                 BUG_ON(!extent_op || !extent_op->update_flags);
2473                 ret = alloc_reserved_tree_block(trans, fs_info,
2474                                                 parent, ref_root,
2475                                                 extent_op->flags_to_set,
2476                                                 &extent_op->key,
2477                                                 ref->level, &ins);
2478         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2479                 ret = __btrfs_inc_extent_ref(trans, fs_info, node,
2480                                              parent, ref_root,
2481                                              ref->level, 0, 1,
2482                                              extent_op);
2483         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2484                 ret = __btrfs_free_extent(trans, fs_info, node,
2485                                           parent, ref_root,
2486                                           ref->level, 0, 1, extent_op);
2487         } else {
2488                 BUG();
2489         }
2490         return ret;
2491 }
2492
2493 /* helper function to actually process a single delayed ref entry */
2494 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2495                                struct btrfs_fs_info *fs_info,
2496                                struct btrfs_delayed_ref_node *node,
2497                                struct btrfs_delayed_extent_op *extent_op,
2498                                int insert_reserved)
2499 {
2500         int ret = 0;
2501
2502         if (trans->aborted) {
2503                 if (insert_reserved)
2504                         btrfs_pin_extent(fs_info, node->bytenr,
2505                                          node->num_bytes, 1);
2506                 return 0;
2507         }
2508
2509         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2510             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2511                 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op,
2512                                            insert_reserved);
2513         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2514                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2515                 ret = run_delayed_data_ref(trans, fs_info, node, extent_op,
2516                                            insert_reserved);
2517         else
2518                 BUG();
2519         return ret;
2520 }
2521
2522 static inline struct btrfs_delayed_ref_node *
2523 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2524 {
2525         struct btrfs_delayed_ref_node *ref;
2526
2527         if (RB_EMPTY_ROOT(&head->ref_tree))
2528                 return NULL;
2529
2530         /*
2531          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2532          * This is to prevent a ref count from going down to zero, which deletes
2533          * the extent item from the extent tree, when there still are references
2534          * to add, which would fail because they would not find the extent item.
2535          */
2536         if (!list_empty(&head->ref_add_list))
2537                 return list_first_entry(&head->ref_add_list,
2538                                 struct btrfs_delayed_ref_node, add_list);
2539
2540         ref = rb_entry(rb_first(&head->ref_tree),
2541                        struct btrfs_delayed_ref_node, ref_node);
2542         ASSERT(list_empty(&ref->add_list));
2543         return ref;
2544 }
2545
2546 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2547                                       struct btrfs_delayed_ref_head *head)
2548 {
2549         spin_lock(&delayed_refs->lock);
2550         head->processing = 0;
2551         delayed_refs->num_heads_ready++;
2552         spin_unlock(&delayed_refs->lock);
2553         btrfs_delayed_ref_unlock(head);
2554 }
2555
2556 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2557                              struct btrfs_fs_info *fs_info,
2558                              struct btrfs_delayed_ref_head *head)
2559 {
2560         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2561         int ret;
2562
2563         if (!extent_op)
2564                 return 0;
2565         head->extent_op = NULL;
2566         if (head->must_insert_reserved) {
2567                 btrfs_free_delayed_extent_op(extent_op);
2568                 return 0;
2569         }
2570         spin_unlock(&head->lock);
2571         ret = run_delayed_extent_op(trans, fs_info, head, extent_op);
2572         btrfs_free_delayed_extent_op(extent_op);
2573         return ret ? ret : 1;
2574 }
2575
2576 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2577                             struct btrfs_fs_info *fs_info,
2578                             struct btrfs_delayed_ref_head *head)
2579 {
2580         struct btrfs_delayed_ref_root *delayed_refs;
2581         int ret;
2582
2583         delayed_refs = &trans->transaction->delayed_refs;
2584
2585         ret = cleanup_extent_op(trans, fs_info, head);
2586         if (ret < 0) {
2587                 unselect_delayed_ref_head(delayed_refs, head);
2588                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2589                 return ret;
2590         } else if (ret) {
2591                 return ret;
2592         }
2593
2594         /*
2595          * Need to drop our head ref lock and re-acquire the delayed ref lock
2596          * and then re-check to make sure nobody got added.
2597          */
2598         spin_unlock(&head->lock);
2599         spin_lock(&delayed_refs->lock);
2600         spin_lock(&head->lock);
2601         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2602                 spin_unlock(&head->lock);
2603                 spin_unlock(&delayed_refs->lock);
2604                 return 1;
2605         }
2606         delayed_refs->num_heads--;
2607         rb_erase(&head->href_node, &delayed_refs->href_root);
2608         RB_CLEAR_NODE(&head->href_node);
2609         spin_unlock(&delayed_refs->lock);
2610         spin_unlock(&head->lock);
2611         atomic_dec(&delayed_refs->num_entries);
2612
2613         trace_run_delayed_ref_head(fs_info, head, 0);
2614
2615         if (head->total_ref_mod < 0) {
2616                 struct btrfs_block_group_cache *cache;
2617
2618                 cache = btrfs_lookup_block_group(fs_info, head->bytenr);
2619                 ASSERT(cache);
2620                 percpu_counter_add(&cache->space_info->total_bytes_pinned,
2621                                    -head->num_bytes);
2622                 btrfs_put_block_group(cache);
2623
2624                 if (head->is_data) {
2625                         spin_lock(&delayed_refs->lock);
2626                         delayed_refs->pending_csums -= head->num_bytes;
2627                         spin_unlock(&delayed_refs->lock);
2628                 }
2629         }
2630
2631         if (head->must_insert_reserved) {
2632                 btrfs_pin_extent(fs_info, head->bytenr,
2633                                  head->num_bytes, 1);
2634                 if (head->is_data) {
2635                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2636                                               head->num_bytes);
2637                 }
2638         }
2639
2640         /* Also free its reserved qgroup space */
2641         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2642                                       head->qgroup_reserved);
2643         btrfs_delayed_ref_unlock(head);
2644         btrfs_put_delayed_ref_head(head);
2645         return 0;
2646 }
2647
2648 /*
2649  * Returns 0 on success or if called with an already aborted transaction.
2650  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2651  */
2652 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2653                                              struct btrfs_fs_info *fs_info,
2654                                              unsigned long nr)
2655 {
2656         struct btrfs_delayed_ref_root *delayed_refs;
2657         struct btrfs_delayed_ref_node *ref;
2658         struct btrfs_delayed_ref_head *locked_ref = NULL;
2659         struct btrfs_delayed_extent_op *extent_op;
2660         ktime_t start = ktime_get();
2661         int ret;
2662         unsigned long count = 0;
2663         unsigned long actual_count = 0;
2664         int must_insert_reserved = 0;
2665
2666         delayed_refs = &trans->transaction->delayed_refs;
2667         while (1) {
2668                 if (!locked_ref) {
2669                         if (count >= nr)
2670                                 break;
2671
2672                         spin_lock(&delayed_refs->lock);
2673                         locked_ref = btrfs_select_ref_head(trans);
2674                         if (!locked_ref) {
2675                                 spin_unlock(&delayed_refs->lock);
2676                                 break;
2677                         }
2678
2679                         /* grab the lock that says we are going to process
2680                          * all the refs for this head */
2681                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2682                         spin_unlock(&delayed_refs->lock);
2683                         /*
2684                          * we may have dropped the spin lock to get the head
2685                          * mutex lock, and that might have given someone else
2686                          * time to free the head.  If that's true, it has been
2687                          * removed from our list and we can move on.
2688                          */
2689                         if (ret == -EAGAIN) {
2690                                 locked_ref = NULL;
2691                                 count++;
2692                                 continue;
2693                         }
2694                 }
2695
2696                 /*
2697                  * We need to try and merge add/drops of the same ref since we
2698                  * can run into issues with relocate dropping the implicit ref
2699                  * and then it being added back again before the drop can
2700                  * finish.  If we merged anything we need to re-loop so we can
2701                  * get a good ref.
2702                  * Or we can get node references of the same type that weren't
2703                  * merged when created due to bumps in the tree mod seq, and
2704                  * we need to merge them to prevent adding an inline extent
2705                  * backref before dropping it (triggering a BUG_ON at
2706                  * insert_inline_extent_backref()).
2707                  */
2708                 spin_lock(&locked_ref->lock);
2709                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2710                                          locked_ref);
2711
2712                 /*
2713                  * locked_ref is the head node, so we have to go one
2714                  * node back for any delayed ref updates
2715                  */
2716                 ref = select_delayed_ref(locked_ref);
2717
2718                 if (ref && ref->seq &&
2719                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2720                         spin_unlock(&locked_ref->lock);
2721                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2722                         locked_ref = NULL;
2723                         cond_resched();
2724                         count++;
2725                         continue;
2726                 }
2727
2728                 /*
2729                  * We're done processing refs in this ref_head, clean everything
2730                  * up and move on to the next ref_head.
2731                  */
2732                 if (!ref) {
2733                         ret = cleanup_ref_head(trans, fs_info, locked_ref);
2734                         if (ret > 0 ) {
2735                                 /* We dropped our lock, we need to loop. */
2736                                 ret = 0;
2737                                 continue;
2738                         } else if (ret) {
2739                                 return ret;
2740                         }
2741                         locked_ref = NULL;
2742                         count++;
2743                         continue;
2744                 }
2745
2746                 actual_count++;
2747                 ref->in_tree = 0;
2748                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2749                 RB_CLEAR_NODE(&ref->ref_node);
2750                 if (!list_empty(&ref->add_list))
2751                         list_del(&ref->add_list);
2752                 /*
2753                  * When we play the delayed ref, also correct the ref_mod on
2754                  * head
2755                  */
2756                 switch (ref->action) {
2757                 case BTRFS_ADD_DELAYED_REF:
2758                 case BTRFS_ADD_DELAYED_EXTENT:
2759                         locked_ref->ref_mod -= ref->ref_mod;
2760                         break;
2761                 case BTRFS_DROP_DELAYED_REF:
2762                         locked_ref->ref_mod += ref->ref_mod;
2763                         break;
2764                 default:
2765                         WARN_ON(1);
2766                 }
2767                 atomic_dec(&delayed_refs->num_entries);
2768
2769                 /*
2770                  * Record the must-insert_reserved flag before we drop the spin
2771                  * lock.
2772                  */
2773                 must_insert_reserved = locked_ref->must_insert_reserved;
2774                 locked_ref->must_insert_reserved = 0;
2775
2776                 extent_op = locked_ref->extent_op;
2777                 locked_ref->extent_op = NULL;
2778                 spin_unlock(&locked_ref->lock);
2779
2780                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2781                                           must_insert_reserved);
2782
2783                 btrfs_free_delayed_extent_op(extent_op);
2784                 if (ret) {
2785                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2786                         btrfs_put_delayed_ref(ref);
2787                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2788                                     ret);
2789                         return ret;
2790                 }
2791
2792                 btrfs_put_delayed_ref(ref);
2793                 count++;
2794                 cond_resched();
2795         }
2796
2797         /*
2798          * We don't want to include ref heads since we can have empty ref heads
2799          * and those will drastically skew our runtime down since we just do
2800          * accounting, no actual extent tree updates.
2801          */
2802         if (actual_count > 0) {
2803                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2804                 u64 avg;
2805
2806                 /*
2807                  * We weigh the current average higher than our current runtime
2808                  * to avoid large swings in the average.
2809                  */
2810                 spin_lock(&delayed_refs->lock);
2811                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2812                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2813                 spin_unlock(&delayed_refs->lock);
2814         }
2815         return 0;
2816 }
2817
2818 #ifdef SCRAMBLE_DELAYED_REFS
2819 /*
2820  * Normally delayed refs get processed in ascending bytenr order. This
2821  * correlates in most cases to the order added. To expose dependencies on this
2822  * order, we start to process the tree in the middle instead of the beginning
2823  */
2824 static u64 find_middle(struct rb_root *root)
2825 {
2826         struct rb_node *n = root->rb_node;
2827         struct btrfs_delayed_ref_node *entry;
2828         int alt = 1;
2829         u64 middle;
2830         u64 first = 0, last = 0;
2831
2832         n = rb_first(root);
2833         if (n) {
2834                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2835                 first = entry->bytenr;
2836         }
2837         n = rb_last(root);
2838         if (n) {
2839                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2840                 last = entry->bytenr;
2841         }
2842         n = root->rb_node;
2843
2844         while (n) {
2845                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2846                 WARN_ON(!entry->in_tree);
2847
2848                 middle = entry->bytenr;
2849
2850                 if (alt)
2851                         n = n->rb_left;
2852                 else
2853                         n = n->rb_right;
2854
2855                 alt = 1 - alt;
2856         }
2857         return middle;
2858 }
2859 #endif
2860
2861 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2862 {
2863         u64 num_bytes;
2864
2865         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2866                              sizeof(struct btrfs_extent_inline_ref));
2867         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2868                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2869
2870         /*
2871          * We don't ever fill up leaves all the way so multiply by 2 just to be
2872          * closer to what we're really going to want to use.
2873          */
2874         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2875 }
2876
2877 /*
2878  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2879  * would require to store the csums for that many bytes.
2880  */
2881 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2882 {
2883         u64 csum_size;
2884         u64 num_csums_per_leaf;
2885         u64 num_csums;
2886
2887         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2888         num_csums_per_leaf = div64_u64(csum_size,
2889                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2890         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2891         num_csums += num_csums_per_leaf - 1;
2892         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2893         return num_csums;
2894 }
2895
2896 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2897                                        struct btrfs_fs_info *fs_info)
2898 {
2899         struct btrfs_block_rsv *global_rsv;
2900         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2901         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2902         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2903         u64 num_bytes, num_dirty_bgs_bytes;
2904         int ret = 0;
2905
2906         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2907         num_heads = heads_to_leaves(fs_info, num_heads);
2908         if (num_heads > 1)
2909                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2910         num_bytes <<= 1;
2911         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2912                                                         fs_info->nodesize;
2913         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2914                                                              num_dirty_bgs);
2915         global_rsv = &fs_info->global_block_rsv;
2916
2917         /*
2918          * If we can't allocate any more chunks lets make sure we have _lots_ of
2919          * wiggle room since running delayed refs can create more delayed refs.
2920          */
2921         if (global_rsv->space_info->full) {
2922                 num_dirty_bgs_bytes <<= 1;
2923                 num_bytes <<= 1;
2924         }
2925
2926         spin_lock(&global_rsv->lock);
2927         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2928                 ret = 1;
2929         spin_unlock(&global_rsv->lock);
2930         return ret;
2931 }
2932
2933 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2934                                        struct btrfs_fs_info *fs_info)
2935 {
2936         u64 num_entries =
2937                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2938         u64 avg_runtime;
2939         u64 val;
2940
2941         smp_mb();
2942         avg_runtime = fs_info->avg_delayed_ref_runtime;
2943         val = num_entries * avg_runtime;
2944         if (val >= NSEC_PER_SEC)
2945                 return 1;
2946         if (val >= NSEC_PER_SEC / 2)
2947                 return 2;
2948
2949         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2950 }
2951
2952 struct async_delayed_refs {
2953         struct btrfs_root *root;
2954         u64 transid;
2955         int count;
2956         int error;
2957         int sync;
2958         struct completion wait;
2959         struct btrfs_work work;
2960 };
2961
2962 static inline struct async_delayed_refs *
2963 to_async_delayed_refs(struct btrfs_work *work)
2964 {
2965         return container_of(work, struct async_delayed_refs, work);
2966 }
2967
2968 static void delayed_ref_async_start(struct btrfs_work *work)
2969 {
2970         struct async_delayed_refs *async = to_async_delayed_refs(work);
2971         struct btrfs_trans_handle *trans;
2972         struct btrfs_fs_info *fs_info = async->root->fs_info;
2973         int ret;
2974
2975         /* if the commit is already started, we don't need to wait here */
2976         if (btrfs_transaction_blocked(fs_info))
2977                 goto done;
2978
2979         trans = btrfs_join_transaction(async->root);
2980         if (IS_ERR(trans)) {
2981                 async->error = PTR_ERR(trans);
2982                 goto done;
2983         }
2984
2985         /*
2986          * trans->sync means that when we call end_transaction, we won't
2987          * wait on delayed refs
2988          */
2989         trans->sync = true;
2990
2991         /* Don't bother flushing if we got into a different transaction */
2992         if (trans->transid > async->transid)
2993                 goto end;
2994
2995         ret = btrfs_run_delayed_refs(trans, fs_info, async->count);
2996         if (ret)
2997                 async->error = ret;
2998 end:
2999         ret = btrfs_end_transaction(trans);
3000         if (ret && !async->error)
3001                 async->error = ret;
3002 done:
3003         if (async->sync)
3004                 complete(&async->wait);
3005         else
3006                 kfree(async);
3007 }
3008
3009 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
3010                                  unsigned long count, u64 transid, int wait)
3011 {
3012         struct async_delayed_refs *async;
3013         int ret;
3014
3015         async = kmalloc(sizeof(*async), GFP_NOFS);
3016         if (!async)
3017                 return -ENOMEM;
3018
3019         async->root = fs_info->tree_root;
3020         async->count = count;
3021         async->error = 0;
3022         async->transid = transid;
3023         if (wait)
3024                 async->sync = 1;
3025         else
3026                 async->sync = 0;
3027         init_completion(&async->wait);
3028
3029         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
3030                         delayed_ref_async_start, NULL, NULL);
3031
3032         btrfs_queue_work(fs_info->extent_workers, &async->work);
3033
3034         if (wait) {
3035                 wait_for_completion(&async->wait);
3036                 ret = async->error;
3037                 kfree(async);
3038                 return ret;
3039         }
3040         return 0;
3041 }
3042
3043 /*
3044  * this starts processing the delayed reference count updates and
3045  * extent insertions we have queued up so far.  count can be
3046  * 0, which means to process everything in the tree at the start
3047  * of the run (but not newly added entries), or it can be some target
3048  * number you'd like to process.
3049  *
3050  * Returns 0 on success or if called with an aborted transaction
3051  * Returns <0 on error and aborts the transaction
3052  */
3053 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
3054                            struct btrfs_fs_info *fs_info, unsigned long count)
3055 {
3056         struct rb_node *node;
3057         struct btrfs_delayed_ref_root *delayed_refs;
3058         struct btrfs_delayed_ref_head *head;
3059         int ret;
3060         int run_all = count == (unsigned long)-1;
3061         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
3062
3063         /* We'll clean this up in btrfs_cleanup_transaction */
3064         if (trans->aborted)
3065                 return 0;
3066
3067         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
3068                 return 0;
3069
3070         delayed_refs = &trans->transaction->delayed_refs;
3071         if (count == 0)
3072                 count = atomic_read(&delayed_refs->num_entries) * 2;
3073
3074 again:
3075 #ifdef SCRAMBLE_DELAYED_REFS
3076         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3077 #endif
3078         trans->can_flush_pending_bgs = false;
3079         ret = __btrfs_run_delayed_refs(trans, fs_info, count);
3080         if (ret < 0) {
3081                 btrfs_abort_transaction(trans, ret);
3082                 return ret;
3083         }
3084
3085         if (run_all) {
3086                 if (!list_empty(&trans->new_bgs))
3087                         btrfs_create_pending_block_groups(trans);
3088
3089                 spin_lock(&delayed_refs->lock);
3090                 node = rb_first(&delayed_refs->href_root);
3091                 if (!node) {
3092                         spin_unlock(&delayed_refs->lock);
3093                         goto out;
3094                 }
3095                 head = rb_entry(node, struct btrfs_delayed_ref_head,
3096                                 href_node);
3097                 refcount_inc(&head->refs);
3098                 spin_unlock(&delayed_refs->lock);
3099
3100                 /* Mutex was contended, block until it's released and retry. */
3101                 mutex_lock(&head->mutex);
3102                 mutex_unlock(&head->mutex);
3103
3104                 btrfs_put_delayed_ref_head(head);
3105                 cond_resched();
3106                 goto again;
3107         }
3108 out:
3109         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3110         return 0;
3111 }
3112
3113 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3114                                 struct btrfs_fs_info *fs_info,
3115                                 u64 bytenr, u64 num_bytes, u64 flags,
3116                                 int level, int is_data)
3117 {
3118         struct btrfs_delayed_extent_op *extent_op;
3119         int ret;
3120
3121         extent_op = btrfs_alloc_delayed_extent_op();
3122         if (!extent_op)
3123                 return -ENOMEM;
3124
3125         extent_op->flags_to_set = flags;
3126         extent_op->update_flags = true;
3127         extent_op->update_key = false;
3128         extent_op->is_data = is_data ? true : false;
3129         extent_op->level = level;
3130
3131         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3132                                           num_bytes, extent_op);
3133         if (ret)
3134                 btrfs_free_delayed_extent_op(extent_op);
3135         return ret;
3136 }
3137
3138 static noinline int check_delayed_ref(struct btrfs_root *root,
3139                                       struct btrfs_path *path,
3140                                       u64 objectid, u64 offset, u64 bytenr)
3141 {
3142         struct btrfs_delayed_ref_head *head;
3143         struct btrfs_delayed_ref_node *ref;
3144         struct btrfs_delayed_data_ref *data_ref;
3145         struct btrfs_delayed_ref_root *delayed_refs;
3146         struct btrfs_transaction *cur_trans;
3147         struct rb_node *node;
3148         int ret = 0;
3149
3150         cur_trans = root->fs_info->running_transaction;
3151         if (!cur_trans)
3152                 return 0;
3153
3154         delayed_refs = &cur_trans->delayed_refs;
3155         spin_lock(&delayed_refs->lock);
3156         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3157         if (!head) {
3158                 spin_unlock(&delayed_refs->lock);
3159                 return 0;
3160         }
3161
3162         if (!mutex_trylock(&head->mutex)) {
3163                 refcount_inc(&head->refs);
3164                 spin_unlock(&delayed_refs->lock);
3165
3166                 btrfs_release_path(path);
3167
3168                 /*
3169                  * Mutex was contended, block until it's released and let
3170                  * caller try again
3171                  */
3172                 mutex_lock(&head->mutex);
3173                 mutex_unlock(&head->mutex);
3174                 btrfs_put_delayed_ref_head(head);
3175                 return -EAGAIN;
3176         }
3177         spin_unlock(&delayed_refs->lock);
3178
3179         spin_lock(&head->lock);
3180         /*
3181          * XXX: We should replace this with a proper search function in the
3182          * future.
3183          */
3184         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
3185                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3186                 /* If it's a shared ref we know a cross reference exists */
3187                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3188                         ret = 1;
3189                         break;
3190                 }
3191
3192                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3193
3194                 /*
3195                  * If our ref doesn't match the one we're currently looking at
3196                  * then we have a cross reference.
3197                  */
3198                 if (data_ref->root != root->root_key.objectid ||
3199                     data_ref->objectid != objectid ||
3200                     data_ref->offset != offset) {
3201                         ret = 1;
3202                         break;
3203                 }
3204         }
3205         spin_unlock(&head->lock);
3206         mutex_unlock(&head->mutex);
3207         return ret;
3208 }
3209
3210 static noinline int check_committed_ref(struct btrfs_root *root,
3211                                         struct btrfs_path *path,
3212                                         u64 objectid, u64 offset, u64 bytenr)
3213 {
3214         struct btrfs_fs_info *fs_info = root->fs_info;
3215         struct btrfs_root *extent_root = fs_info->extent_root;
3216         struct extent_buffer *leaf;
3217         struct btrfs_extent_data_ref *ref;
3218         struct btrfs_extent_inline_ref *iref;
3219         struct btrfs_extent_item *ei;
3220         struct btrfs_key key;
3221         u32 item_size;
3222         int type;
3223         int ret;
3224
3225         key.objectid = bytenr;
3226         key.offset = (u64)-1;
3227         key.type = BTRFS_EXTENT_ITEM_KEY;
3228
3229         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3230         if (ret < 0)
3231                 goto out;
3232         BUG_ON(ret == 0); /* Corruption */
3233
3234         ret = -ENOENT;
3235         if (path->slots[0] == 0)
3236                 goto out;
3237
3238         path->slots[0]--;
3239         leaf = path->nodes[0];
3240         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3241
3242         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3243                 goto out;
3244
3245         ret = 1;
3246         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3247 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3248         if (item_size < sizeof(*ei)) {
3249                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3250                 goto out;
3251         }
3252 #endif
3253         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3254
3255         if (item_size != sizeof(*ei) +
3256             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3257                 goto out;
3258
3259         if (btrfs_extent_generation(leaf, ei) <=
3260             btrfs_root_last_snapshot(&root->root_item))
3261                 goto out;
3262
3263         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3264
3265         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3266         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3267                 goto out;
3268
3269         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3270         if (btrfs_extent_refs(leaf, ei) !=
3271             btrfs_extent_data_ref_count(leaf, ref) ||
3272             btrfs_extent_data_ref_root(leaf, ref) !=
3273             root->root_key.objectid ||
3274             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3275             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3276                 goto out;
3277
3278         ret = 0;
3279 out:
3280         return ret;
3281 }
3282
3283 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3284                           u64 bytenr)
3285 {
3286         struct btrfs_path *path;
3287         int ret;
3288         int ret2;
3289
3290         path = btrfs_alloc_path();
3291         if (!path)
3292                 return -ENOENT;
3293
3294         do {
3295                 ret = check_committed_ref(root, path, objectid,
3296                                           offset, bytenr);
3297                 if (ret && ret != -ENOENT)
3298                         goto out;
3299
3300                 ret2 = check_delayed_ref(root, path, objectid,
3301                                          offset, bytenr);
3302         } while (ret2 == -EAGAIN);
3303
3304         if (ret2 && ret2 != -ENOENT) {
3305                 ret = ret2;
3306                 goto out;
3307         }
3308
3309         if (ret != -ENOENT || ret2 != -ENOENT)
3310                 ret = 0;
3311 out:
3312         btrfs_free_path(path);
3313         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3314                 WARN_ON(ret > 0);
3315         return ret;
3316 }
3317
3318 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3319                            struct btrfs_root *root,
3320                            struct extent_buffer *buf,
3321                            int full_backref, int inc)
3322 {
3323         struct btrfs_fs_info *fs_info = root->fs_info;
3324         u64 bytenr;
3325         u64 num_bytes;
3326         u64 parent;
3327         u64 ref_root;
3328         u32 nritems;
3329         struct btrfs_key key;
3330         struct btrfs_file_extent_item *fi;
3331         int i;
3332         int level;
3333         int ret = 0;
3334         int (*process_func)(struct btrfs_trans_handle *,
3335                             struct btrfs_root *,
3336                             u64, u64, u64, u64, u64, u64);
3337
3338
3339         if (btrfs_is_testing(fs_info))
3340                 return 0;
3341
3342         ref_root = btrfs_header_owner(buf);
3343         nritems = btrfs_header_nritems(buf);
3344         level = btrfs_header_level(buf);
3345
3346         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3347                 return 0;
3348
3349         if (inc)
3350                 process_func = btrfs_inc_extent_ref;
3351         else
3352                 process_func = btrfs_free_extent;
3353
3354         if (full_backref)
3355                 parent = buf->start;
3356         else
3357                 parent = 0;
3358
3359         for (i = 0; i < nritems; i++) {
3360                 if (level == 0) {
3361                         btrfs_item_key_to_cpu(buf, &key, i);
3362                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3363                                 continue;
3364                         fi = btrfs_item_ptr(buf, i,
3365                                             struct btrfs_file_extent_item);
3366                         if (btrfs_file_extent_type(buf, fi) ==
3367                             BTRFS_FILE_EXTENT_INLINE)
3368                                 continue;
3369                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3370                         if (bytenr == 0)
3371                                 continue;
3372
3373                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3374                         key.offset -= btrfs_file_extent_offset(buf, fi);
3375                         ret = process_func(trans, root, bytenr, num_bytes,
3376                                            parent, ref_root, key.objectid,
3377                                            key.offset);
3378                         if (ret)
3379                                 goto fail;
3380                 } else {
3381                         bytenr = btrfs_node_blockptr(buf, i);
3382                         num_bytes = fs_info->nodesize;
3383                         ret = process_func(trans, root, bytenr, num_bytes,
3384                                            parent, ref_root, level - 1, 0);
3385                         if (ret)
3386                                 goto fail;
3387                 }
3388         }
3389         return 0;
3390 fail:
3391         return ret;
3392 }
3393
3394 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3395                   struct extent_buffer *buf, int full_backref)
3396 {
3397         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3398 }
3399
3400 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3401                   struct extent_buffer *buf, int full_backref)
3402 {
3403         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3404 }
3405
3406 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3407                                  struct btrfs_fs_info *fs_info,
3408                                  struct btrfs_path *path,
3409                                  struct btrfs_block_group_cache *cache)
3410 {
3411         int ret;
3412         struct btrfs_root *extent_root = fs_info->extent_root;
3413         unsigned long bi;
3414         struct extent_buffer *leaf;
3415
3416         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3417         if (ret) {
3418                 if (ret > 0)
3419                         ret = -ENOENT;
3420                 goto fail;
3421         }
3422
3423         leaf = path->nodes[0];
3424         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3425         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3426         btrfs_mark_buffer_dirty(leaf);
3427 fail:
3428         btrfs_release_path(path);
3429         return ret;
3430
3431 }
3432
3433 static struct btrfs_block_group_cache *
3434 next_block_group(struct btrfs_fs_info *fs_info,
3435                  struct btrfs_block_group_cache *cache)
3436 {
3437         struct rb_node *node;
3438
3439         spin_lock(&fs_info->block_group_cache_lock);
3440
3441         /* If our block group was removed, we need a full search. */
3442         if (RB_EMPTY_NODE(&cache->cache_node)) {
3443                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3444
3445                 spin_unlock(&fs_info->block_group_cache_lock);
3446                 btrfs_put_block_group(cache);
3447                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3448         }
3449         node = rb_next(&cache->cache_node);
3450         btrfs_put_block_group(cache);
3451         if (node) {
3452                 cache = rb_entry(node, struct btrfs_block_group_cache,
3453                                  cache_node);
3454                 btrfs_get_block_group(cache);
3455         } else
3456                 cache = NULL;
3457         spin_unlock(&fs_info->block_group_cache_lock);
3458         return cache;
3459 }
3460
3461 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3462                             struct btrfs_trans_handle *trans,
3463                             struct btrfs_path *path)
3464 {
3465         struct btrfs_fs_info *fs_info = block_group->fs_info;
3466         struct btrfs_root *root = fs_info->tree_root;
3467         struct inode *inode = NULL;
3468         struct extent_changeset *data_reserved = NULL;
3469         u64 alloc_hint = 0;
3470         int dcs = BTRFS_DC_ERROR;
3471         u64 num_pages = 0;
3472         int retries = 0;
3473         int ret = 0;
3474
3475         /*
3476          * If this block group is smaller than 100 megs don't bother caching the
3477          * block group.
3478          */
3479         if (block_group->key.offset < (100 * SZ_1M)) {
3480                 spin_lock(&block_group->lock);
3481                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3482                 spin_unlock(&block_group->lock);
3483                 return 0;
3484         }
3485
3486         if (trans->aborted)
3487                 return 0;
3488 again:
3489         inode = lookup_free_space_inode(fs_info, block_group, path);
3490         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3491                 ret = PTR_ERR(inode);
3492                 btrfs_release_path(path);
3493                 goto out;
3494         }
3495
3496         if (IS_ERR(inode)) {
3497                 BUG_ON(retries);
3498                 retries++;
3499
3500                 if (block_group->ro)
3501                         goto out_free;
3502
3503                 ret = create_free_space_inode(fs_info, trans, block_group,
3504                                               path);
3505                 if (ret)
3506                         goto out_free;
3507                 goto again;
3508         }
3509
3510         /*
3511          * We want to set the generation to 0, that way if anything goes wrong
3512          * from here on out we know not to trust this cache when we load up next
3513          * time.
3514          */
3515         BTRFS_I(inode)->generation = 0;
3516         ret = btrfs_update_inode(trans, root, inode);
3517         if (ret) {
3518                 /*
3519                  * So theoretically we could recover from this, simply set the
3520                  * super cache generation to 0 so we know to invalidate the
3521                  * cache, but then we'd have to keep track of the block groups
3522                  * that fail this way so we know we _have_ to reset this cache
3523                  * before the next commit or risk reading stale cache.  So to
3524                  * limit our exposure to horrible edge cases lets just abort the
3525                  * transaction, this only happens in really bad situations
3526                  * anyway.
3527                  */
3528                 btrfs_abort_transaction(trans, ret);
3529                 goto out_put;
3530         }
3531         WARN_ON(ret);
3532
3533         /* We've already setup this transaction, go ahead and exit */
3534         if (block_group->cache_generation == trans->transid &&
3535             i_size_read(inode)) {
3536                 dcs = BTRFS_DC_SETUP;
3537                 goto out_put;
3538         }
3539
3540         if (i_size_read(inode) > 0) {
3541                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3542                                         &fs_info->global_block_rsv);
3543                 if (ret)
3544                         goto out_put;
3545
3546                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3547                 if (ret)
3548                         goto out_put;
3549         }
3550
3551         spin_lock(&block_group->lock);
3552         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3553             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3554                 /*
3555                  * don't bother trying to write stuff out _if_
3556                  * a) we're not cached,
3557                  * b) we're with nospace_cache mount option,
3558                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3559                  */
3560                 dcs = BTRFS_DC_WRITTEN;
3561                 spin_unlock(&block_group->lock);
3562                 goto out_put;
3563         }
3564         spin_unlock(&block_group->lock);
3565
3566         /*
3567          * We hit an ENOSPC when setting up the cache in this transaction, just
3568          * skip doing the setup, we've already cleared the cache so we're safe.
3569          */
3570         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3571                 ret = -ENOSPC;
3572                 goto out_put;
3573         }
3574
3575         /*
3576          * Try to preallocate enough space based on how big the block group is.
3577          * Keep in mind this has to include any pinned space which could end up
3578          * taking up quite a bit since it's not folded into the other space
3579          * cache.
3580          */
3581         num_pages = div_u64(block_group->key.offset, SZ_256M);
3582         if (!num_pages)
3583                 num_pages = 1;
3584
3585         num_pages *= 16;
3586         num_pages *= PAGE_SIZE;
3587
3588         ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3589         if (ret)
3590                 goto out_put;
3591
3592         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3593                                               num_pages, num_pages,
3594                                               &alloc_hint);
3595         /*
3596          * Our cache requires contiguous chunks so that we don't modify a bunch
3597          * of metadata or split extents when writing the cache out, which means
3598          * we can enospc if we are heavily fragmented in addition to just normal
3599          * out of space conditions.  So if we hit this just skip setting up any
3600          * other block groups for this transaction, maybe we'll unpin enough
3601          * space the next time around.
3602          */
3603         if (!ret)
3604                 dcs = BTRFS_DC_SETUP;
3605         else if (ret == -ENOSPC)
3606                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3607
3608 out_put:
3609         iput(inode);
3610 out_free:
3611         btrfs_release_path(path);
3612 out:
3613         spin_lock(&block_group->lock);
3614         if (!ret && dcs == BTRFS_DC_SETUP)
3615                 block_group->cache_generation = trans->transid;
3616         block_group->disk_cache_state = dcs;
3617         spin_unlock(&block_group->lock);
3618
3619         extent_changeset_free(data_reserved);
3620         return ret;
3621 }
3622
3623 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3624                             struct btrfs_fs_info *fs_info)
3625 {
3626         struct btrfs_block_group_cache *cache, *tmp;
3627         struct btrfs_transaction *cur_trans = trans->transaction;
3628         struct btrfs_path *path;
3629
3630         if (list_empty(&cur_trans->dirty_bgs) ||
3631             !btrfs_test_opt(fs_info, SPACE_CACHE))
3632                 return 0;
3633
3634         path = btrfs_alloc_path();
3635         if (!path)
3636                 return -ENOMEM;
3637
3638         /* Could add new block groups, use _safe just in case */
3639         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3640                                  dirty_list) {
3641                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3642                         cache_save_setup(cache, trans, path);
3643         }
3644
3645         btrfs_free_path(path);
3646         return 0;
3647 }
3648
3649 /*
3650  * transaction commit does final block group cache writeback during a
3651  * critical section where nothing is allowed to change the FS.  This is
3652  * required in order for the cache to actually match the block group,
3653  * but can introduce a lot of latency into the commit.
3654  *
3655  * So, btrfs_start_dirty_block_groups is here to kick off block group
3656  * cache IO.  There's a chance we'll have to redo some of it if the
3657  * block group changes again during the commit, but it greatly reduces
3658  * the commit latency by getting rid of the easy block groups while
3659  * we're still allowing others to join the commit.
3660  */
3661 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3662 {
3663         struct btrfs_fs_info *fs_info = trans->fs_info;
3664         struct btrfs_block_group_cache *cache;
3665         struct btrfs_transaction *cur_trans = trans->transaction;
3666         int ret = 0;
3667         int should_put;
3668         struct btrfs_path *path = NULL;
3669         LIST_HEAD(dirty);
3670         struct list_head *io = &cur_trans->io_bgs;
3671         int num_started = 0;
3672         int loops = 0;
3673
3674         spin_lock(&cur_trans->dirty_bgs_lock);
3675         if (list_empty(&cur_trans->dirty_bgs)) {
3676                 spin_unlock(&cur_trans->dirty_bgs_lock);
3677                 return 0;
3678         }
3679         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3680         spin_unlock(&cur_trans->dirty_bgs_lock);
3681
3682 again:
3683         /*
3684          * make sure all the block groups on our dirty list actually
3685          * exist
3686          */
3687         btrfs_create_pending_block_groups(trans);
3688
3689         if (!path) {
3690                 path = btrfs_alloc_path();
3691                 if (!path)
3692                         return -ENOMEM;
3693         }
3694
3695         /*
3696          * cache_write_mutex is here only to save us from balance or automatic
3697          * removal of empty block groups deleting this block group while we are
3698          * writing out the cache
3699          */
3700         mutex_lock(&trans->transaction->cache_write_mutex);
3701         while (!list_empty(&dirty)) {
3702                 cache = list_first_entry(&dirty,
3703                                          struct btrfs_block_group_cache,
3704                                          dirty_list);
3705                 /*
3706                  * this can happen if something re-dirties a block
3707                  * group that is already under IO.  Just wait for it to
3708                  * finish and then do it all again
3709                  */
3710                 if (!list_empty(&cache->io_list)) {
3711                         list_del_init(&cache->io_list);
3712                         btrfs_wait_cache_io(trans, cache, path);
3713                         btrfs_put_block_group(cache);
3714                 }
3715
3716
3717                 /*
3718                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3719                  * if it should update the cache_state.  Don't delete
3720                  * until after we wait.
3721                  *
3722                  * Since we're not running in the commit critical section
3723                  * we need the dirty_bgs_lock to protect from update_block_group
3724                  */
3725                 spin_lock(&cur_trans->dirty_bgs_lock);
3726                 list_del_init(&cache->dirty_list);
3727                 spin_unlock(&cur_trans->dirty_bgs_lock);
3728
3729                 should_put = 1;
3730
3731                 cache_save_setup(cache, trans, path);
3732
3733                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3734                         cache->io_ctl.inode = NULL;
3735                         ret = btrfs_write_out_cache(fs_info, trans,
3736                                                     cache, path);
3737                         if (ret == 0 && cache->io_ctl.inode) {
3738                                 num_started++;
3739                                 should_put = 0;
3740
3741                                 /*
3742                                  * The cache_write_mutex is protecting the
3743                                  * io_list, also refer to the definition of
3744                                  * btrfs_transaction::io_bgs for more details
3745                                  */
3746                                 list_add_tail(&cache->io_list, io);
3747                         } else {
3748                                 /*
3749                                  * if we failed to write the cache, the
3750                                  * generation will be bad and life goes on
3751                                  */
3752                                 ret = 0;
3753                         }
3754                 }
3755                 if (!ret) {
3756                         ret = write_one_cache_group(trans, fs_info,
3757                                                     path, cache);
3758                         /*
3759                          * Our block group might still be attached to the list
3760                          * of new block groups in the transaction handle of some
3761                          * other task (struct btrfs_trans_handle->new_bgs). This
3762                          * means its block group item isn't yet in the extent
3763                          * tree. If this happens ignore the error, as we will
3764                          * try again later in the critical section of the
3765                          * transaction commit.
3766                          */
3767                         if (ret == -ENOENT) {
3768                                 ret = 0;
3769                                 spin_lock(&cur_trans->dirty_bgs_lock);
3770                                 if (list_empty(&cache->dirty_list)) {
3771                                         list_add_tail(&cache->dirty_list,
3772                                                       &cur_trans->dirty_bgs);
3773                                         btrfs_get_block_group(cache);
3774                                 }
3775                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3776                         } else if (ret) {
3777                                 btrfs_abort_transaction(trans, ret);
3778                         }
3779                 }
3780
3781                 /* if its not on the io list, we need to put the block group */
3782                 if (should_put)
3783                         btrfs_put_block_group(cache);
3784
3785                 if (ret)
3786                         break;
3787
3788                 /*
3789                  * Avoid blocking other tasks for too long. It might even save
3790                  * us from writing caches for block groups that are going to be
3791                  * removed.
3792                  */
3793                 mutex_unlock(&trans->transaction->cache_write_mutex);
3794                 mutex_lock(&trans->transaction->cache_write_mutex);
3795         }
3796         mutex_unlock(&trans->transaction->cache_write_mutex);
3797
3798         /*
3799          * go through delayed refs for all the stuff we've just kicked off
3800          * and then loop back (just once)
3801          */
3802         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
3803         if (!ret && loops == 0) {
3804                 loops++;
3805                 spin_lock(&cur_trans->dirty_bgs_lock);
3806                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3807                 /*
3808                  * dirty_bgs_lock protects us from concurrent block group
3809                  * deletes too (not just cache_write_mutex).
3810                  */
3811                 if (!list_empty(&dirty)) {
3812                         spin_unlock(&cur_trans->dirty_bgs_lock);
3813                         goto again;
3814                 }
3815                 spin_unlock(&cur_trans->dirty_bgs_lock);
3816         } else if (ret < 0) {
3817                 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3818         }
3819
3820         btrfs_free_path(path);
3821         return ret;
3822 }
3823
3824 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3825                                    struct btrfs_fs_info *fs_info)
3826 {
3827         struct btrfs_block_group_cache *cache;
3828         struct btrfs_transaction *cur_trans = trans->transaction;
3829         int ret = 0;
3830         int should_put;
3831         struct btrfs_path *path;
3832         struct list_head *io = &cur_trans->io_bgs;
3833         int num_started = 0;
3834
3835         path = btrfs_alloc_path();
3836         if (!path)
3837                 return -ENOMEM;
3838
3839         /*
3840          * Even though we are in the critical section of the transaction commit,
3841          * we can still have concurrent tasks adding elements to this
3842          * transaction's list of dirty block groups. These tasks correspond to
3843          * endio free space workers started when writeback finishes for a
3844          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3845          * allocate new block groups as a result of COWing nodes of the root
3846          * tree when updating the free space inode. The writeback for the space
3847          * caches is triggered by an earlier call to
3848          * btrfs_start_dirty_block_groups() and iterations of the following
3849          * loop.
3850          * Also we want to do the cache_save_setup first and then run the
3851          * delayed refs to make sure we have the best chance at doing this all
3852          * in one shot.
3853          */
3854         spin_lock(&cur_trans->dirty_bgs_lock);
3855         while (!list_empty(&cur_trans->dirty_bgs)) {
3856                 cache = list_first_entry(&cur_trans->dirty_bgs,
3857                                          struct btrfs_block_group_cache,
3858                                          dirty_list);
3859
3860                 /*
3861                  * this can happen if cache_save_setup re-dirties a block
3862                  * group that is already under IO.  Just wait for it to
3863                  * finish and then do it all again
3864                  */
3865                 if (!list_empty(&cache->io_list)) {
3866                         spin_unlock(&cur_trans->dirty_bgs_lock);
3867                         list_del_init(&cache->io_list);
3868                         btrfs_wait_cache_io(trans, cache, path);
3869                         btrfs_put_block_group(cache);
3870                         spin_lock(&cur_trans->dirty_bgs_lock);
3871                 }
3872
3873                 /*
3874                  * don't remove from the dirty list until after we've waited
3875                  * on any pending IO
3876                  */
3877                 list_del_init(&cache->dirty_list);
3878                 spin_unlock(&cur_trans->dirty_bgs_lock);
3879                 should_put = 1;
3880
3881                 cache_save_setup(cache, trans, path);
3882
3883                 if (!ret)
3884                         ret = btrfs_run_delayed_refs(trans, fs_info,
3885                                                      (unsigned long) -1);
3886
3887                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3888                         cache->io_ctl.inode = NULL;
3889                         ret = btrfs_write_out_cache(fs_info, trans,
3890                                                     cache, path);
3891                         if (ret == 0 && cache->io_ctl.inode) {
3892                                 num_started++;
3893                                 should_put = 0;
3894                                 list_add_tail(&cache->io_list, io);
3895                         } else {
3896                                 /*
3897                                  * if we failed to write the cache, the
3898                                  * generation will be bad and life goes on
3899                                  */
3900                                 ret = 0;
3901                         }
3902                 }
3903                 if (!ret) {
3904                         ret = write_one_cache_group(trans, fs_info,
3905                                                     path, cache);
3906                         /*
3907                          * One of the free space endio workers might have
3908                          * created a new block group while updating a free space
3909                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3910                          * and hasn't released its transaction handle yet, in
3911                          * which case the new block group is still attached to
3912                          * its transaction handle and its creation has not
3913                          * finished yet (no block group item in the extent tree
3914                          * yet, etc). If this is the case, wait for all free
3915                          * space endio workers to finish and retry. This is a
3916                          * a very rare case so no need for a more efficient and
3917                          * complex approach.
3918                          */
3919                         if (ret == -ENOENT) {
3920                                 wait_event(cur_trans->writer_wait,
3921                                    atomic_read(&cur_trans->num_writers) == 1);
3922                                 ret = write_one_cache_group(trans, fs_info,
3923                                                             path, cache);
3924                         }
3925                         if (ret)
3926                                 btrfs_abort_transaction(trans, ret);
3927                 }
3928
3929                 /* if its not on the io list, we need to put the block group */
3930                 if (should_put)
3931                         btrfs_put_block_group(cache);
3932                 spin_lock(&cur_trans->dirty_bgs_lock);
3933         }
3934         spin_unlock(&cur_trans->dirty_bgs_lock);
3935
3936         /*
3937          * Refer to the definition of io_bgs member for details why it's safe
3938          * to use it without any locking
3939          */
3940         while (!list_empty(io)) {
3941                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3942                                          io_list);
3943                 list_del_init(&cache->io_list);
3944                 btrfs_wait_cache_io(trans, cache, path);
3945                 btrfs_put_block_group(cache);
3946         }
3947
3948         btrfs_free_path(path);
3949         return ret;
3950 }
3951
3952 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3953 {
3954         struct btrfs_block_group_cache *block_group;
3955         int readonly = 0;
3956
3957         block_group = btrfs_lookup_block_group(fs_info, bytenr);
3958         if (!block_group || block_group->ro)
3959                 readonly = 1;
3960         if (block_group)
3961                 btrfs_put_block_group(block_group);
3962         return readonly;
3963 }
3964
3965 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3966 {
3967         struct btrfs_block_group_cache *bg;
3968         bool ret = true;
3969
3970         bg = btrfs_lookup_block_group(fs_info, bytenr);
3971         if (!bg)
3972                 return false;
3973
3974         spin_lock(&bg->lock);
3975         if (bg->ro)
3976                 ret = false;
3977         else
3978                 atomic_inc(&bg->nocow_writers);
3979         spin_unlock(&bg->lock);
3980
3981         /* no put on block group, done by btrfs_dec_nocow_writers */
3982         if (!ret)
3983                 btrfs_put_block_group(bg);
3984
3985         return ret;
3986
3987 }
3988
3989 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3990 {
3991         struct btrfs_block_group_cache *bg;
3992
3993         bg = btrfs_lookup_block_group(fs_info, bytenr);
3994         ASSERT(bg);
3995         if (atomic_dec_and_test(&bg->nocow_writers))
3996                 wake_up_atomic_t(&bg->nocow_writers);
3997         /*
3998          * Once for our lookup and once for the lookup done by a previous call
3999          * to btrfs_inc_nocow_writers()
4000          */
4001         btrfs_put_block_group(bg);
4002         btrfs_put_block_group(bg);
4003 }
4004
4005 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
4006 {
4007         wait_on_atomic_t(&bg->nocow_writers, atomic_t_wait,
4008                          TASK_UNINTERRUPTIBLE);
4009 }
4010
4011 static const char *alloc_name(u64 flags)
4012 {
4013         switch (flags) {
4014         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
4015                 return "mixed";
4016         case BTRFS_BLOCK_GROUP_METADATA:
4017                 return "metadata";
4018         case BTRFS_BLOCK_GROUP_DATA:
4019                 return "data";
4020         case BTRFS_BLOCK_GROUP_SYSTEM:
4021                 return "system";
4022         default:
4023                 WARN_ON(1);
4024                 return "invalid-combination";
4025         };
4026 }
4027
4028 static int create_space_info(struct btrfs_fs_info *info, u64 flags,
4029                              struct btrfs_space_info **new)
4030 {
4031
4032         struct btrfs_space_info *space_info;
4033         int i;
4034         int ret;
4035
4036         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
4037         if (!space_info)
4038                 return -ENOMEM;
4039
4040         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
4041                                  GFP_KERNEL);
4042         if (ret) {
4043                 kfree(space_info);
4044                 return ret;
4045         }
4046
4047         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
4048                 INIT_LIST_HEAD(&space_info->block_groups[i]);
4049         init_rwsem(&space_info->groups_sem);
4050         spin_lock_init(&space_info->lock);
4051         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
4052         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4053         init_waitqueue_head(&space_info->wait);
4054         INIT_LIST_HEAD(&space_info->ro_bgs);
4055         INIT_LIST_HEAD(&space_info->tickets);
4056         INIT_LIST_HEAD(&space_info->priority_tickets);
4057
4058         ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
4059                                     info->space_info_kobj, "%s",
4060                                     alloc_name(space_info->flags));
4061         if (ret) {
4062                 percpu_counter_destroy(&space_info->total_bytes_pinned);
4063                 kfree(space_info);
4064                 return ret;
4065         }
4066
4067         *new = space_info;
4068         list_add_rcu(&space_info->list, &info->space_info);
4069         if (flags & BTRFS_BLOCK_GROUP_DATA)
4070                 info->data_sinfo = space_info;
4071
4072         return ret;
4073 }
4074
4075 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4076                              u64 total_bytes, u64 bytes_used,
4077                              u64 bytes_readonly,
4078                              struct btrfs_space_info **space_info)
4079 {
4080         struct btrfs_space_info *found;
4081         int factor;
4082
4083         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
4084                      BTRFS_BLOCK_GROUP_RAID10))
4085                 factor = 2;
4086         else
4087                 factor = 1;
4088
4089         found = __find_space_info(info, flags);
4090         ASSERT(found);
4091         spin_lock(&found->lock);
4092         found->total_bytes += total_bytes;
4093         found->disk_total += total_bytes * factor;
4094         found->bytes_used += bytes_used;
4095         found->disk_used += bytes_used * factor;
4096         found->bytes_readonly += bytes_readonly;
4097         if (total_bytes > 0)
4098                 found->full = 0;
4099         space_info_add_new_bytes(info, found, total_bytes -
4100                                  bytes_used - bytes_readonly);
4101         spin_unlock(&found->lock);
4102         *space_info = found;
4103 }
4104
4105 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4106 {
4107         u64 extra_flags = chunk_to_extended(flags) &
4108                                 BTRFS_EXTENDED_PROFILE_MASK;
4109
4110         write_seqlock(&fs_info->profiles_lock);
4111         if (flags & BTRFS_BLOCK_GROUP_DATA)
4112                 fs_info->avail_data_alloc_bits |= extra_flags;
4113         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4114                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4115         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4116                 fs_info->avail_system_alloc_bits |= extra_flags;
4117         write_sequnlock(&fs_info->profiles_lock);
4118 }
4119
4120 /*
4121  * returns target flags in extended format or 0 if restripe for this
4122  * chunk_type is not in progress
4123  *
4124  * should be called with either volume_mutex or balance_lock held
4125  */
4126 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4127 {
4128         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4129         u64 target = 0;
4130
4131         if (!bctl)
4132                 return 0;
4133
4134         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4135             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4136                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4137         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4138                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4139                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4140         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4141                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4142                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4143         }
4144
4145         return target;
4146 }
4147
4148 /*
4149  * @flags: available profiles in extended format (see ctree.h)
4150  *
4151  * Returns reduced profile in chunk format.  If profile changing is in
4152  * progress (either running or paused) picks the target profile (if it's
4153  * already available), otherwise falls back to plain reducing.
4154  */
4155 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4156 {
4157         u64 num_devices = fs_info->fs_devices->rw_devices;
4158         u64 target;
4159         u64 raid_type;
4160         u64 allowed = 0;
4161
4162         /*
4163          * see if restripe for this chunk_type is in progress, if so
4164          * try to reduce to the target profile
4165          */
4166         spin_lock(&fs_info->balance_lock);
4167         target = get_restripe_target(fs_info, flags);
4168         if (target) {
4169                 /* pick target profile only if it's already available */
4170                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4171                         spin_unlock(&fs_info->balance_lock);
4172                         return extended_to_chunk(target);
4173                 }
4174         }
4175         spin_unlock(&fs_info->balance_lock);
4176
4177         /* First, mask out the RAID levels which aren't possible */
4178         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4179                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4180                         allowed |= btrfs_raid_group[raid_type];
4181         }
4182         allowed &= flags;
4183
4184         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4185                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4186         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4187                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4188         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4189                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4190         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4191                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4192         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4193                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4194
4195         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4196
4197         return extended_to_chunk(flags | allowed);
4198 }
4199
4200 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4201 {
4202         unsigned seq;
4203         u64 flags;
4204
4205         do {
4206                 flags = orig_flags;
4207                 seq = read_seqbegin(&fs_info->profiles_lock);
4208
4209                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4210                         flags |= fs_info->avail_data_alloc_bits;
4211                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4212                         flags |= fs_info->avail_system_alloc_bits;
4213                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4214                         flags |= fs_info->avail_metadata_alloc_bits;
4215         } while (read_seqretry(&fs_info->profiles_lock, seq));
4216
4217         return btrfs_reduce_alloc_profile(fs_info, flags);
4218 }
4219
4220 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4221 {
4222         struct btrfs_fs_info *fs_info = root->fs_info;
4223         u64 flags;
4224         u64 ret;
4225
4226         if (data)
4227                 flags = BTRFS_BLOCK_GROUP_DATA;
4228         else if (root == fs_info->chunk_root)
4229                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4230         else
4231                 flags = BTRFS_BLOCK_GROUP_METADATA;
4232
4233         ret = get_alloc_profile(fs_info, flags);
4234         return ret;
4235 }
4236
4237 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4238 {
4239         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4240 }
4241
4242 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4243 {
4244         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4245 }
4246
4247 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4248 {
4249         return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4250 }
4251
4252 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4253                                  bool may_use_included)
4254 {
4255         ASSERT(s_info);
4256         return s_info->bytes_used + s_info->bytes_reserved +
4257                 s_info->bytes_pinned + s_info->bytes_readonly +
4258                 (may_use_included ? s_info->bytes_may_use : 0);
4259 }
4260
4261 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4262 {
4263         struct btrfs_root *root = inode->root;
4264         struct btrfs_fs_info *fs_info = root->fs_info;
4265         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4266         u64 used;
4267         int ret = 0;
4268         int need_commit = 2;
4269         int have_pinned_space;
4270
4271         /* make sure bytes are sectorsize aligned */
4272         bytes = ALIGN(bytes, fs_info->sectorsize);
4273
4274         if (btrfs_is_free_space_inode(inode)) {
4275                 need_commit = 0;
4276                 ASSERT(current->journal_info);
4277         }
4278
4279 again:
4280         /* make sure we have enough space to handle the data first */
4281         spin_lock(&data_sinfo->lock);
4282         used = btrfs_space_info_used(data_sinfo, true);
4283
4284         if (used + bytes > data_sinfo->total_bytes) {
4285                 struct btrfs_trans_handle *trans;
4286
4287                 /*
4288                  * if we don't have enough free bytes in this space then we need
4289                  * to alloc a new chunk.
4290                  */
4291                 if (!data_sinfo->full) {
4292                         u64 alloc_target;
4293
4294                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4295                         spin_unlock(&data_sinfo->lock);
4296
4297                         alloc_target = btrfs_data_alloc_profile(fs_info);
4298                         /*
4299                          * It is ugly that we don't call nolock join
4300                          * transaction for the free space inode case here.
4301                          * But it is safe because we only do the data space
4302                          * reservation for the free space cache in the
4303                          * transaction context, the common join transaction
4304                          * just increase the counter of the current transaction
4305                          * handler, doesn't try to acquire the trans_lock of
4306                          * the fs.
4307                          */
4308                         trans = btrfs_join_transaction(root);
4309                         if (IS_ERR(trans))
4310                                 return PTR_ERR(trans);
4311
4312                         ret = do_chunk_alloc(trans, fs_info, alloc_target,
4313                                              CHUNK_ALLOC_NO_FORCE);
4314                         btrfs_end_transaction(trans);
4315                         if (ret < 0) {
4316                                 if (ret != -ENOSPC)
4317                                         return ret;
4318                                 else {
4319                                         have_pinned_space = 1;
4320                                         goto commit_trans;
4321                                 }
4322                         }
4323
4324                         goto again;
4325                 }
4326
4327                 /*
4328                  * If we don't have enough pinned space to deal with this
4329                  * allocation, and no removed chunk in current transaction,
4330                  * don't bother committing the transaction.
4331                  */
4332                 have_pinned_space = percpu_counter_compare(
4333                         &data_sinfo->total_bytes_pinned,
4334                         used + bytes - data_sinfo->total_bytes);
4335                 spin_unlock(&data_sinfo->lock);
4336
4337                 /* commit the current transaction and try again */
4338 commit_trans:
4339                 if (need_commit) {
4340                         need_commit--;
4341
4342                         if (need_commit > 0) {
4343                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4344                                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4345                                                          (u64)-1);
4346                         }
4347
4348                         trans = btrfs_join_transaction(root);
4349                         if (IS_ERR(trans))
4350                                 return PTR_ERR(trans);
4351                         if (have_pinned_space >= 0 ||
4352                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4353                                      &trans->transaction->flags) ||
4354                             need_commit > 0) {
4355                                 ret = btrfs_commit_transaction(trans);
4356                                 if (ret)
4357                                         return ret;
4358                                 /*
4359                                  * The cleaner kthread might still be doing iput
4360                                  * operations. Wait for it to finish so that
4361                                  * more space is released.
4362                                  */
4363                                 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4364                                 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4365                                 goto again;
4366                         } else {
4367                                 btrfs_end_transaction(trans);
4368                         }
4369                 }
4370
4371                 trace_btrfs_space_reservation(fs_info,
4372                                               "space_info:enospc",
4373                                               data_sinfo->flags, bytes, 1);
4374                 return -ENOSPC;
4375         }
4376         data_sinfo->bytes_may_use += bytes;
4377         trace_btrfs_space_reservation(fs_info, "space_info",
4378                                       data_sinfo->flags, bytes, 1);
4379         spin_unlock(&data_sinfo->lock);
4380
4381         return ret;
4382 }
4383
4384 int btrfs_check_data_free_space(struct inode *inode,
4385                         struct extent_changeset **reserved, u64 start, u64 len)
4386 {
4387         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4388         int ret;
4389
4390         /* align the range */
4391         len = round_up(start + len, fs_info->sectorsize) -
4392               round_down(start, fs_info->sectorsize);
4393         start = round_down(start, fs_info->sectorsize);
4394
4395         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4396         if (ret < 0)
4397                 return ret;
4398
4399         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4400         ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4401         if (ret < 0)
4402                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4403         else
4404                 ret = 0;
4405         return ret;
4406 }
4407
4408 /*
4409  * Called if we need to clear a data reservation for this inode
4410  * Normally in a error case.
4411  *
4412  * This one will *NOT* use accurate qgroup reserved space API, just for case
4413  * which we can't sleep and is sure it won't affect qgroup reserved space.
4414  * Like clear_bit_hook().
4415  */
4416 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4417                                             u64 len)
4418 {
4419         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4420         struct btrfs_space_info *data_sinfo;
4421
4422         /* Make sure the range is aligned to sectorsize */
4423         len = round_up(start + len, fs_info->sectorsize) -
4424               round_down(start, fs_info->sectorsize);
4425         start = round_down(start, fs_info->sectorsize);
4426
4427         data_sinfo = fs_info->data_sinfo;
4428         spin_lock(&data_sinfo->lock);
4429         if (WARN_ON(data_sinfo->bytes_may_use < len))
4430                 data_sinfo->bytes_may_use = 0;
4431         else
4432                 data_sinfo->bytes_may_use -= len;
4433         trace_btrfs_space_reservation(fs_info, "space_info",
4434                                       data_sinfo->flags, len, 0);
4435         spin_unlock(&data_sinfo->lock);
4436 }
4437
4438 /*
4439  * Called if we need to clear a data reservation for this inode
4440  * Normally in a error case.
4441  *
4442  * This one will handle the per-inode data rsv map for accurate reserved
4443  * space framework.
4444  */
4445 void btrfs_free_reserved_data_space(struct inode *inode,
4446                         struct extent_changeset *reserved, u64 start, u64 len)
4447 {
4448         struct btrfs_root *root = BTRFS_I(inode)->root;
4449
4450         /* Make sure the range is aligned to sectorsize */
4451         len = round_up(start + len, root->fs_info->sectorsize) -
4452               round_down(start, root->fs_info->sectorsize);
4453         start = round_down(start, root->fs_info->sectorsize);
4454
4455         btrfs_free_reserved_data_space_noquota(inode, start, len);
4456         btrfs_qgroup_free_data(inode, reserved, start, len);
4457 }
4458
4459 static void force_metadata_allocation(struct btrfs_fs_info *info)
4460 {
4461         struct list_head *head = &info->space_info;
4462         struct btrfs_space_info *found;
4463
4464         rcu_read_lock();
4465         list_for_each_entry_rcu(found, head, list) {
4466                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4467                         found->force_alloc = CHUNK_ALLOC_FORCE;
4468         }
4469         rcu_read_unlock();
4470 }
4471
4472 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4473 {
4474         return (global->size << 1);
4475 }
4476
4477 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4478                               struct btrfs_space_info *sinfo, int force)
4479 {
4480         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4481         u64 bytes_used = btrfs_space_info_used(sinfo, false);
4482         u64 thresh;
4483
4484         if (force == CHUNK_ALLOC_FORCE)
4485                 return 1;
4486
4487         /*
4488          * We need to take into account the global rsv because for all intents
4489          * and purposes it's used space.  Don't worry about locking the
4490          * global_rsv, it doesn't change except when the transaction commits.
4491          */
4492         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4493                 bytes_used += calc_global_rsv_need_space(global_rsv);
4494
4495         /*
4496          * in limited mode, we want to have some free space up to
4497          * about 1% of the FS size.
4498          */
4499         if (force == CHUNK_ALLOC_LIMITED) {
4500                 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4501                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4502
4503                 if (sinfo->total_bytes - bytes_used < thresh)
4504                         return 1;
4505         }
4506
4507         if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4508                 return 0;
4509         return 1;
4510 }
4511
4512 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4513 {
4514         u64 num_dev;
4515
4516         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4517                     BTRFS_BLOCK_GROUP_RAID0 |
4518                     BTRFS_BLOCK_GROUP_RAID5 |
4519                     BTRFS_BLOCK_GROUP_RAID6))
4520                 num_dev = fs_info->fs_devices->rw_devices;
4521         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4522                 num_dev = 2;
4523         else
4524                 num_dev = 1;    /* DUP or single */
4525
4526         return num_dev;
4527 }
4528
4529 /*
4530  * If @is_allocation is true, reserve space in the system space info necessary
4531  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4532  * removing a chunk.
4533  */
4534 void check_system_chunk(struct btrfs_trans_handle *trans,
4535                         struct btrfs_fs_info *fs_info, u64 type)
4536 {
4537         struct btrfs_space_info *info;
4538         u64 left;
4539         u64 thresh;
4540         int ret = 0;
4541         u64 num_devs;
4542
4543         /*
4544          * Needed because we can end up allocating a system chunk and for an
4545          * atomic and race free space reservation in the chunk block reserve.
4546          */
4547         ASSERT(mutex_is_locked(&fs_info->chunk_mutex));
4548
4549         info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4550         spin_lock(&info->lock);
4551         left = info->total_bytes - btrfs_space_info_used(info, true);
4552         spin_unlock(&info->lock);
4553
4554         num_devs = get_profile_num_devs(fs_info, type);
4555
4556         /* num_devs device items to update and 1 chunk item to add or remove */
4557         thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4558                 btrfs_calc_trans_metadata_size(fs_info, 1);
4559
4560         if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4561                 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4562                            left, thresh, type);
4563                 dump_space_info(fs_info, info, 0, 0);
4564         }
4565
4566         if (left < thresh) {
4567                 u64 flags = btrfs_system_alloc_profile(fs_info);
4568
4569                 /*
4570                  * Ignore failure to create system chunk. We might end up not
4571                  * needing it, as we might not need to COW all nodes/leafs from
4572                  * the paths we visit in the chunk tree (they were already COWed
4573                  * or created in the current transaction for example).
4574                  */
4575                 ret = btrfs_alloc_chunk(trans, fs_info, flags);
4576         }
4577
4578         if (!ret) {
4579                 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4580                                           &fs_info->chunk_block_rsv,
4581                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4582                 if (!ret)
4583                         trans->chunk_bytes_reserved += thresh;
4584         }
4585 }
4586
4587 /*
4588  * If force is CHUNK_ALLOC_FORCE:
4589  *    - return 1 if it successfully allocates a chunk,
4590  *    - return errors including -ENOSPC otherwise.
4591  * If force is NOT CHUNK_ALLOC_FORCE:
4592  *    - return 0 if it doesn't need to allocate a new chunk,
4593  *    - return 1 if it successfully allocates a chunk,
4594  *    - return errors including -ENOSPC otherwise.
4595  */
4596 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4597                           struct btrfs_fs_info *fs_info, u64 flags, int force)
4598 {
4599         struct btrfs_space_info *space_info;
4600         int wait_for_alloc = 0;
4601         int ret = 0;
4602
4603         /* Don't re-enter if we're already allocating a chunk */
4604         if (trans->allocating_chunk)
4605                 return -ENOSPC;
4606
4607         space_info = __find_space_info(fs_info, flags);
4608         if (!space_info) {
4609                 ret = create_space_info(fs_info, flags, &space_info);
4610                 if (ret)
4611                         return ret;
4612         }
4613
4614 again:
4615         spin_lock(&space_info->lock);
4616         if (force < space_info->force_alloc)
4617                 force = space_info->force_alloc;
4618         if (space_info->full) {
4619                 if (should_alloc_chunk(fs_info, space_info, force))
4620                         ret = -ENOSPC;
4621                 else
4622                         ret = 0;
4623                 spin_unlock(&space_info->lock);
4624                 return ret;
4625         }
4626
4627         if (!should_alloc_chunk(fs_info, space_info, force)) {
4628                 spin_unlock(&space_info->lock);
4629                 return 0;
4630         } else if (space_info->chunk_alloc) {
4631                 wait_for_alloc = 1;
4632         } else {
4633                 space_info->chunk_alloc = 1;
4634         }
4635
4636         spin_unlock(&space_info->lock);
4637
4638         mutex_lock(&fs_info->chunk_mutex);
4639
4640         /*
4641          * The chunk_mutex is held throughout the entirety of a chunk
4642          * allocation, so once we've acquired the chunk_mutex we know that the
4643          * other guy is done and we need to recheck and see if we should
4644          * allocate.
4645          */
4646         if (wait_for_alloc) {
4647                 mutex_unlock(&fs_info->chunk_mutex);
4648                 wait_for_alloc = 0;
4649                 goto again;
4650         }
4651
4652         trans->allocating_chunk = true;
4653
4654         /*
4655          * If we have mixed data/metadata chunks we want to make sure we keep
4656          * allocating mixed chunks instead of individual chunks.
4657          */
4658         if (btrfs_mixed_space_info(space_info))
4659                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4660
4661         /*
4662          * if we're doing a data chunk, go ahead and make sure that
4663          * we keep a reasonable number of metadata chunks allocated in the
4664          * FS as well.
4665          */
4666         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4667                 fs_info->data_chunk_allocations++;
4668                 if (!(fs_info->data_chunk_allocations %
4669                       fs_info->metadata_ratio))
4670                         force_metadata_allocation(fs_info);
4671         }
4672
4673         /*
4674          * Check if we have enough space in SYSTEM chunk because we may need
4675          * to update devices.
4676          */
4677         check_system_chunk(trans, fs_info, flags);
4678
4679         ret = btrfs_alloc_chunk(trans, fs_info, flags);
4680         trans->allocating_chunk = false;
4681
4682         spin_lock(&space_info->lock);
4683         if (ret < 0 && ret != -ENOSPC)
4684                 goto out;
4685         if (ret)
4686                 space_info->full = 1;
4687         else
4688                 ret = 1;
4689
4690         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4691 out:
4692         space_info->chunk_alloc = 0;
4693         spin_unlock(&space_info->lock);
4694         mutex_unlock(&fs_info->chunk_mutex);
4695         /*
4696          * When we allocate a new chunk we reserve space in the chunk block
4697          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4698          * add new nodes/leafs to it if we end up needing to do it when
4699          * inserting the chunk item and updating device items as part of the
4700          * second phase of chunk allocation, performed by
4701          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4702          * large number of new block groups to create in our transaction
4703          * handle's new_bgs list to avoid exhausting the chunk block reserve
4704          * in extreme cases - like having a single transaction create many new
4705          * block groups when starting to write out the free space caches of all
4706          * the block groups that were made dirty during the lifetime of the
4707          * transaction.
4708          */
4709         if (trans->can_flush_pending_bgs &&
4710             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4711                 btrfs_create_pending_block_groups(trans);
4712                 btrfs_trans_release_chunk_metadata(trans);
4713         }
4714         return ret;
4715 }
4716
4717 static int can_overcommit(struct btrfs_fs_info *fs_info,
4718                           struct btrfs_space_info *space_info, u64 bytes,
4719                           enum btrfs_reserve_flush_enum flush,
4720                           bool system_chunk)
4721 {
4722         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4723         u64 profile;
4724         u64 space_size;
4725         u64 avail;
4726         u64 used;
4727
4728         /* Don't overcommit when in mixed mode. */
4729         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4730                 return 0;
4731
4732         if (system_chunk)
4733                 profile = btrfs_system_alloc_profile(fs_info);
4734         else
4735                 profile = btrfs_metadata_alloc_profile(fs_info);
4736
4737         used = btrfs_space_info_used(space_info, false);
4738
4739         /*
4740          * We only want to allow over committing if we have lots of actual space
4741          * free, but if we don't have enough space to handle the global reserve
4742          * space then we could end up having a real enospc problem when trying
4743          * to allocate a chunk or some other such important allocation.
4744          */
4745         spin_lock(&global_rsv->lock);
4746         space_size = calc_global_rsv_need_space(global_rsv);
4747         spin_unlock(&global_rsv->lock);
4748         if (used + space_size >= space_info->total_bytes)
4749                 return 0;
4750
4751         used += space_info->bytes_may_use;
4752
4753         avail = atomic64_read(&fs_info->free_chunk_space);
4754
4755         /*
4756          * If we have dup, raid1 or raid10 then only half of the free
4757          * space is actually useable.  For raid56, the space info used
4758          * doesn't include the parity drive, so we don't have to
4759          * change the math
4760          */
4761         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4762                        BTRFS_BLOCK_GROUP_RAID1 |
4763                        BTRFS_BLOCK_GROUP_RAID10))
4764                 avail >>= 1;
4765
4766         /*
4767          * If we aren't flushing all things, let us overcommit up to
4768          * 1/2th of the space. If we can flush, don't let us overcommit
4769          * too much, let it overcommit up to 1/8 of the space.
4770          */
4771         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4772                 avail >>= 3;
4773         else
4774                 avail >>= 1;
4775
4776         if (used + bytes < space_info->total_bytes + avail)
4777                 return 1;
4778         return 0;
4779 }
4780
4781 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4782                                          unsigned long nr_pages, int nr_items)
4783 {
4784         struct super_block *sb = fs_info->sb;
4785
4786         if (down_read_trylock(&sb->s_umount)) {
4787                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4788                 up_read(&sb->s_umount);
4789         } else {
4790                 /*
4791                  * We needn't worry the filesystem going from r/w to r/o though
4792                  * we don't acquire ->s_umount mutex, because the filesystem
4793                  * should guarantee the delalloc inodes list be empty after
4794                  * the filesystem is readonly(all dirty pages are written to
4795                  * the disk).
4796                  */
4797                 btrfs_start_delalloc_roots(fs_info, 0, nr_items);
4798                 if (!current->journal_info)
4799                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4800         }
4801 }
4802
4803 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4804                                         u64 to_reclaim)
4805 {
4806         u64 bytes;
4807         u64 nr;
4808
4809         bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4810         nr = div64_u64(to_reclaim, bytes);
4811         if (!nr)
4812                 nr = 1;
4813         return nr;
4814 }
4815
4816 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4817
4818 /*
4819  * shrink metadata reservation for delalloc
4820  */
4821 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4822                             u64 orig, bool wait_ordered)
4823 {
4824         struct btrfs_space_info *space_info;
4825         struct btrfs_trans_handle *trans;
4826         u64 delalloc_bytes;
4827         u64 max_reclaim;
4828         u64 items;
4829         long time_left;
4830         unsigned long nr_pages;
4831         int loops;
4832         enum btrfs_reserve_flush_enum flush;
4833
4834         /* Calc the number of the pages we need flush for space reservation */
4835         items = calc_reclaim_items_nr(fs_info, to_reclaim);
4836         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4837
4838         trans = (struct btrfs_trans_handle *)current->journal_info;
4839         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4840
4841         delalloc_bytes = percpu_counter_sum_positive(
4842                                                 &fs_info->delalloc_bytes);
4843         if (delalloc_bytes == 0) {
4844                 if (trans)
4845                         return;
4846                 if (wait_ordered)
4847                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4848                 return;
4849         }
4850
4851         loops = 0;
4852         while (delalloc_bytes && loops < 3) {
4853                 max_reclaim = min(delalloc_bytes, to_reclaim);
4854                 nr_pages = max_reclaim >> PAGE_SHIFT;
4855                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4856                 /*
4857                  * We need to wait for the async pages to actually start before
4858                  * we do anything.
4859                  */
4860                 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4861                 if (!max_reclaim)
4862                         goto skip_async;
4863
4864                 if (max_reclaim <= nr_pages)
4865                         max_reclaim = 0;
4866                 else
4867                         max_reclaim -= nr_pages;
4868
4869                 wait_event(fs_info->async_submit_wait,
4870                            atomic_read(&fs_info->async_delalloc_pages) <=
4871                            (int)max_reclaim);
4872 skip_async:
4873                 if (!trans)
4874                         flush = BTRFS_RESERVE_FLUSH_ALL;
4875                 else
4876                         flush = BTRFS_RESERVE_NO_FLUSH;
4877                 spin_lock(&space_info->lock);
4878                 if (list_empty(&space_info->tickets) &&
4879                     list_empty(&space_info->priority_tickets)) {
4880                         spin_unlock(&space_info->lock);
4881                         break;
4882                 }
4883                 spin_unlock(&space_info->lock);
4884
4885                 loops++;
4886                 if (wait_ordered && !trans) {
4887                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4888                 } else {
4889                         time_left = schedule_timeout_killable(1);
4890                         if (time_left)
4891                                 break;
4892                 }
4893                 delalloc_bytes = percpu_counter_sum_positive(
4894                                                 &fs_info->delalloc_bytes);
4895         }
4896 }
4897
4898 struct reserve_ticket {
4899         u64 bytes;
4900         int error;
4901         struct list_head list;
4902         wait_queue_head_t wait;
4903 };
4904
4905 /**
4906  * maybe_commit_transaction - possibly commit the transaction if its ok to
4907  * @root - the root we're allocating for
4908  * @bytes - the number of bytes we want to reserve
4909  * @force - force the commit
4910  *
4911  * This will check to make sure that committing the transaction will actually
4912  * get us somewhere and then commit the transaction if it does.  Otherwise it
4913  * will return -ENOSPC.
4914  */
4915 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4916                                   struct btrfs_space_info *space_info)
4917 {
4918         struct reserve_ticket *ticket = NULL;
4919         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4920         struct btrfs_trans_handle *trans;
4921         u64 bytes;
4922
4923         trans = (struct btrfs_trans_handle *)current->journal_info;
4924         if (trans)
4925                 return -EAGAIN;
4926
4927         spin_lock(&space_info->lock);
4928         if (!list_empty(&space_info->priority_tickets))
4929                 ticket = list_first_entry(&space_info->priority_tickets,
4930                                           struct reserve_ticket, list);
4931         else if (!list_empty(&space_info->tickets))
4932                 ticket = list_first_entry(&space_info->tickets,
4933                                           struct reserve_ticket, list);
4934         bytes = (ticket) ? ticket->bytes : 0;
4935         spin_unlock(&space_info->lock);
4936
4937         if (!bytes)
4938                 return 0;
4939
4940         /* See if there is enough pinned space to make this reservation */
4941         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4942                                    bytes) >= 0)
4943                 goto commit;
4944
4945         /*
4946          * See if there is some space in the delayed insertion reservation for
4947          * this reservation.
4948          */
4949         if (space_info != delayed_rsv->space_info)
4950                 return -ENOSPC;
4951
4952         spin_lock(&delayed_rsv->lock);
4953         if (delayed_rsv->size > bytes)
4954                 bytes = 0;
4955         else
4956                 bytes -= delayed_rsv->size;
4957         spin_unlock(&delayed_rsv->lock);
4958
4959         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4960                                    bytes) < 0) {
4961                 return -ENOSPC;
4962         }
4963
4964 commit:
4965         trans = btrfs_join_transaction(fs_info->extent_root);
4966         if (IS_ERR(trans))
4967                 return -ENOSPC;
4968
4969         return btrfs_commit_transaction(trans);
4970 }
4971
4972 /*
4973  * Try to flush some data based on policy set by @state. This is only advisory
4974  * and may fail for various reasons. The caller is supposed to examine the
4975  * state of @space_info to detect the outcome.
4976  */
4977 static void flush_space(struct btrfs_fs_info *fs_info,
4978                        struct btrfs_space_info *space_info, u64 num_bytes,
4979                        int state)
4980 {
4981         struct btrfs_root *root = fs_info->extent_root;
4982         struct btrfs_trans_handle *trans;
4983         int nr;
4984         int ret = 0;
4985
4986         switch (state) {
4987         case FLUSH_DELAYED_ITEMS_NR:
4988         case FLUSH_DELAYED_ITEMS:
4989                 if (state == FLUSH_DELAYED_ITEMS_NR)
4990                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4991                 else
4992                         nr = -1;
4993
4994                 trans = btrfs_join_transaction(root);
4995                 if (IS_ERR(trans)) {
4996                         ret = PTR_ERR(trans);
4997                         break;
4998                 }
4999                 ret = btrfs_run_delayed_items_nr(trans, nr);
5000                 btrfs_end_transaction(trans);
5001                 break;
5002         case FLUSH_DELALLOC:
5003         case FLUSH_DELALLOC_WAIT:
5004                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
5005                                 state == FLUSH_DELALLOC_WAIT);
5006                 break;
5007         case ALLOC_CHUNK:
5008                 trans = btrfs_join_transaction(root);
5009                 if (IS_ERR(trans)) {
5010                         ret = PTR_ERR(trans);
5011                         break;
5012                 }
5013                 ret = do_chunk_alloc(trans, fs_info,
5014                                      btrfs_metadata_alloc_profile(fs_info),
5015                                      CHUNK_ALLOC_NO_FORCE);
5016                 btrfs_end_transaction(trans);
5017                 if (ret > 0 || ret == -ENOSPC)
5018                         ret = 0;
5019                 break;
5020         case COMMIT_TRANS:
5021                 ret = may_commit_transaction(fs_info, space_info);
5022                 break;
5023         default:
5024                 ret = -ENOSPC;
5025                 break;
5026         }
5027
5028         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
5029                                 ret);
5030         return;
5031 }
5032
5033 static inline u64
5034 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
5035                                  struct btrfs_space_info *space_info,
5036                                  bool system_chunk)
5037 {
5038         struct reserve_ticket *ticket;
5039         u64 used;
5040         u64 expected;
5041         u64 to_reclaim = 0;
5042
5043         list_for_each_entry(ticket, &space_info->tickets, list)
5044                 to_reclaim += ticket->bytes;
5045         list_for_each_entry(ticket, &space_info->priority_tickets, list)
5046                 to_reclaim += ticket->bytes;
5047         if (to_reclaim)
5048                 return to_reclaim;
5049
5050         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
5051         if (can_overcommit(fs_info, space_info, to_reclaim,
5052                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5053                 return 0;
5054
5055         used = btrfs_space_info_used(space_info, true);
5056
5057         if (can_overcommit(fs_info, space_info, SZ_1M,
5058                            BTRFS_RESERVE_FLUSH_ALL, system_chunk))
5059                 expected = div_factor_fine(space_info->total_bytes, 95);
5060         else
5061                 expected = div_factor_fine(space_info->total_bytes, 90);
5062
5063         if (used > expected)
5064                 to_reclaim = used - expected;
5065         else
5066                 to_reclaim = 0;
5067         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5068                                      space_info->bytes_reserved);
5069         return to_reclaim;
5070 }
5071
5072 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5073                                         struct btrfs_space_info *space_info,
5074                                         u64 used, bool system_chunk)
5075 {
5076         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5077
5078         /* If we're just plain full then async reclaim just slows us down. */
5079         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5080                 return 0;
5081
5082         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5083                                               system_chunk))
5084                 return 0;
5085
5086         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5087                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5088 }
5089
5090 static void wake_all_tickets(struct list_head *head)
5091 {
5092         struct reserve_ticket *ticket;
5093
5094         while (!list_empty(head)) {
5095                 ticket = list_first_entry(head, struct reserve_ticket, list);
5096                 list_del_init(&ticket->list);
5097                 ticket->error = -ENOSPC;
5098                 wake_up(&ticket->wait);
5099         }
5100 }
5101
5102 /*
5103  * This is for normal flushers, we can wait all goddamned day if we want to.  We
5104  * will loop and continuously try to flush as long as we are making progress.
5105  * We count progress as clearing off tickets each time we have to loop.
5106  */
5107 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5108 {
5109         struct btrfs_fs_info *fs_info;
5110         struct btrfs_space_info *space_info;
5111         u64 to_reclaim;
5112         int flush_state;
5113         int commit_cycles = 0;
5114         u64 last_tickets_id;
5115
5116         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5117         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5118
5119         spin_lock(&space_info->lock);
5120         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5121                                                       false);
5122         if (!to_reclaim) {
5123                 space_info->flush = 0;
5124                 spin_unlock(&space_info->lock);
5125                 return;
5126         }
5127         last_tickets_id = space_info->tickets_id;
5128         spin_unlock(&space_info->lock);
5129
5130         flush_state = FLUSH_DELAYED_ITEMS_NR;
5131         do {
5132                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5133                 spin_lock(&space_info->lock);
5134                 if (list_empty(&space_info->tickets)) {
5135                         space_info->flush = 0;
5136                         spin_unlock(&space_info->lock);
5137                         return;
5138                 }
5139                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5140                                                               space_info,
5141                                                               false);
5142                 if (last_tickets_id == space_info->tickets_id) {
5143                         flush_state++;
5144                 } else {
5145                         last_tickets_id = space_info->tickets_id;
5146                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5147                         if (commit_cycles)
5148                                 commit_cycles--;
5149                 }
5150
5151                 if (flush_state > COMMIT_TRANS) {
5152                         commit_cycles++;
5153                         if (commit_cycles > 2) {
5154                                 wake_all_tickets(&space_info->tickets);
5155                                 space_info->flush = 0;
5156                         } else {
5157                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5158                         }
5159                 }
5160                 spin_unlock(&space_info->lock);
5161         } while (flush_state <= COMMIT_TRANS);
5162 }
5163
5164 void btrfs_init_async_reclaim_work(struct work_struct *work)
5165 {
5166         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5167 }
5168
5169 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5170                                             struct btrfs_space_info *space_info,
5171                                             struct reserve_ticket *ticket)
5172 {
5173         u64 to_reclaim;
5174         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5175
5176         spin_lock(&space_info->lock);
5177         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5178                                                       false);
5179         if (!to_reclaim) {
5180                 spin_unlock(&space_info->lock);
5181                 return;
5182         }
5183         spin_unlock(&space_info->lock);
5184
5185         do {
5186                 flush_space(fs_info, space_info, to_reclaim, flush_state);
5187                 flush_state++;
5188                 spin_lock(&space_info->lock);
5189                 if (ticket->bytes == 0) {
5190                         spin_unlock(&space_info->lock);
5191                         return;
5192                 }
5193                 spin_unlock(&space_info->lock);
5194
5195                 /*
5196                  * Priority flushers can't wait on delalloc without
5197                  * deadlocking.
5198                  */
5199                 if (flush_state == FLUSH_DELALLOC ||
5200                     flush_state == FLUSH_DELALLOC_WAIT)
5201                         flush_state = ALLOC_CHUNK;
5202         } while (flush_state < COMMIT_TRANS);
5203 }
5204
5205 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5206                                struct btrfs_space_info *space_info,
5207                                struct reserve_ticket *ticket, u64 orig_bytes)
5208
5209 {
5210         DEFINE_WAIT(wait);
5211         int ret = 0;
5212
5213         spin_lock(&space_info->lock);
5214         while (ticket->bytes > 0 && ticket->error == 0) {
5215                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5216                 if (ret) {
5217                         ret = -EINTR;
5218                         break;
5219                 }
5220                 spin_unlock(&space_info->lock);
5221
5222                 schedule();
5223
5224                 finish_wait(&ticket->wait, &wait);
5225                 spin_lock(&space_info->lock);
5226         }
5227         if (!ret)
5228                 ret = ticket->error;
5229         if (!list_empty(&ticket->list))
5230                 list_del_init(&ticket->list);
5231         if (ticket->bytes && ticket->bytes < orig_bytes) {
5232                 u64 num_bytes = orig_bytes - ticket->bytes;
5233                 space_info->bytes_may_use -= num_bytes;
5234                 trace_btrfs_space_reservation(fs_info, "space_info",
5235                                               space_info->flags, num_bytes, 0);
5236         }
5237         spin_unlock(&space_info->lock);
5238
5239         return ret;
5240 }
5241
5242 /**
5243  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5244  * @root - the root we're allocating for
5245  * @space_info - the space info we want to allocate from
5246  * @orig_bytes - the number of bytes we want
5247  * @flush - whether or not we can flush to make our reservation
5248  *
5249  * This will reserve orig_bytes number of bytes from the space info associated
5250  * with the block_rsv.  If there is not enough space it will make an attempt to
5251  * flush out space to make room.  It will do this by flushing delalloc if
5252  * possible or committing the transaction.  If flush is 0 then no attempts to
5253  * regain reservations will be made and this will fail if there is not enough
5254  * space already.
5255  */
5256 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5257                                     struct btrfs_space_info *space_info,
5258                                     u64 orig_bytes,
5259                                     enum btrfs_reserve_flush_enum flush,
5260                                     bool system_chunk)
5261 {
5262         struct reserve_ticket ticket;
5263         u64 used;
5264         int ret = 0;
5265
5266         ASSERT(orig_bytes);
5267         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5268
5269         spin_lock(&space_info->lock);
5270         ret = -ENOSPC;
5271         used = btrfs_space_info_used(space_info, true);
5272
5273         /*
5274          * If we have enough space then hooray, make our reservation and carry
5275          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5276          * If not things get more complicated.
5277          */
5278         if (used + orig_bytes <= space_info->total_bytes) {
5279                 space_info->bytes_may_use += orig_bytes;
5280                 trace_btrfs_space_reservation(fs_info, "space_info",
5281                                               space_info->flags, orig_bytes, 1);
5282                 ret = 0;
5283         } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5284                                   system_chunk)) {
5285                 space_info->bytes_may_use += orig_bytes;
5286                 trace_btrfs_space_reservation(fs_info, "space_info",
5287                                               space_info->flags, orig_bytes, 1);
5288                 ret = 0;
5289         }
5290
5291         /*
5292          * If we couldn't make a reservation then setup our reservation ticket
5293          * and kick the async worker if it's not already running.
5294          *
5295          * If we are a priority flusher then we just need to add our ticket to
5296          * the list and we will do our own flushing further down.
5297          */
5298         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5299                 ticket.bytes = orig_bytes;
5300                 ticket.error = 0;
5301                 init_waitqueue_head(&ticket.wait);
5302                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5303                         list_add_tail(&ticket.list, &space_info->tickets);
5304                         if (!space_info->flush) {
5305                                 space_info->flush = 1;
5306                                 trace_btrfs_trigger_flush(fs_info,
5307                                                           space_info->flags,
5308                                                           orig_bytes, flush,
5309                                                           "enospc");
5310                                 queue_work(system_unbound_wq,
5311                                            &fs_info->async_reclaim_work);
5312                         }
5313                 } else {
5314                         list_add_tail(&ticket.list,
5315                                       &space_info->priority_tickets);
5316                 }
5317         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5318                 used += orig_bytes;
5319                 /*
5320                  * We will do the space reservation dance during log replay,
5321                  * which means we won't have fs_info->fs_root set, so don't do
5322                  * the async reclaim as we will panic.
5323                  */
5324                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5325                     need_do_async_reclaim(fs_info, space_info,
5326                                           used, system_chunk) &&
5327                     !work_busy(&fs_info->async_reclaim_work)) {
5328                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
5329                                                   orig_bytes, flush, "preempt");
5330                         queue_work(system_unbound_wq,
5331                                    &fs_info->async_reclaim_work);
5332                 }
5333         }
5334         spin_unlock(&space_info->lock);
5335         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5336                 return ret;
5337
5338         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5339                 return wait_reserve_ticket(fs_info, space_info, &ticket,
5340                                            orig_bytes);
5341
5342         ret = 0;
5343         priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5344         spin_lock(&space_info->lock);
5345         if (ticket.bytes) {
5346                 if (ticket.bytes < orig_bytes) {
5347                         u64 num_bytes = orig_bytes - ticket.bytes;
5348                         space_info->bytes_may_use -= num_bytes;
5349                         trace_btrfs_space_reservation(fs_info, "space_info",
5350                                                       space_info->flags,
5351                                                       num_bytes, 0);
5352
5353                 }
5354                 list_del_init(&ticket.list);
5355                 ret = -ENOSPC;
5356         }
5357         spin_unlock(&space_info->lock);
5358         ASSERT(list_empty(&ticket.list));
5359         return ret;
5360 }
5361
5362 /**
5363  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5364  * @root - the root we're allocating for
5365  * @block_rsv - the block_rsv we're allocating for
5366  * @orig_bytes - the number of bytes we want
5367  * @flush - whether or not we can flush to make our reservation
5368  *
5369  * This will reserve orgi_bytes number of bytes from the space info associated
5370  * with the block_rsv.  If there is not enough space it will make an attempt to
5371  * flush out space to make room.  It will do this by flushing delalloc if
5372  * possible or committing the transaction.  If flush is 0 then no attempts to
5373  * regain reservations will be made and this will fail if there is not enough
5374  * space already.
5375  */
5376 static int reserve_metadata_bytes(struct btrfs_root *root,
5377                                   struct btrfs_block_rsv *block_rsv,
5378                                   u64 orig_bytes,
5379                                   enum btrfs_reserve_flush_enum flush)
5380 {
5381         struct btrfs_fs_info *fs_info = root->fs_info;
5382         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5383         int ret;
5384         bool system_chunk = (root == fs_info->chunk_root);
5385
5386         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5387                                        orig_bytes, flush, system_chunk);
5388         if (ret == -ENOSPC &&
5389             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5390                 if (block_rsv != global_rsv &&
5391                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5392                         ret = 0;
5393         }
5394         if (ret == -ENOSPC) {
5395                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5396                                               block_rsv->space_info->flags,
5397                                               orig_bytes, 1);
5398
5399                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5400                         dump_space_info(fs_info, block_rsv->space_info,
5401                                         orig_bytes, 0);
5402         }
5403         return ret;
5404 }
5405
5406 static struct btrfs_block_rsv *get_block_rsv(
5407                                         const struct btrfs_trans_handle *trans,
5408                                         const struct btrfs_root *root)
5409 {
5410         struct btrfs_fs_info *fs_info = root->fs_info;
5411         struct btrfs_block_rsv *block_rsv = NULL;
5412
5413         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5414             (root == fs_info->csum_root && trans->adding_csums) ||
5415             (root == fs_info->uuid_root))
5416                 block_rsv = trans->block_rsv;
5417
5418         if (!block_rsv)
5419                 block_rsv = root->block_rsv;
5420
5421         if (!block_rsv)
5422                 block_rsv = &fs_info->empty_block_rsv;
5423
5424         return block_rsv;
5425 }
5426
5427 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5428                                u64 num_bytes)
5429 {
5430         int ret = -ENOSPC;
5431         spin_lock(&block_rsv->lock);
5432         if (block_rsv->reserved >= num_bytes) {
5433                 block_rsv->reserved -= num_bytes;
5434                 if (block_rsv->reserved < block_rsv->size)
5435                         block_rsv->full = 0;
5436                 ret = 0;
5437         }
5438         spin_unlock(&block_rsv->lock);
5439         return ret;
5440 }
5441
5442 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5443                                 u64 num_bytes, int update_size)
5444 {
5445         spin_lock(&block_rsv->lock);
5446         block_rsv->reserved += num_bytes;
5447         if (update_size)
5448                 block_rsv->size += num_bytes;
5449         else if (block_rsv->reserved >= block_rsv->size)
5450                 block_rsv->full = 1;
5451         spin_unlock(&block_rsv->lock);
5452 }
5453
5454 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5455                              struct btrfs_block_rsv *dest, u64 num_bytes,
5456                              int min_factor)
5457 {
5458         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5459         u64 min_bytes;
5460
5461         if (global_rsv->space_info != dest->space_info)
5462                 return -ENOSPC;
5463
5464         spin_lock(&global_rsv->lock);
5465         min_bytes = div_factor(global_rsv->size, min_factor);
5466         if (global_rsv->reserved < min_bytes + num_bytes) {
5467                 spin_unlock(&global_rsv->lock);
5468                 return -ENOSPC;
5469         }
5470         global_rsv->reserved -= num_bytes;
5471         if (global_rsv->reserved < global_rsv->size)
5472                 global_rsv->full = 0;
5473         spin_unlock(&global_rsv->lock);
5474
5475         block_rsv_add_bytes(dest, num_bytes, 1);
5476         return 0;
5477 }
5478
5479 /*
5480  * This is for space we already have accounted in space_info->bytes_may_use, so
5481  * basically when we're returning space from block_rsv's.
5482  */
5483 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5484                                      struct btrfs_space_info *space_info,
5485                                      u64 num_bytes)
5486 {
5487         struct reserve_ticket *ticket;
5488         struct list_head *head;
5489         u64 used;
5490         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5491         bool check_overcommit = false;
5492
5493         spin_lock(&space_info->lock);
5494         head = &space_info->priority_tickets;
5495
5496         /*
5497          * If we are over our limit then we need to check and see if we can
5498          * overcommit, and if we can't then we just need to free up our space
5499          * and not satisfy any requests.
5500          */
5501         used = btrfs_space_info_used(space_info, true);
5502         if (used - num_bytes >= space_info->total_bytes)
5503                 check_overcommit = true;
5504 again:
5505         while (!list_empty(head) && num_bytes) {
5506                 ticket = list_first_entry(head, struct reserve_ticket,
5507                                           list);
5508                 /*
5509                  * We use 0 bytes because this space is already reserved, so
5510                  * adding the ticket space would be a double count.
5511                  */
5512                 if (check_overcommit &&
5513                     !can_overcommit(fs_info, space_info, 0, flush, false))
5514                         break;
5515                 if (num_bytes >= ticket->bytes) {
5516                         list_del_init(&ticket->list);
5517                         num_bytes -= ticket->bytes;
5518                         ticket->bytes = 0;
5519                         space_info->tickets_id++;
5520                         wake_up(&ticket->wait);
5521                 } else {
5522                         ticket->bytes -= num_bytes;
5523                         num_bytes = 0;
5524                 }
5525         }
5526
5527         if (num_bytes && head == &space_info->priority_tickets) {
5528                 head = &space_info->tickets;
5529                 flush = BTRFS_RESERVE_FLUSH_ALL;
5530                 goto again;
5531         }
5532         space_info->bytes_may_use -= num_bytes;
5533         trace_btrfs_space_reservation(fs_info, "space_info",
5534                                       space_info->flags, num_bytes, 0);
5535         spin_unlock(&space_info->lock);
5536 }
5537
5538 /*
5539  * This is for newly allocated space that isn't accounted in
5540  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5541  * we use this helper.
5542  */
5543 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5544                                      struct btrfs_space_info *space_info,
5545                                      u64 num_bytes)
5546 {
5547         struct reserve_ticket *ticket;
5548         struct list_head *head = &space_info->priority_tickets;
5549
5550 again:
5551         while (!list_empty(head) && num_bytes) {
5552                 ticket = list_first_entry(head, struct reserve_ticket,
5553                                           list);
5554                 if (num_bytes >= ticket->bytes) {
5555                         trace_btrfs_space_reservation(fs_info, "space_info",
5556                                                       space_info->flags,
5557                                                       ticket->bytes, 1);
5558                         list_del_init(&ticket->list);
5559                         num_bytes -= ticket->bytes;
5560                         space_info->bytes_may_use += ticket->bytes;
5561                         ticket->bytes = 0;
5562                         space_info->tickets_id++;
5563                         wake_up(&ticket->wait);
5564                 } else {
5565                         trace_btrfs_space_reservation(fs_info, "space_info",
5566                                                       space_info->flags,
5567                                                       num_bytes, 1);
5568                         space_info->bytes_may_use += num_bytes;
5569                         ticket->bytes -= num_bytes;
5570                         num_bytes = 0;
5571                 }
5572         }
5573
5574         if (num_bytes && head == &space_info->priority_tickets) {
5575                 head = &space_info->tickets;
5576                 goto again;
5577         }
5578 }
5579
5580 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5581                                     struct btrfs_block_rsv *block_rsv,
5582                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5583 {
5584         struct btrfs_space_info *space_info = block_rsv->space_info;
5585         u64 ret;
5586
5587         spin_lock(&block_rsv->lock);
5588         if (num_bytes == (u64)-1)
5589                 num_bytes = block_rsv->size;
5590         block_rsv->size -= num_bytes;
5591         if (block_rsv->reserved >= block_rsv->size) {
5592                 num_bytes = block_rsv->reserved - block_rsv->size;
5593                 block_rsv->reserved = block_rsv->size;
5594                 block_rsv->full = 1;
5595         } else {
5596                 num_bytes = 0;
5597         }
5598         spin_unlock(&block_rsv->lock);
5599
5600         ret = num_bytes;
5601         if (num_bytes > 0) {
5602                 if (dest) {
5603                         spin_lock(&dest->lock);
5604                         if (!dest->full) {
5605                                 u64 bytes_to_add;
5606
5607                                 bytes_to_add = dest->size - dest->reserved;
5608                                 bytes_to_add = min(num_bytes, bytes_to_add);
5609                                 dest->reserved += bytes_to_add;
5610                                 if (dest->reserved >= dest->size)
5611                                         dest->full = 1;
5612                                 num_bytes -= bytes_to_add;
5613                         }
5614                         spin_unlock(&dest->lock);
5615                 }
5616                 if (num_bytes)
5617                         space_info_add_old_bytes(fs_info, space_info,
5618                                                  num_bytes);
5619         }
5620         return ret;
5621 }
5622
5623 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5624                             struct btrfs_block_rsv *dst, u64 num_bytes,
5625                             int update_size)
5626 {
5627         int ret;
5628
5629         ret = block_rsv_use_bytes(src, num_bytes);
5630         if (ret)
5631                 return ret;
5632
5633         block_rsv_add_bytes(dst, num_bytes, update_size);
5634         return 0;
5635 }
5636
5637 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5638 {
5639         memset(rsv, 0, sizeof(*rsv));
5640         spin_lock_init(&rsv->lock);
5641         rsv->type = type;
5642 }
5643
5644 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5645                                    struct btrfs_block_rsv *rsv,
5646                                    unsigned short type)
5647 {
5648         btrfs_init_block_rsv(rsv, type);
5649         rsv->space_info = __find_space_info(fs_info,
5650                                             BTRFS_BLOCK_GROUP_METADATA);
5651 }
5652
5653 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5654                                               unsigned short type)
5655 {
5656         struct btrfs_block_rsv *block_rsv;
5657
5658         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5659         if (!block_rsv)
5660                 return NULL;
5661
5662         btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5663         return block_rsv;
5664 }
5665
5666 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5667                           struct btrfs_block_rsv *rsv)
5668 {
5669         if (!rsv)
5670                 return;
5671         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5672         kfree(rsv);
5673 }
5674
5675 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5676 {
5677         kfree(rsv);
5678 }
5679
5680 int btrfs_block_rsv_add(struct btrfs_root *root,
5681                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5682                         enum btrfs_reserve_flush_enum flush)
5683 {
5684         int ret;
5685
5686         if (num_bytes == 0)
5687                 return 0;
5688
5689         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5690         if (!ret) {
5691                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5692                 return 0;
5693         }
5694
5695         return ret;
5696 }
5697
5698 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5699 {
5700         u64 num_bytes = 0;
5701         int ret = -ENOSPC;
5702
5703         if (!block_rsv)
5704                 return 0;
5705
5706         spin_lock(&block_rsv->lock);
5707         num_bytes = div_factor(block_rsv->size, min_factor);
5708         if (block_rsv->reserved >= num_bytes)
5709                 ret = 0;
5710         spin_unlock(&block_rsv->lock);
5711
5712         return ret;
5713 }
5714
5715 int btrfs_block_rsv_refill(struct btrfs_root *root,
5716                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5717                            enum btrfs_reserve_flush_enum flush)
5718 {
5719         u64 num_bytes = 0;
5720         int ret = -ENOSPC;
5721
5722         if (!block_rsv)
5723                 return 0;
5724
5725         spin_lock(&block_rsv->lock);
5726         num_bytes = min_reserved;
5727         if (block_rsv->reserved >= num_bytes)
5728                 ret = 0;
5729         else
5730                 num_bytes -= block_rsv->reserved;
5731         spin_unlock(&block_rsv->lock);
5732
5733         if (!ret)
5734                 return 0;
5735
5736         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5737         if (!ret) {
5738                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5739                 return 0;
5740         }
5741
5742         return ret;
5743 }
5744
5745 /**
5746  * btrfs_inode_rsv_refill - refill the inode block rsv.
5747  * @inode - the inode we are refilling.
5748  * @flush - the flusing restriction.
5749  *
5750  * Essentially the same as btrfs_block_rsv_refill, except it uses the
5751  * block_rsv->size as the minimum size.  We'll either refill the missing amount
5752  * or return if we already have enough space.  This will also handle the resreve
5753  * tracepoint for the reserved amount.
5754  */
5755 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5756                                   enum btrfs_reserve_flush_enum flush)
5757 {
5758         struct btrfs_root *root = inode->root;
5759         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5760         u64 num_bytes = 0;
5761         int ret = -ENOSPC;
5762
5763         spin_lock(&block_rsv->lock);
5764         if (block_rsv->reserved < block_rsv->size)
5765                 num_bytes = block_rsv->size - block_rsv->reserved;
5766         spin_unlock(&block_rsv->lock);
5767
5768         if (num_bytes == 0)
5769                 return 0;
5770
5771         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5772         if (!ret) {
5773                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5774                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5775                                               btrfs_ino(inode), num_bytes, 1);
5776         }
5777         return ret;
5778 }
5779
5780 /**
5781  * btrfs_inode_rsv_release - release any excessive reservation.
5782  * @inode - the inode we need to release from.
5783  *
5784  * This is the same as btrfs_block_rsv_release, except that it handles the
5785  * tracepoint for the reservation.
5786  */
5787 static void btrfs_inode_rsv_release(struct btrfs_inode *inode)
5788 {
5789         struct btrfs_fs_info *fs_info = inode->root->fs_info;
5790         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5791         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5792         u64 released = 0;
5793
5794         /*
5795          * Since we statically set the block_rsv->size we just want to say we
5796          * are releasing 0 bytes, and then we'll just get the reservation over
5797          * the size free'd.
5798          */
5799         released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0);
5800         if (released > 0)
5801                 trace_btrfs_space_reservation(fs_info, "delalloc",
5802                                               btrfs_ino(inode), released, 0);
5803 }
5804
5805 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5806                              struct btrfs_block_rsv *block_rsv,
5807                              u64 num_bytes)
5808 {
5809         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5810
5811         if (global_rsv == block_rsv ||
5812             block_rsv->space_info != global_rsv->space_info)
5813                 global_rsv = NULL;
5814         block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes);
5815 }
5816
5817 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5818 {
5819         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5820         struct btrfs_space_info *sinfo = block_rsv->space_info;
5821         u64 num_bytes;
5822
5823         /*
5824          * The global block rsv is based on the size of the extent tree, the
5825          * checksum tree and the root tree.  If the fs is empty we want to set
5826          * it to a minimal amount for safety.
5827          */
5828         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5829                 btrfs_root_used(&fs_info->csum_root->root_item) +
5830                 btrfs_root_used(&fs_info->tree_root->root_item);
5831         num_bytes = max_t(u64, num_bytes, SZ_16M);
5832
5833         spin_lock(&sinfo->lock);
5834         spin_lock(&block_rsv->lock);
5835
5836         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5837
5838         if (block_rsv->reserved < block_rsv->size) {
5839                 num_bytes = btrfs_space_info_used(sinfo, true);
5840                 if (sinfo->total_bytes > num_bytes) {
5841                         num_bytes = sinfo->total_bytes - num_bytes;
5842                         num_bytes = min(num_bytes,
5843                                         block_rsv->size - block_rsv->reserved);
5844                         block_rsv->reserved += num_bytes;
5845                         sinfo->bytes_may_use += num_bytes;
5846                         trace_btrfs_space_reservation(fs_info, "space_info",
5847                                                       sinfo->flags, num_bytes,
5848                                                       1);
5849                 }
5850         } else if (block_rsv->reserved > block_rsv->size) {
5851                 num_bytes = block_rsv->reserved - block_rsv->size;
5852                 sinfo->bytes_may_use -= num_bytes;
5853                 trace_btrfs_space_reservation(fs_info, "space_info",
5854                                       sinfo->flags, num_bytes, 0);
5855                 block_rsv->reserved = block_rsv->size;
5856         }
5857
5858         if (block_rsv->reserved == block_rsv->size)
5859                 block_rsv->full = 1;
5860         else
5861                 block_rsv->full = 0;
5862
5863         spin_unlock(&block_rsv->lock);
5864         spin_unlock(&sinfo->lock);
5865 }
5866
5867 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5868 {
5869         struct btrfs_space_info *space_info;
5870
5871         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5872         fs_info->chunk_block_rsv.space_info = space_info;
5873
5874         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5875         fs_info->global_block_rsv.space_info = space_info;
5876         fs_info->trans_block_rsv.space_info = space_info;
5877         fs_info->empty_block_rsv.space_info = space_info;
5878         fs_info->delayed_block_rsv.space_info = space_info;
5879
5880         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5881         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5882         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5883         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5884         if (fs_info->quota_root)
5885                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5886         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5887
5888         update_global_block_rsv(fs_info);
5889 }
5890
5891 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5892 {
5893         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5894                                 (u64)-1);
5895         WARN_ON(fs_info->trans_block_rsv.size > 0);
5896         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5897         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5898         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5899         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5900         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5901 }
5902
5903
5904 /*
5905  * To be called after all the new block groups attached to the transaction
5906  * handle have been created (btrfs_create_pending_block_groups()).
5907  */
5908 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5909 {
5910         struct btrfs_fs_info *fs_info = trans->fs_info;
5911
5912         if (!trans->chunk_bytes_reserved)
5913                 return;
5914
5915         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5916
5917         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5918                                 trans->chunk_bytes_reserved);
5919         trans->chunk_bytes_reserved = 0;
5920 }
5921
5922 /* Can only return 0 or -ENOSPC */
5923 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5924                                   struct btrfs_inode *inode)
5925 {
5926         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5927         struct btrfs_root *root = inode->root;
5928         /*
5929          * We always use trans->block_rsv here as we will have reserved space
5930          * for our orphan when starting the transaction, using get_block_rsv()
5931          * here will sometimes make us choose the wrong block rsv as we could be
5932          * doing a reloc inode for a non refcounted root.
5933          */
5934         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5935         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5936
5937         /*
5938          * We need to hold space in order to delete our orphan item once we've
5939          * added it, so this takes the reservation so we can release it later
5940          * when we are truly done with the orphan item.
5941          */
5942         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5943
5944         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5945                         num_bytes, 1);
5946         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5947 }
5948
5949 void btrfs_orphan_release_metadata(struct btrfs_inode *inode)
5950 {
5951         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
5952         struct btrfs_root *root = inode->root;
5953         u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
5954
5955         trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode),
5956                         num_bytes, 0);
5957         btrfs_block_rsv_release(fs_info, root->orphan_block_rsv, num_bytes);
5958 }
5959
5960 /*
5961  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5962  * root: the root of the parent directory
5963  * rsv: block reservation
5964  * items: the number of items that we need do reservation
5965  * qgroup_reserved: used to return the reserved size in qgroup
5966  *
5967  * This function is used to reserve the space for snapshot/subvolume
5968  * creation and deletion. Those operations are different with the
5969  * common file/directory operations, they change two fs/file trees
5970  * and root tree, the number of items that the qgroup reserves is
5971  * different with the free space reservation. So we can not use
5972  * the space reservation mechanism in start_transaction().
5973  */
5974 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5975                                      struct btrfs_block_rsv *rsv,
5976                                      int items,
5977                                      u64 *qgroup_reserved,
5978                                      bool use_global_rsv)
5979 {
5980         u64 num_bytes;
5981         int ret;
5982         struct btrfs_fs_info *fs_info = root->fs_info;
5983         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5984
5985         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5986                 /* One for parent inode, two for dir entries */
5987                 num_bytes = 3 * fs_info->nodesize;
5988                 ret = btrfs_qgroup_reserve_meta(root, num_bytes, true);
5989                 if (ret)
5990                         return ret;
5991         } else {
5992                 num_bytes = 0;
5993         }
5994
5995         *qgroup_reserved = num_bytes;
5996
5997         num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5998         rsv->space_info = __find_space_info(fs_info,
5999                                             BTRFS_BLOCK_GROUP_METADATA);
6000         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
6001                                   BTRFS_RESERVE_FLUSH_ALL);
6002
6003         if (ret == -ENOSPC && use_global_rsv)
6004                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
6005
6006         if (ret && *qgroup_reserved)
6007                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
6008
6009         return ret;
6010 }
6011
6012 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
6013                                       struct btrfs_block_rsv *rsv)
6014 {
6015         btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
6016 }
6017
6018 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
6019                                                  struct btrfs_inode *inode)
6020 {
6021         struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
6022         u64 reserve_size = 0;
6023         u64 csum_leaves;
6024         unsigned outstanding_extents;
6025
6026         lockdep_assert_held(&inode->lock);
6027         outstanding_extents = inode->outstanding_extents;
6028         if (outstanding_extents)
6029                 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
6030                                                 outstanding_extents + 1);
6031         csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
6032                                                  inode->csum_bytes);
6033         reserve_size += btrfs_calc_trans_metadata_size(fs_info,
6034                                                        csum_leaves);
6035
6036         spin_lock(&block_rsv->lock);
6037         block_rsv->size = reserve_size;
6038         spin_unlock(&block_rsv->lock);
6039 }
6040
6041 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
6042 {
6043         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6044         struct btrfs_root *root = inode->root;
6045         unsigned nr_extents;
6046         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
6047         int ret = 0;
6048         bool delalloc_lock = true;
6049
6050         /* If we are a free space inode we need to not flush since we will be in
6051          * the middle of a transaction commit.  We also don't need the delalloc
6052          * mutex since we won't race with anybody.  We need this mostly to make
6053          * lockdep shut its filthy mouth.
6054          *
6055          * If we have a transaction open (can happen if we call truncate_block
6056          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
6057          */
6058         if (btrfs_is_free_space_inode(inode)) {
6059                 flush = BTRFS_RESERVE_NO_FLUSH;
6060                 delalloc_lock = false;
6061         } else {
6062                 if (current->journal_info)
6063                         flush = BTRFS_RESERVE_FLUSH_LIMIT;
6064
6065                 if (btrfs_transaction_in_commit(fs_info))
6066                         schedule_timeout(1);
6067         }
6068
6069         if (delalloc_lock)
6070                 mutex_lock(&inode->delalloc_mutex);
6071
6072         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6073
6074         /* Add our new extents and calculate the new rsv size. */
6075         spin_lock(&inode->lock);
6076         nr_extents = count_max_extents(num_bytes);
6077         btrfs_mod_outstanding_extents(inode, nr_extents);
6078         inode->csum_bytes += num_bytes;
6079         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6080         spin_unlock(&inode->lock);
6081
6082         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
6083                 ret = btrfs_qgroup_reserve_meta(root,
6084                                 nr_extents * fs_info->nodesize, true);
6085                 if (ret)
6086                         goto out_fail;
6087         }
6088
6089         ret = btrfs_inode_rsv_refill(inode, flush);
6090         if (unlikely(ret)) {
6091                 btrfs_qgroup_free_meta(root,
6092                                        nr_extents * fs_info->nodesize);
6093                 goto out_fail;
6094         }
6095
6096         if (delalloc_lock)
6097                 mutex_unlock(&inode->delalloc_mutex);
6098         return 0;
6099
6100 out_fail:
6101         spin_lock(&inode->lock);
6102         nr_extents = count_max_extents(num_bytes);
6103         btrfs_mod_outstanding_extents(inode, -nr_extents);
6104         inode->csum_bytes -= num_bytes;
6105         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6106         spin_unlock(&inode->lock);
6107
6108         btrfs_inode_rsv_release(inode);
6109         if (delalloc_lock)
6110                 mutex_unlock(&inode->delalloc_mutex);
6111         return ret;
6112 }
6113
6114 /**
6115  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6116  * @inode: the inode to release the reservation for.
6117  * @num_bytes: the number of bytes we are releasing.
6118  *
6119  * This will release the metadata reservation for an inode.  This can be called
6120  * once we complete IO for a given set of bytes to release their metadata
6121  * reservations, or on error for the same reason.
6122  */
6123 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes)
6124 {
6125         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6126
6127         num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6128         spin_lock(&inode->lock);
6129         inode->csum_bytes -= num_bytes;
6130         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6131         spin_unlock(&inode->lock);
6132
6133         if (btrfs_is_testing(fs_info))
6134                 return;
6135
6136         btrfs_inode_rsv_release(inode);
6137 }
6138
6139 /**
6140  * btrfs_delalloc_release_extents - release our outstanding_extents
6141  * @inode: the inode to balance the reservation for.
6142  * @num_bytes: the number of bytes we originally reserved with
6143  *
6144  * When we reserve space we increase outstanding_extents for the extents we may
6145  * add.  Once we've set the range as delalloc or created our ordered extents we
6146  * have outstanding_extents to track the real usage, so we use this to free our
6147  * temporarily tracked outstanding_extents.  This _must_ be used in conjunction
6148  * with btrfs_delalloc_reserve_metadata.
6149  */
6150 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes)
6151 {
6152         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6153         unsigned num_extents;
6154
6155         spin_lock(&inode->lock);
6156         num_extents = count_max_extents(num_bytes);
6157         btrfs_mod_outstanding_extents(inode, -num_extents);
6158         btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6159         spin_unlock(&inode->lock);
6160
6161         if (btrfs_is_testing(fs_info))
6162                 return;
6163
6164         btrfs_inode_rsv_release(inode);
6165 }
6166
6167 /**
6168  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6169  * delalloc
6170  * @inode: inode we're writing to
6171  * @start: start range we are writing to
6172  * @len: how long the range we are writing to
6173  * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6174  *            current reservation.
6175  *
6176  * This will do the following things
6177  *
6178  * o reserve space in data space info for num bytes
6179  *   and reserve precious corresponding qgroup space
6180  *   (Done in check_data_free_space)
6181  *
6182  * o reserve space for metadata space, based on the number of outstanding
6183  *   extents and how much csums will be needed
6184  *   also reserve metadata space in a per root over-reserve method.
6185  * o add to the inodes->delalloc_bytes
6186  * o add it to the fs_info's delalloc inodes list.
6187  *   (Above 3 all done in delalloc_reserve_metadata)
6188  *
6189  * Return 0 for success
6190  * Return <0 for error(-ENOSPC or -EQUOT)
6191  */
6192 int btrfs_delalloc_reserve_space(struct inode *inode,
6193                         struct extent_changeset **reserved, u64 start, u64 len)
6194 {
6195         int ret;
6196
6197         ret = btrfs_check_data_free_space(inode, reserved, start, len);
6198         if (ret < 0)
6199                 return ret;
6200         ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6201         if (ret < 0)
6202                 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6203         return ret;
6204 }
6205
6206 /**
6207  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6208  * @inode: inode we're releasing space for
6209  * @start: start position of the space already reserved
6210  * @len: the len of the space already reserved
6211  * @release_bytes: the len of the space we consumed or didn't use
6212  *
6213  * This function will release the metadata space that was not used and will
6214  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6215  * list if there are no delalloc bytes left.
6216  * Also it will handle the qgroup reserved space.
6217  */
6218 void btrfs_delalloc_release_space(struct inode *inode,
6219                                   struct extent_changeset *reserved,
6220                                   u64 start, u64 len)
6221 {
6222         btrfs_delalloc_release_metadata(BTRFS_I(inode), len);
6223         btrfs_free_reserved_data_space(inode, reserved, start, len);
6224 }
6225
6226 static int update_block_group(struct btrfs_trans_handle *trans,
6227                               struct btrfs_fs_info *info, u64 bytenr,
6228                               u64 num_bytes, int alloc)
6229 {
6230         struct btrfs_block_group_cache *cache = NULL;
6231         u64 total = num_bytes;
6232         u64 old_val;
6233         u64 byte_in_group;
6234         int factor;
6235
6236         /* block accounting for super block */
6237         spin_lock(&info->delalloc_root_lock);
6238         old_val = btrfs_super_bytes_used(info->super_copy);
6239         if (alloc)
6240                 old_val += num_bytes;
6241         else
6242                 old_val -= num_bytes;
6243         btrfs_set_super_bytes_used(info->super_copy, old_val);
6244         spin_unlock(&info->delalloc_root_lock);
6245
6246         while (total) {
6247                 cache = btrfs_lookup_block_group(info, bytenr);
6248                 if (!cache)
6249                         return -ENOENT;
6250                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6251                                     BTRFS_BLOCK_GROUP_RAID1 |
6252                                     BTRFS_BLOCK_GROUP_RAID10))
6253                         factor = 2;
6254                 else
6255                         factor = 1;
6256                 /*
6257                  * If this block group has free space cache written out, we
6258                  * need to make sure to load it if we are removing space.  This
6259                  * is because we need the unpinning stage to actually add the
6260                  * space back to the block group, otherwise we will leak space.
6261                  */
6262                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6263                         cache_block_group(cache, 1);
6264
6265                 byte_in_group = bytenr - cache->key.objectid;
6266                 WARN_ON(byte_in_group > cache->key.offset);
6267
6268                 spin_lock(&cache->space_info->lock);
6269                 spin_lock(&cache->lock);
6270
6271                 if (btrfs_test_opt(info, SPACE_CACHE) &&
6272                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6273                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6274
6275                 old_val = btrfs_block_group_used(&cache->item);
6276                 num_bytes = min(total, cache->key.offset - byte_in_group);
6277                 if (alloc) {
6278                         old_val += num_bytes;
6279                         btrfs_set_block_group_used(&cache->item, old_val);
6280                         cache->reserved -= num_bytes;
6281                         cache->space_info->bytes_reserved -= num_bytes;
6282                         cache->space_info->bytes_used += num_bytes;
6283                         cache->space_info->disk_used += num_bytes * factor;
6284                         spin_unlock(&cache->lock);
6285                         spin_unlock(&cache->space_info->lock);
6286                 } else {
6287                         old_val -= num_bytes;
6288                         btrfs_set_block_group_used(&cache->item, old_val);
6289                         cache->pinned += num_bytes;
6290                         cache->space_info->bytes_pinned += num_bytes;
6291                         cache->space_info->bytes_used -= num_bytes;
6292                         cache->space_info->disk_used -= num_bytes * factor;
6293                         spin_unlock(&cache->lock);
6294                         spin_unlock(&cache->space_info->lock);
6295
6296                         trace_btrfs_space_reservation(info, "pinned",
6297                                                       cache->space_info->flags,
6298                                                       num_bytes, 1);
6299                         percpu_counter_add(&cache->space_info->total_bytes_pinned,
6300                                            num_bytes);
6301                         set_extent_dirty(info->pinned_extents,
6302                                          bytenr, bytenr + num_bytes - 1,
6303                                          GFP_NOFS | __GFP_NOFAIL);
6304                 }
6305
6306                 spin_lock(&trans->transaction->dirty_bgs_lock);
6307                 if (list_empty(&cache->dirty_list)) {
6308                         list_add_tail(&cache->dirty_list,
6309                                       &trans->transaction->dirty_bgs);
6310                                 trans->transaction->num_dirty_bgs++;
6311                         btrfs_get_block_group(cache);
6312                 }
6313                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6314
6315                 /*
6316                  * No longer have used bytes in this block group, queue it for
6317                  * deletion. We do this after adding the block group to the
6318                  * dirty list to avoid races between cleaner kthread and space
6319                  * cache writeout.
6320                  */
6321                 if (!alloc && old_val == 0) {
6322                         spin_lock(&info->unused_bgs_lock);
6323                         if (list_empty(&cache->bg_list)) {
6324                                 btrfs_get_block_group(cache);
6325                                 list_add_tail(&cache->bg_list,
6326                                               &info->unused_bgs);
6327                         }
6328                         spin_unlock(&info->unused_bgs_lock);
6329                 }
6330
6331                 btrfs_put_block_group(cache);
6332                 total -= num_bytes;
6333                 bytenr += num_bytes;
6334         }
6335         return 0;
6336 }
6337
6338 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6339 {
6340         struct btrfs_block_group_cache *cache;
6341         u64 bytenr;
6342
6343         spin_lock(&fs_info->block_group_cache_lock);
6344         bytenr = fs_info->first_logical_byte;
6345         spin_unlock(&fs_info->block_group_cache_lock);
6346
6347         if (bytenr < (u64)-1)
6348                 return bytenr;
6349
6350         cache = btrfs_lookup_first_block_group(fs_info, search_start);
6351         if (!cache)
6352                 return 0;
6353
6354         bytenr = cache->key.objectid;
6355         btrfs_put_block_group(cache);
6356
6357         return bytenr;
6358 }
6359
6360 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6361                            struct btrfs_block_group_cache *cache,
6362                            u64 bytenr, u64 num_bytes, int reserved)
6363 {
6364         spin_lock(&cache->space_info->lock);
6365         spin_lock(&cache->lock);
6366         cache->pinned += num_bytes;
6367         cache->space_info->bytes_pinned += num_bytes;
6368         if (reserved) {
6369                 cache->reserved -= num_bytes;
6370                 cache->space_info->bytes_reserved -= num_bytes;
6371         }
6372         spin_unlock(&cache->lock);
6373         spin_unlock(&cache->space_info->lock);
6374
6375         trace_btrfs_space_reservation(fs_info, "pinned",
6376                                       cache->space_info->flags, num_bytes, 1);
6377         percpu_counter_add(&cache->space_info->total_bytes_pinned, num_bytes);
6378         set_extent_dirty(fs_info->pinned_extents, bytenr,
6379                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6380         return 0;
6381 }
6382
6383 /*
6384  * this function must be called within transaction
6385  */
6386 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6387                      u64 bytenr, u64 num_bytes, int reserved)
6388 {
6389         struct btrfs_block_group_cache *cache;
6390
6391         cache = btrfs_lookup_block_group(fs_info, bytenr);
6392         BUG_ON(!cache); /* Logic error */
6393
6394         pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6395
6396         btrfs_put_block_group(cache);
6397         return 0;
6398 }
6399
6400 /*
6401  * this function must be called within transaction
6402  */
6403 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6404                                     u64 bytenr, u64 num_bytes)
6405 {
6406         struct btrfs_block_group_cache *cache;
6407         int ret;
6408
6409         cache = btrfs_lookup_block_group(fs_info, bytenr);
6410         if (!cache)
6411                 return -EINVAL;
6412
6413         /*
6414          * pull in the free space cache (if any) so that our pin
6415          * removes the free space from the cache.  We have load_only set
6416          * to one because the slow code to read in the free extents does check
6417          * the pinned extents.
6418          */
6419         cache_block_group(cache, 1);
6420
6421         pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6422
6423         /* remove us from the free space cache (if we're there at all) */
6424         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6425         btrfs_put_block_group(cache);
6426         return ret;
6427 }
6428
6429 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6430                                    u64 start, u64 num_bytes)
6431 {
6432         int ret;
6433         struct btrfs_block_group_cache *block_group;
6434         struct btrfs_caching_control *caching_ctl;
6435
6436         block_group = btrfs_lookup_block_group(fs_info, start);
6437         if (!block_group)
6438                 return -EINVAL;
6439
6440         cache_block_group(block_group, 0);
6441         caching_ctl = get_caching_control(block_group);
6442
6443         if (!caching_ctl) {
6444                 /* Logic error */
6445                 BUG_ON(!block_group_cache_done(block_group));
6446                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6447         } else {
6448                 mutex_lock(&caching_ctl->mutex);
6449
6450                 if (start >= caching_ctl->progress) {
6451                         ret = add_excluded_extent(fs_info, start, num_bytes);
6452                 } else if (start + num_bytes <= caching_ctl->progress) {
6453                         ret = btrfs_remove_free_space(block_group,
6454                                                       start, num_bytes);
6455                 } else {
6456                         num_bytes = caching_ctl->progress - start;
6457                         ret = btrfs_remove_free_space(block_group,
6458                                                       start, num_bytes);
6459                         if (ret)
6460                                 goto out_lock;
6461
6462                         num_bytes = (start + num_bytes) -
6463                                 caching_ctl->progress;
6464                         start = caching_ctl->progress;
6465                         ret = add_excluded_extent(fs_info, start, num_bytes);
6466                 }
6467 out_lock:
6468                 mutex_unlock(&caching_ctl->mutex);
6469                 put_caching_control(caching_ctl);
6470         }
6471         btrfs_put_block_group(block_group);
6472         return ret;
6473 }
6474
6475 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6476                                  struct extent_buffer *eb)
6477 {
6478         struct btrfs_file_extent_item *item;
6479         struct btrfs_key key;
6480         int found_type;
6481         int i;
6482
6483         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6484                 return 0;
6485
6486         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6487                 btrfs_item_key_to_cpu(eb, &key, i);
6488                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6489                         continue;
6490                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6491                 found_type = btrfs_file_extent_type(eb, item);
6492                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6493                         continue;
6494                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6495                         continue;
6496                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6497                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6498                 __exclude_logged_extent(fs_info, key.objectid, key.offset);
6499         }
6500
6501         return 0;
6502 }
6503
6504 static void
6505 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6506 {
6507         atomic_inc(&bg->reservations);
6508 }
6509
6510 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6511                                         const u64 start)
6512 {
6513         struct btrfs_block_group_cache *bg;
6514
6515         bg = btrfs_lookup_block_group(fs_info, start);
6516         ASSERT(bg);
6517         if (atomic_dec_and_test(&bg->reservations))
6518                 wake_up_atomic_t(&bg->reservations);
6519         btrfs_put_block_group(bg);
6520 }
6521
6522 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6523 {
6524         struct btrfs_space_info *space_info = bg->space_info;
6525
6526         ASSERT(bg->ro);
6527
6528         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6529                 return;
6530
6531         /*
6532          * Our block group is read only but before we set it to read only,
6533          * some task might have had allocated an extent from it already, but it
6534          * has not yet created a respective ordered extent (and added it to a
6535          * root's list of ordered extents).
6536          * Therefore wait for any task currently allocating extents, since the
6537          * block group's reservations counter is incremented while a read lock
6538          * on the groups' semaphore is held and decremented after releasing
6539          * the read access on that semaphore and creating the ordered extent.
6540          */
6541         down_write(&space_info->groups_sem);
6542         up_write(&space_info->groups_sem);
6543
6544         wait_on_atomic_t(&bg->reservations, atomic_t_wait,
6545                          TASK_UNINTERRUPTIBLE);
6546 }
6547
6548 /**
6549  * btrfs_add_reserved_bytes - update the block_group and space info counters
6550  * @cache:      The cache we are manipulating
6551  * @ram_bytes:  The number of bytes of file content, and will be same to
6552  *              @num_bytes except for the compress path.
6553  * @num_bytes:  The number of bytes in question
6554  * @delalloc:   The blocks are allocated for the delalloc write
6555  *
6556  * This is called by the allocator when it reserves space. If this is a
6557  * reservation and the block group has become read only we cannot make the
6558  * reservation and return -EAGAIN, otherwise this function always succeeds.
6559  */
6560 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6561                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6562 {
6563         struct btrfs_space_info *space_info = cache->space_info;
6564         int ret = 0;
6565
6566         spin_lock(&space_info->lock);
6567         spin_lock(&cache->lock);
6568         if (cache->ro) {
6569                 ret = -EAGAIN;
6570         } else {
6571                 cache->reserved += num_bytes;
6572                 space_info->bytes_reserved += num_bytes;
6573
6574                 trace_btrfs_space_reservation(cache->fs_info,
6575                                 "space_info", space_info->flags,
6576                                 ram_bytes, 0);
6577                 space_info->bytes_may_use -= ram_bytes;
6578                 if (delalloc)
6579                         cache->delalloc_bytes += num_bytes;
6580         }
6581         spin_unlock(&cache->lock);
6582         spin_unlock(&space_info->lock);
6583         return ret;
6584 }
6585
6586 /**
6587  * btrfs_free_reserved_bytes - update the block_group and space info counters
6588  * @cache:      The cache we are manipulating
6589  * @num_bytes:  The number of bytes in question
6590  * @delalloc:   The blocks are allocated for the delalloc write
6591  *
6592  * This is called by somebody who is freeing space that was never actually used
6593  * on disk.  For example if you reserve some space for a new leaf in transaction
6594  * A and before transaction A commits you free that leaf, you call this with
6595  * reserve set to 0 in order to clear the reservation.
6596  */
6597
6598 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6599                                      u64 num_bytes, int delalloc)
6600 {
6601         struct btrfs_space_info *space_info = cache->space_info;
6602         int ret = 0;
6603
6604         spin_lock(&space_info->lock);
6605         spin_lock(&cache->lock);
6606         if (cache->ro)
6607                 space_info->bytes_readonly += num_bytes;
6608         cache->reserved -= num_bytes;
6609         space_info->bytes_reserved -= num_bytes;
6610
6611         if (delalloc)
6612                 cache->delalloc_bytes -= num_bytes;
6613         spin_unlock(&cache->lock);
6614         spin_unlock(&space_info->lock);
6615         return ret;
6616 }
6617 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6618 {
6619         struct btrfs_caching_control *next;
6620         struct btrfs_caching_control *caching_ctl;
6621         struct btrfs_block_group_cache *cache;
6622
6623         down_write(&fs_info->commit_root_sem);
6624
6625         list_for_each_entry_safe(caching_ctl, next,
6626                                  &fs_info->caching_block_groups, list) {
6627                 cache = caching_ctl->block_group;
6628                 if (block_group_cache_done(cache)) {
6629                         cache->last_byte_to_unpin = (u64)-1;
6630                         list_del_init(&caching_ctl->list);
6631                         put_caching_control(caching_ctl);
6632                 } else {
6633                         cache->last_byte_to_unpin = caching_ctl->progress;
6634                 }
6635         }
6636
6637         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6638                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6639         else
6640                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6641
6642         up_write(&fs_info->commit_root_sem);
6643
6644         update_global_block_rsv(fs_info);
6645 }
6646
6647 /*
6648  * Returns the free cluster for the given space info and sets empty_cluster to
6649  * what it should be based on the mount options.
6650  */
6651 static struct btrfs_free_cluster *
6652 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6653                    struct btrfs_space_info *space_info, u64 *empty_cluster)
6654 {
6655         struct btrfs_free_cluster *ret = NULL;
6656
6657         *empty_cluster = 0;
6658         if (btrfs_mixed_space_info(space_info))
6659                 return ret;
6660
6661         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6662                 ret = &fs_info->meta_alloc_cluster;
6663                 if (btrfs_test_opt(fs_info, SSD))
6664                         *empty_cluster = SZ_2M;
6665                 else
6666                         *empty_cluster = SZ_64K;
6667         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6668                    btrfs_test_opt(fs_info, SSD_SPREAD)) {
6669                 *empty_cluster = SZ_2M;
6670                 ret = &fs_info->data_alloc_cluster;
6671         }
6672
6673         return ret;
6674 }
6675
6676 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6677                               u64 start, u64 end,
6678                               const bool return_free_space)
6679 {
6680         struct btrfs_block_group_cache *cache = NULL;
6681         struct btrfs_space_info *space_info;
6682         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6683         struct btrfs_free_cluster *cluster = NULL;
6684         u64 len;
6685         u64 total_unpinned = 0;
6686         u64 empty_cluster = 0;
6687         bool readonly;
6688
6689         while (start <= end) {
6690                 readonly = false;
6691                 if (!cache ||
6692                     start >= cache->key.objectid + cache->key.offset) {
6693                         if (cache)
6694                                 btrfs_put_block_group(cache);
6695                         total_unpinned = 0;
6696                         cache = btrfs_lookup_block_group(fs_info, start);
6697                         BUG_ON(!cache); /* Logic error */
6698
6699                         cluster = fetch_cluster_info(fs_info,
6700                                                      cache->space_info,
6701                                                      &empty_cluster);
6702                         empty_cluster <<= 1;
6703                 }
6704
6705                 len = cache->key.objectid + cache->key.offset - start;
6706                 len = min(len, end + 1 - start);
6707
6708                 if (start < cache->last_byte_to_unpin) {
6709                         len = min(len, cache->last_byte_to_unpin - start);
6710                         if (return_free_space)
6711                                 btrfs_add_free_space(cache, start, len);
6712                 }
6713
6714                 start += len;
6715                 total_unpinned += len;
6716                 space_info = cache->space_info;
6717
6718                 /*
6719                  * If this space cluster has been marked as fragmented and we've
6720                  * unpinned enough in this block group to potentially allow a
6721                  * cluster to be created inside of it go ahead and clear the
6722                  * fragmented check.
6723                  */
6724                 if (cluster && cluster->fragmented &&
6725                     total_unpinned > empty_cluster) {
6726                         spin_lock(&cluster->lock);
6727                         cluster->fragmented = 0;
6728                         spin_unlock(&cluster->lock);
6729                 }
6730
6731                 spin_lock(&space_info->lock);
6732                 spin_lock(&cache->lock);
6733                 cache->pinned -= len;
6734                 space_info->bytes_pinned -= len;
6735
6736                 trace_btrfs_space_reservation(fs_info, "pinned",
6737                                               space_info->flags, len, 0);
6738                 space_info->max_extent_size = 0;
6739                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6740                 if (cache->ro) {
6741                         space_info->bytes_readonly += len;
6742                         readonly = true;
6743                 }
6744                 spin_unlock(&cache->lock);
6745                 if (!readonly && return_free_space &&
6746                     global_rsv->space_info == space_info) {
6747                         u64 to_add = len;
6748
6749                         spin_lock(&global_rsv->lock);
6750                         if (!global_rsv->full) {
6751                                 to_add = min(len, global_rsv->size -
6752                                              global_rsv->reserved);
6753                                 global_rsv->reserved += to_add;
6754                                 space_info->bytes_may_use += to_add;
6755                                 if (global_rsv->reserved >= global_rsv->size)
6756                                         global_rsv->full = 1;
6757                                 trace_btrfs_space_reservation(fs_info,
6758                                                               "space_info",
6759                                                               space_info->flags,
6760                                                               to_add, 1);
6761                                 len -= to_add;
6762                         }
6763                         spin_unlock(&global_rsv->lock);
6764                         /* Add to any tickets we may have */
6765                         if (len)
6766                                 space_info_add_new_bytes(fs_info, space_info,
6767                                                          len);
6768                 }
6769                 spin_unlock(&space_info->lock);
6770         }
6771
6772         if (cache)
6773                 btrfs_put_block_group(cache);
6774         return 0;
6775 }
6776
6777 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6778                                struct btrfs_fs_info *fs_info)
6779 {
6780         struct btrfs_block_group_cache *block_group, *tmp;
6781         struct list_head *deleted_bgs;
6782         struct extent_io_tree *unpin;
6783         u64 start;
6784         u64 end;
6785         int ret;
6786
6787         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6788                 unpin = &fs_info->freed_extents[1];
6789         else
6790                 unpin = &fs_info->freed_extents[0];
6791
6792         while (!trans->aborted) {
6793                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6794                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6795                                             EXTENT_DIRTY, NULL);
6796                 if (ret) {
6797                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6798                         break;
6799                 }
6800
6801                 if (btrfs_test_opt(fs_info, DISCARD))
6802                         ret = btrfs_discard_extent(fs_info, start,
6803                                                    end + 1 - start, NULL);
6804
6805                 clear_extent_dirty(unpin, start, end);
6806                 unpin_extent_range(fs_info, start, end, true);
6807                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6808                 cond_resched();
6809         }
6810
6811         /*
6812          * Transaction is finished.  We don't need the lock anymore.  We
6813          * do need to clean up the block groups in case of a transaction
6814          * abort.
6815          */
6816         deleted_bgs = &trans->transaction->deleted_bgs;
6817         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6818                 u64 trimmed = 0;
6819
6820                 ret = -EROFS;
6821                 if (!trans->aborted)
6822                         ret = btrfs_discard_extent(fs_info,
6823                                                    block_group->key.objectid,
6824                                                    block_group->key.offset,
6825                                                    &trimmed);
6826
6827                 list_del_init(&block_group->bg_list);
6828                 btrfs_put_block_group_trimming(block_group);
6829                 btrfs_put_block_group(block_group);
6830
6831                 if (ret) {
6832                         const char *errstr = btrfs_decode_error(ret);
6833                         btrfs_warn(fs_info,
6834                            "discard failed while removing blockgroup: errno=%d %s",
6835                                    ret, errstr);
6836                 }
6837         }
6838
6839         return 0;
6840 }
6841
6842 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6843                                 struct btrfs_fs_info *info,
6844                                 struct btrfs_delayed_ref_node *node, u64 parent,
6845                                 u64 root_objectid, u64 owner_objectid,
6846                                 u64 owner_offset, int refs_to_drop,
6847                                 struct btrfs_delayed_extent_op *extent_op)
6848 {
6849         struct btrfs_key key;
6850         struct btrfs_path *path;
6851         struct btrfs_root *extent_root = info->extent_root;
6852         struct extent_buffer *leaf;
6853         struct btrfs_extent_item *ei;
6854         struct btrfs_extent_inline_ref *iref;
6855         int ret;
6856         int is_data;
6857         int extent_slot = 0;
6858         int found_extent = 0;
6859         int num_to_del = 1;
6860         u32 item_size;
6861         u64 refs;
6862         u64 bytenr = node->bytenr;
6863         u64 num_bytes = node->num_bytes;
6864         int last_ref = 0;
6865         bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6866
6867         path = btrfs_alloc_path();
6868         if (!path)
6869                 return -ENOMEM;
6870
6871         path->reada = READA_FORWARD;
6872         path->leave_spinning = 1;
6873
6874         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6875         BUG_ON(!is_data && refs_to_drop != 1);
6876
6877         if (is_data)
6878                 skinny_metadata = false;
6879
6880         ret = lookup_extent_backref(trans, info, path, &iref,
6881                                     bytenr, num_bytes, parent,
6882                                     root_objectid, owner_objectid,
6883                                     owner_offset);
6884         if (ret == 0) {
6885                 extent_slot = path->slots[0];
6886                 while (extent_slot >= 0) {
6887                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6888                                               extent_slot);
6889                         if (key.objectid != bytenr)
6890                                 break;
6891                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6892                             key.offset == num_bytes) {
6893                                 found_extent = 1;
6894                                 break;
6895                         }
6896                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6897                             key.offset == owner_objectid) {
6898                                 found_extent = 1;
6899                                 break;
6900                         }
6901                         if (path->slots[0] - extent_slot > 5)
6902                                 break;
6903                         extent_slot--;
6904                 }
6905 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6906                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6907                 if (found_extent && item_size < sizeof(*ei))
6908                         found_extent = 0;
6909 #endif
6910                 if (!found_extent) {
6911                         BUG_ON(iref);
6912                         ret = remove_extent_backref(trans, info, path, NULL,
6913                                                     refs_to_drop,
6914                                                     is_data, &last_ref);
6915                         if (ret) {
6916                                 btrfs_abort_transaction(trans, ret);
6917                                 goto out;
6918                         }
6919                         btrfs_release_path(path);
6920                         path->leave_spinning = 1;
6921
6922                         key.objectid = bytenr;
6923                         key.type = BTRFS_EXTENT_ITEM_KEY;
6924                         key.offset = num_bytes;
6925
6926                         if (!is_data && skinny_metadata) {
6927                                 key.type = BTRFS_METADATA_ITEM_KEY;
6928                                 key.offset = owner_objectid;
6929                         }
6930
6931                         ret = btrfs_search_slot(trans, extent_root,
6932                                                 &key, path, -1, 1);
6933                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6934                                 /*
6935                                  * Couldn't find our skinny metadata item,
6936                                  * see if we have ye olde extent item.
6937                                  */
6938                                 path->slots[0]--;
6939                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6940                                                       path->slots[0]);
6941                                 if (key.objectid == bytenr &&
6942                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6943                                     key.offset == num_bytes)
6944                                         ret = 0;
6945                         }
6946
6947                         if (ret > 0 && skinny_metadata) {
6948                                 skinny_metadata = false;
6949                                 key.objectid = bytenr;
6950                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6951                                 key.offset = num_bytes;
6952                                 btrfs_release_path(path);
6953                                 ret = btrfs_search_slot(trans, extent_root,
6954                                                         &key, path, -1, 1);
6955                         }
6956
6957                         if (ret) {
6958                                 btrfs_err(info,
6959                                           "umm, got %d back from search, was looking for %llu",
6960                                           ret, bytenr);
6961                                 if (ret > 0)
6962                                         btrfs_print_leaf(path->nodes[0]);
6963                         }
6964                         if (ret < 0) {
6965                                 btrfs_abort_transaction(trans, ret);
6966                                 goto out;
6967                         }
6968                         extent_slot = path->slots[0];
6969                 }
6970         } else if (WARN_ON(ret == -ENOENT)) {
6971                 btrfs_print_leaf(path->nodes[0]);
6972                 btrfs_err(info,
6973                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6974                         bytenr, parent, root_objectid, owner_objectid,
6975                         owner_offset);
6976                 btrfs_abort_transaction(trans, ret);
6977                 goto out;
6978         } else {
6979                 btrfs_abort_transaction(trans, ret);
6980                 goto out;
6981         }
6982
6983         leaf = path->nodes[0];
6984         item_size = btrfs_item_size_nr(leaf, extent_slot);
6985 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6986         if (item_size < sizeof(*ei)) {
6987                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6988                 ret = convert_extent_item_v0(trans, info, path, owner_objectid,
6989                                              0);
6990                 if (ret < 0) {
6991                         btrfs_abort_transaction(trans, ret);
6992                         goto out;
6993                 }
6994
6995                 btrfs_release_path(path);
6996                 path->leave_spinning = 1;
6997
6998                 key.objectid = bytenr;
6999                 key.type = BTRFS_EXTENT_ITEM_KEY;
7000                 key.offset = num_bytes;
7001
7002                 ret = btrfs_search_slot(trans, extent_root, &key, path,
7003                                         -1, 1);
7004                 if (ret) {
7005                         btrfs_err(info,
7006                                   "umm, got %d back from search, was looking for %llu",
7007                                 ret, bytenr);
7008                         btrfs_print_leaf(path->nodes[0]);
7009                 }
7010                 if (ret < 0) {
7011                         btrfs_abort_transaction(trans, ret);
7012                         goto out;
7013                 }
7014
7015                 extent_slot = path->slots[0];
7016                 leaf = path->nodes[0];
7017                 item_size = btrfs_item_size_nr(leaf, extent_slot);
7018         }
7019 #endif
7020         BUG_ON(item_size < sizeof(*ei));
7021         ei = btrfs_item_ptr(leaf, extent_slot,
7022                             struct btrfs_extent_item);
7023         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7024             key.type == BTRFS_EXTENT_ITEM_KEY) {
7025                 struct btrfs_tree_block_info *bi;
7026                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7027                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7028                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7029         }
7030
7031         refs = btrfs_extent_refs(leaf, ei);
7032         if (refs < refs_to_drop) {
7033                 btrfs_err(info,
7034                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7035                           refs_to_drop, refs, bytenr);
7036                 ret = -EINVAL;
7037                 btrfs_abort_transaction(trans, ret);
7038                 goto out;
7039         }
7040         refs -= refs_to_drop;
7041
7042         if (refs > 0) {
7043                 if (extent_op)
7044                         __run_delayed_extent_op(extent_op, leaf, ei);
7045                 /*
7046                  * In the case of inline back ref, reference count will
7047                  * be updated by remove_extent_backref
7048                  */
7049                 if (iref) {
7050                         BUG_ON(!found_extent);
7051                 } else {
7052                         btrfs_set_extent_refs(leaf, ei, refs);
7053                         btrfs_mark_buffer_dirty(leaf);
7054                 }
7055                 if (found_extent) {
7056                         ret = remove_extent_backref(trans, info, path,
7057                                                     iref, refs_to_drop,
7058                                                     is_data, &last_ref);
7059                         if (ret) {
7060                                 btrfs_abort_transaction(trans, ret);
7061                                 goto out;
7062                         }
7063                 }
7064         } else {
7065                 if (found_extent) {
7066                         BUG_ON(is_data && refs_to_drop !=
7067                                extent_data_ref_count(path, iref));
7068                         if (iref) {
7069                                 BUG_ON(path->slots[0] != extent_slot);
7070                         } else {
7071                                 BUG_ON(path->slots[0] != extent_slot + 1);
7072                                 path->slots[0] = extent_slot;
7073                                 num_to_del = 2;
7074                         }
7075                 }
7076
7077                 last_ref = 1;
7078                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7079                                       num_to_del);
7080                 if (ret) {
7081                         btrfs_abort_transaction(trans, ret);
7082                         goto out;
7083                 }
7084                 btrfs_release_path(path);
7085
7086                 if (is_data) {
7087                         ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
7088                         if (ret) {
7089                                 btrfs_abort_transaction(trans, ret);
7090                                 goto out;
7091                         }
7092                 }
7093
7094                 ret = add_to_free_space_tree(trans, info, bytenr, num_bytes);
7095                 if (ret) {
7096                         btrfs_abort_transaction(trans, ret);
7097                         goto out;
7098                 }
7099
7100                 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
7101                 if (ret) {
7102                         btrfs_abort_transaction(trans, ret);
7103                         goto out;
7104                 }
7105         }
7106         btrfs_release_path(path);
7107
7108 out:
7109         btrfs_free_path(path);
7110         return ret;
7111 }
7112
7113 /*
7114  * when we free an block, it is possible (and likely) that we free the last
7115  * delayed ref for that extent as well.  This searches the delayed ref tree for
7116  * a given extent, and if there are no other delayed refs to be processed, it
7117  * removes it from the tree.
7118  */
7119 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7120                                       u64 bytenr)
7121 {
7122         struct btrfs_delayed_ref_head *head;
7123         struct btrfs_delayed_ref_root *delayed_refs;
7124         int ret = 0;
7125
7126         delayed_refs = &trans->transaction->delayed_refs;
7127         spin_lock(&delayed_refs->lock);
7128         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7129         if (!head)
7130                 goto out_delayed_unlock;
7131
7132         spin_lock(&head->lock);
7133         if (!RB_EMPTY_ROOT(&head->ref_tree))
7134                 goto out;
7135
7136         if (head->extent_op) {
7137                 if (!head->must_insert_reserved)
7138                         goto out;
7139                 btrfs_free_delayed_extent_op(head->extent_op);
7140                 head->extent_op = NULL;
7141         }
7142
7143         /*
7144          * waiting for the lock here would deadlock.  If someone else has it
7145          * locked they are already in the process of dropping it anyway
7146          */
7147         if (!mutex_trylock(&head->mutex))
7148                 goto out;
7149
7150         /*
7151          * at this point we have a head with no other entries.  Go
7152          * ahead and process it.
7153          */
7154         rb_erase(&head->href_node, &delayed_refs->href_root);
7155         RB_CLEAR_NODE(&head->href_node);
7156         atomic_dec(&delayed_refs->num_entries);
7157
7158         /*
7159          * we don't take a ref on the node because we're removing it from the
7160          * tree, so we just steal the ref the tree was holding.
7161          */
7162         delayed_refs->num_heads--;
7163         if (head->processing == 0)
7164                 delayed_refs->num_heads_ready--;
7165         head->processing = 0;
7166         spin_unlock(&head->lock);
7167         spin_unlock(&delayed_refs->lock);
7168
7169         BUG_ON(head->extent_op);
7170         if (head->must_insert_reserved)
7171                 ret = 1;
7172
7173         mutex_unlock(&head->mutex);
7174         btrfs_put_delayed_ref_head(head);
7175         return ret;
7176 out:
7177         spin_unlock(&head->lock);
7178
7179 out_delayed_unlock:
7180         spin_unlock(&delayed_refs->lock);
7181         return 0;
7182 }
7183
7184 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7185                            struct btrfs_root *root,
7186                            struct extent_buffer *buf,
7187                            u64 parent, int last_ref)
7188 {
7189         struct btrfs_fs_info *fs_info = root->fs_info;
7190         int pin = 1;
7191         int ret;
7192
7193         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7194                 int old_ref_mod, new_ref_mod;
7195
7196                 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7197                                    root->root_key.objectid,
7198                                    btrfs_header_level(buf), 0,
7199                                    BTRFS_DROP_DELAYED_REF);
7200                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, buf->start,
7201                                                  buf->len, parent,
7202                                                  root->root_key.objectid,
7203                                                  btrfs_header_level(buf),
7204                                                  BTRFS_DROP_DELAYED_REF, NULL,
7205                                                  &old_ref_mod, &new_ref_mod);
7206                 BUG_ON(ret); /* -ENOMEM */
7207                 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7208         }
7209
7210         if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7211                 struct btrfs_block_group_cache *cache;
7212
7213                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7214                         ret = check_ref_cleanup(trans, buf->start);
7215                         if (!ret)
7216                                 goto out;
7217                 }
7218
7219                 pin = 0;
7220                 cache = btrfs_lookup_block_group(fs_info, buf->start);
7221
7222                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7223                         pin_down_extent(fs_info, cache, buf->start,
7224                                         buf->len, 1);
7225                         btrfs_put_block_group(cache);
7226                         goto out;
7227                 }
7228
7229                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7230
7231                 btrfs_add_free_space(cache, buf->start, buf->len);
7232                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7233                 btrfs_put_block_group(cache);
7234                 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7235         }
7236 out:
7237         if (pin)
7238                 add_pinned_bytes(fs_info, buf->len, btrfs_header_level(buf),
7239                                  root->root_key.objectid);
7240
7241         if (last_ref) {
7242                 /*
7243                  * Deleting the buffer, clear the corrupt flag since it doesn't
7244                  * matter anymore.
7245                  */
7246                 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7247         }
7248 }
7249
7250 /* Can return -ENOMEM */
7251 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7252                       struct btrfs_root *root,
7253                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7254                       u64 owner, u64 offset)
7255 {
7256         struct btrfs_fs_info *fs_info = root->fs_info;
7257         int old_ref_mod, new_ref_mod;
7258         int ret;
7259
7260         if (btrfs_is_testing(fs_info))
7261                 return 0;
7262
7263         if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7264                 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7265                                    root_objectid, owner, offset,
7266                                    BTRFS_DROP_DELAYED_REF);
7267
7268         /*
7269          * tree log blocks never actually go into the extent allocation
7270          * tree, just update pinning info and exit early.
7271          */
7272         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7273                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7274                 /* unlocks the pinned mutex */
7275                 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7276                 old_ref_mod = new_ref_mod = 0;
7277                 ret = 0;
7278         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7279                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7280                                                  num_bytes, parent,
7281                                                  root_objectid, (int)owner,
7282                                                  BTRFS_DROP_DELAYED_REF, NULL,
7283                                                  &old_ref_mod, &new_ref_mod);
7284         } else {
7285                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7286                                                  num_bytes, parent,
7287                                                  root_objectid, owner, offset,
7288                                                  0, BTRFS_DROP_DELAYED_REF,
7289                                                  &old_ref_mod, &new_ref_mod);
7290         }
7291
7292         if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0)
7293                 add_pinned_bytes(fs_info, num_bytes, owner, root_objectid);
7294
7295         return ret;
7296 }
7297
7298 /*
7299  * when we wait for progress in the block group caching, its because
7300  * our allocation attempt failed at least once.  So, we must sleep
7301  * and let some progress happen before we try again.
7302  *
7303  * This function will sleep at least once waiting for new free space to
7304  * show up, and then it will check the block group free space numbers
7305  * for our min num_bytes.  Another option is to have it go ahead
7306  * and look in the rbtree for a free extent of a given size, but this
7307  * is a good start.
7308  *
7309  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7310  * any of the information in this block group.
7311  */
7312 static noinline void
7313 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7314                                 u64 num_bytes)
7315 {
7316         struct btrfs_caching_control *caching_ctl;
7317
7318         caching_ctl = get_caching_control(cache);
7319         if (!caching_ctl)
7320                 return;
7321
7322         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7323                    (cache->free_space_ctl->free_space >= num_bytes));
7324
7325         put_caching_control(caching_ctl);
7326 }
7327
7328 static noinline int
7329 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7330 {
7331         struct btrfs_caching_control *caching_ctl;
7332         int ret = 0;
7333
7334         caching_ctl = get_caching_control(cache);
7335         if (!caching_ctl)
7336                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7337
7338         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7339         if (cache->cached == BTRFS_CACHE_ERROR)
7340                 ret = -EIO;
7341         put_caching_control(caching_ctl);
7342         return ret;
7343 }
7344
7345 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7346         [BTRFS_RAID_RAID10]     = "raid10",
7347         [BTRFS_RAID_RAID1]      = "raid1",
7348         [BTRFS_RAID_DUP]        = "dup",
7349         [BTRFS_RAID_RAID0]      = "raid0",
7350         [BTRFS_RAID_SINGLE]     = "single",
7351         [BTRFS_RAID_RAID5]      = "raid5",
7352         [BTRFS_RAID_RAID6]      = "raid6",
7353 };
7354
7355 static const char *get_raid_name(enum btrfs_raid_types type)
7356 {
7357         if (type >= BTRFS_NR_RAID_TYPES)
7358                 return NULL;
7359
7360         return btrfs_raid_type_names[type];
7361 }
7362
7363 enum btrfs_loop_type {
7364         LOOP_CACHING_NOWAIT = 0,
7365         LOOP_CACHING_WAIT = 1,
7366         LOOP_ALLOC_CHUNK = 2,
7367         LOOP_NO_EMPTY_SIZE = 3,
7368 };
7369
7370 static inline void
7371 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7372                        int delalloc)
7373 {
7374         if (delalloc)
7375                 down_read(&cache->data_rwsem);
7376 }
7377
7378 static inline void
7379 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7380                        int delalloc)
7381 {
7382         btrfs_get_block_group(cache);
7383         if (delalloc)
7384                 down_read(&cache->data_rwsem);
7385 }
7386
7387 static struct btrfs_block_group_cache *
7388 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7389                    struct btrfs_free_cluster *cluster,
7390                    int delalloc)
7391 {
7392         struct btrfs_block_group_cache *used_bg = NULL;
7393
7394         spin_lock(&cluster->refill_lock);
7395         while (1) {
7396                 used_bg = cluster->block_group;
7397                 if (!used_bg)
7398                         return NULL;
7399
7400                 if (used_bg == block_group)
7401                         return used_bg;
7402
7403                 btrfs_get_block_group(used_bg);
7404
7405                 if (!delalloc)
7406                         return used_bg;
7407
7408                 if (down_read_trylock(&used_bg->data_rwsem))
7409                         return used_bg;
7410
7411                 spin_unlock(&cluster->refill_lock);
7412
7413                 /* We should only have one-level nested. */
7414                 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7415
7416                 spin_lock(&cluster->refill_lock);
7417                 if (used_bg == cluster->block_group)
7418                         return used_bg;
7419
7420                 up_read(&used_bg->data_rwsem);
7421                 btrfs_put_block_group(used_bg);
7422         }
7423 }
7424
7425 static inline void
7426 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7427                          int delalloc)
7428 {
7429         if (delalloc)
7430                 up_read(&cache->data_rwsem);
7431         btrfs_put_block_group(cache);
7432 }
7433
7434 /*
7435  * walks the btree of allocated extents and find a hole of a given size.
7436  * The key ins is changed to record the hole:
7437  * ins->objectid == start position
7438  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7439  * ins->offset == the size of the hole.
7440  * Any available blocks before search_start are skipped.
7441  *
7442  * If there is no suitable free space, we will record the max size of
7443  * the free space extent currently.
7444  */
7445 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7446                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7447                                 u64 hint_byte, struct btrfs_key *ins,
7448                                 u64 flags, int delalloc)
7449 {
7450         int ret = 0;
7451         struct btrfs_root *root = fs_info->extent_root;
7452         struct btrfs_free_cluster *last_ptr = NULL;
7453         struct btrfs_block_group_cache *block_group = NULL;
7454         u64 search_start = 0;
7455         u64 max_extent_size = 0;
7456         u64 empty_cluster = 0;
7457         struct btrfs_space_info *space_info;
7458         int loop = 0;
7459         int index = btrfs_bg_flags_to_raid_index(flags);
7460         bool failed_cluster_refill = false;
7461         bool failed_alloc = false;
7462         bool use_cluster = true;
7463         bool have_caching_bg = false;
7464         bool orig_have_caching_bg = false;
7465         bool full_search = false;
7466
7467         WARN_ON(num_bytes < fs_info->sectorsize);
7468         ins->type = BTRFS_EXTENT_ITEM_KEY;
7469         ins->objectid = 0;
7470         ins->offset = 0;
7471
7472         trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7473
7474         space_info = __find_space_info(fs_info, flags);
7475         if (!space_info) {
7476                 btrfs_err(fs_info, "No space info for %llu", flags);
7477                 return -ENOSPC;
7478         }
7479
7480         /*
7481          * If our free space is heavily fragmented we may not be able to make
7482          * big contiguous allocations, so instead of doing the expensive search
7483          * for free space, simply return ENOSPC with our max_extent_size so we
7484          * can go ahead and search for a more manageable chunk.
7485          *
7486          * If our max_extent_size is large enough for our allocation simply
7487          * disable clustering since we will likely not be able to find enough
7488          * space to create a cluster and induce latency trying.
7489          */
7490         if (unlikely(space_info->max_extent_size)) {
7491                 spin_lock(&space_info->lock);
7492                 if (space_info->max_extent_size &&
7493                     num_bytes > space_info->max_extent_size) {
7494                         ins->offset = space_info->max_extent_size;
7495                         spin_unlock(&space_info->lock);
7496                         return -ENOSPC;
7497                 } else if (space_info->max_extent_size) {
7498                         use_cluster = false;
7499                 }
7500                 spin_unlock(&space_info->lock);
7501         }
7502
7503         last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7504         if (last_ptr) {
7505                 spin_lock(&last_ptr->lock);
7506                 if (last_ptr->block_group)
7507                         hint_byte = last_ptr->window_start;
7508                 if (last_ptr->fragmented) {
7509                         /*
7510                          * We still set window_start so we can keep track of the
7511                          * last place we found an allocation to try and save
7512                          * some time.
7513                          */
7514                         hint_byte = last_ptr->window_start;
7515                         use_cluster = false;
7516                 }
7517                 spin_unlock(&last_ptr->lock);
7518         }
7519
7520         search_start = max(search_start, first_logical_byte(fs_info, 0));
7521         search_start = max(search_start, hint_byte);
7522         if (search_start == hint_byte) {
7523                 block_group = btrfs_lookup_block_group(fs_info, search_start);
7524                 /*
7525                  * we don't want to use the block group if it doesn't match our
7526                  * allocation bits, or if its not cached.
7527                  *
7528                  * However if we are re-searching with an ideal block group
7529                  * picked out then we don't care that the block group is cached.
7530                  */
7531                 if (block_group && block_group_bits(block_group, flags) &&
7532                     block_group->cached != BTRFS_CACHE_NO) {
7533                         down_read(&space_info->groups_sem);
7534                         if (list_empty(&block_group->list) ||
7535                             block_group->ro) {
7536                                 /*
7537                                  * someone is removing this block group,
7538                                  * we can't jump into the have_block_group
7539                                  * target because our list pointers are not
7540                                  * valid
7541                                  */
7542                                 btrfs_put_block_group(block_group);
7543                                 up_read(&space_info->groups_sem);
7544                         } else {
7545                                 index = btrfs_bg_flags_to_raid_index(
7546                                                 block_group->flags);
7547                                 btrfs_lock_block_group(block_group, delalloc);
7548                                 goto have_block_group;
7549                         }
7550                 } else if (block_group) {
7551                         btrfs_put_block_group(block_group);
7552                 }
7553         }
7554 search:
7555         have_caching_bg = false;
7556         if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7557                 full_search = true;
7558         down_read(&space_info->groups_sem);
7559         list_for_each_entry(block_group, &space_info->block_groups[index],
7560                             list) {
7561                 u64 offset;
7562                 int cached;
7563
7564                 /* If the block group is read-only, we can skip it entirely. */
7565                 if (unlikely(block_group->ro))
7566                         continue;
7567
7568                 btrfs_grab_block_group(block_group, delalloc);
7569                 search_start = block_group->key.objectid;
7570
7571                 /*
7572                  * this can happen if we end up cycling through all the
7573                  * raid types, but we want to make sure we only allocate
7574                  * for the proper type.
7575                  */
7576                 if (!block_group_bits(block_group, flags)) {
7577                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7578                                 BTRFS_BLOCK_GROUP_RAID1 |
7579                                 BTRFS_BLOCK_GROUP_RAID5 |
7580                                 BTRFS_BLOCK_GROUP_RAID6 |
7581                                 BTRFS_BLOCK_GROUP_RAID10;
7582
7583                         /*
7584                          * if they asked for extra copies and this block group
7585                          * doesn't provide them, bail.  This does allow us to
7586                          * fill raid0 from raid1.
7587                          */
7588                         if ((flags & extra) && !(block_group->flags & extra))
7589                                 goto loop;
7590                 }
7591
7592 have_block_group:
7593                 cached = block_group_cache_done(block_group);
7594                 if (unlikely(!cached)) {
7595                         have_caching_bg = true;
7596                         ret = cache_block_group(block_group, 0);
7597                         BUG_ON(ret < 0);
7598                         ret = 0;
7599                 }
7600
7601                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7602                         goto loop;
7603
7604                 /*
7605                  * Ok we want to try and use the cluster allocator, so
7606                  * lets look there
7607                  */
7608                 if (last_ptr && use_cluster) {
7609                         struct btrfs_block_group_cache *used_block_group;
7610                         unsigned long aligned_cluster;
7611                         /*
7612                          * the refill lock keeps out other
7613                          * people trying to start a new cluster
7614                          */
7615                         used_block_group = btrfs_lock_cluster(block_group,
7616                                                               last_ptr,
7617                                                               delalloc);
7618                         if (!used_block_group)
7619                                 goto refill_cluster;
7620
7621                         if (used_block_group != block_group &&
7622                             (used_block_group->ro ||
7623                              !block_group_bits(used_block_group, flags)))
7624                                 goto release_cluster;
7625
7626                         offset = btrfs_alloc_from_cluster(used_block_group,
7627                                                 last_ptr,
7628                                                 num_bytes,
7629                                                 used_block_group->key.objectid,
7630                                                 &max_extent_size);
7631                         if (offset) {
7632                                 /* we have a block, we're done */
7633                                 spin_unlock(&last_ptr->refill_lock);
7634                                 trace_btrfs_reserve_extent_cluster(fs_info,
7635                                                 used_block_group,
7636                                                 search_start, num_bytes);
7637                                 if (used_block_group != block_group) {
7638                                         btrfs_release_block_group(block_group,
7639                                                                   delalloc);
7640                                         block_group = used_block_group;
7641                                 }
7642                                 goto checks;
7643                         }
7644
7645                         WARN_ON(last_ptr->block_group != used_block_group);
7646 release_cluster:
7647                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7648                          * set up a new clusters, so lets just skip it
7649                          * and let the allocator find whatever block
7650                          * it can find.  If we reach this point, we
7651                          * will have tried the cluster allocator
7652                          * plenty of times and not have found
7653                          * anything, so we are likely way too
7654                          * fragmented for the clustering stuff to find
7655                          * anything.
7656                          *
7657                          * However, if the cluster is taken from the
7658                          * current block group, release the cluster
7659                          * first, so that we stand a better chance of
7660                          * succeeding in the unclustered
7661                          * allocation.  */
7662                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7663                             used_block_group != block_group) {
7664                                 spin_unlock(&last_ptr->refill_lock);
7665                                 btrfs_release_block_group(used_block_group,
7666                                                           delalloc);
7667                                 goto unclustered_alloc;
7668                         }
7669
7670                         /*
7671                          * this cluster didn't work out, free it and
7672                          * start over
7673                          */
7674                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7675
7676                         if (used_block_group != block_group)
7677                                 btrfs_release_block_group(used_block_group,
7678                                                           delalloc);
7679 refill_cluster:
7680                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7681                                 spin_unlock(&last_ptr->refill_lock);
7682                                 goto unclustered_alloc;
7683                         }
7684
7685                         aligned_cluster = max_t(unsigned long,
7686                                                 empty_cluster + empty_size,
7687                                               block_group->full_stripe_len);
7688
7689                         /* allocate a cluster in this block group */
7690                         ret = btrfs_find_space_cluster(fs_info, block_group,
7691                                                        last_ptr, search_start,
7692                                                        num_bytes,
7693                                                        aligned_cluster);
7694                         if (ret == 0) {
7695                                 /*
7696                                  * now pull our allocation out of this
7697                                  * cluster
7698                                  */
7699                                 offset = btrfs_alloc_from_cluster(block_group,
7700                                                         last_ptr,
7701                                                         num_bytes,
7702                                                         search_start,
7703                                                         &max_extent_size);
7704                                 if (offset) {
7705                                         /* we found one, proceed */
7706                                         spin_unlock(&last_ptr->refill_lock);
7707                                         trace_btrfs_reserve_extent_cluster(fs_info,
7708                                                 block_group, search_start,
7709                                                 num_bytes);
7710                                         goto checks;
7711                                 }
7712                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7713                                    && !failed_cluster_refill) {
7714                                 spin_unlock(&last_ptr->refill_lock);
7715
7716                                 failed_cluster_refill = true;
7717                                 wait_block_group_cache_progress(block_group,
7718                                        num_bytes + empty_cluster + empty_size);
7719                                 goto have_block_group;
7720                         }
7721
7722                         /*
7723                          * at this point we either didn't find a cluster
7724                          * or we weren't able to allocate a block from our
7725                          * cluster.  Free the cluster we've been trying
7726                          * to use, and go to the next block group
7727                          */
7728                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7729                         spin_unlock(&last_ptr->refill_lock);
7730                         goto loop;
7731                 }
7732
7733 unclustered_alloc:
7734                 /*
7735                  * We are doing an unclustered alloc, set the fragmented flag so
7736                  * we don't bother trying to setup a cluster again until we get
7737                  * more space.
7738                  */
7739                 if (unlikely(last_ptr)) {
7740                         spin_lock(&last_ptr->lock);
7741                         last_ptr->fragmented = 1;
7742                         spin_unlock(&last_ptr->lock);
7743                 }
7744                 if (cached) {
7745                         struct btrfs_free_space_ctl *ctl =
7746                                 block_group->free_space_ctl;
7747
7748                         spin_lock(&ctl->tree_lock);
7749                         if (ctl->free_space <
7750                             num_bytes + empty_cluster + empty_size) {
7751                                 if (ctl->free_space > max_extent_size)
7752                                         max_extent_size = ctl->free_space;
7753                                 spin_unlock(&ctl->tree_lock);
7754                                 goto loop;
7755                         }
7756                         spin_unlock(&ctl->tree_lock);
7757                 }
7758
7759                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7760                                                     num_bytes, empty_size,
7761                                                     &max_extent_size);
7762                 /*
7763                  * If we didn't find a chunk, and we haven't failed on this
7764                  * block group before, and this block group is in the middle of
7765                  * caching and we are ok with waiting, then go ahead and wait
7766                  * for progress to be made, and set failed_alloc to true.
7767                  *
7768                  * If failed_alloc is true then we've already waited on this
7769                  * block group once and should move on to the next block group.
7770                  */
7771                 if (!offset && !failed_alloc && !cached &&
7772                     loop > LOOP_CACHING_NOWAIT) {
7773                         wait_block_group_cache_progress(block_group,
7774                                                 num_bytes + empty_size);
7775                         failed_alloc = true;
7776                         goto have_block_group;
7777                 } else if (!offset) {
7778                         goto loop;
7779                 }
7780 checks:
7781                 search_start = ALIGN(offset, fs_info->stripesize);
7782
7783                 /* move on to the next group */
7784                 if (search_start + num_bytes >
7785                     block_group->key.objectid + block_group->key.offset) {
7786                         btrfs_add_free_space(block_group, offset, num_bytes);
7787                         goto loop;
7788                 }
7789
7790                 if (offset < search_start)
7791                         btrfs_add_free_space(block_group, offset,
7792                                              search_start - offset);
7793                 BUG_ON(offset > search_start);
7794
7795                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7796                                 num_bytes, delalloc);
7797                 if (ret == -EAGAIN) {
7798                         btrfs_add_free_space(block_group, offset, num_bytes);
7799                         goto loop;
7800                 }
7801                 btrfs_inc_block_group_reservations(block_group);
7802
7803                 /* we are all good, lets return */
7804                 ins->objectid = search_start;
7805                 ins->offset = num_bytes;
7806
7807                 trace_btrfs_reserve_extent(fs_info, block_group,
7808                                            search_start, num_bytes);
7809                 btrfs_release_block_group(block_group, delalloc);
7810                 break;
7811 loop:
7812                 failed_cluster_refill = false;
7813                 failed_alloc = false;
7814                 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7815                        index);
7816                 btrfs_release_block_group(block_group, delalloc);
7817                 cond_resched();
7818         }
7819         up_read(&space_info->groups_sem);
7820
7821         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7822                 && !orig_have_caching_bg)
7823                 orig_have_caching_bg = true;
7824
7825         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7826                 goto search;
7827
7828         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7829                 goto search;
7830
7831         /*
7832          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7833          *                      caching kthreads as we move along
7834          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7835          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7836          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7837          *                      again
7838          */
7839         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7840                 index = 0;
7841                 if (loop == LOOP_CACHING_NOWAIT) {
7842                         /*
7843                          * We want to skip the LOOP_CACHING_WAIT step if we
7844                          * don't have any uncached bgs and we've already done a
7845                          * full search through.
7846                          */
7847                         if (orig_have_caching_bg || !full_search)
7848                                 loop = LOOP_CACHING_WAIT;
7849                         else
7850                                 loop = LOOP_ALLOC_CHUNK;
7851                 } else {
7852                         loop++;
7853                 }
7854
7855                 if (loop == LOOP_ALLOC_CHUNK) {
7856                         struct btrfs_trans_handle *trans;
7857                         int exist = 0;
7858
7859                         trans = current->journal_info;
7860                         if (trans)
7861                                 exist = 1;
7862                         else
7863                                 trans = btrfs_join_transaction(root);
7864
7865                         if (IS_ERR(trans)) {
7866                                 ret = PTR_ERR(trans);
7867                                 goto out;
7868                         }
7869
7870                         ret = do_chunk_alloc(trans, fs_info, flags,
7871                                              CHUNK_ALLOC_FORCE);
7872
7873                         /*
7874                          * If we can't allocate a new chunk we've already looped
7875                          * through at least once, move on to the NO_EMPTY_SIZE
7876                          * case.
7877                          */
7878                         if (ret == -ENOSPC)
7879                                 loop = LOOP_NO_EMPTY_SIZE;
7880
7881                         /*
7882                          * Do not bail out on ENOSPC since we
7883                          * can do more things.
7884                          */
7885                         if (ret < 0 && ret != -ENOSPC)
7886                                 btrfs_abort_transaction(trans, ret);
7887                         else
7888                                 ret = 0;
7889                         if (!exist)
7890                                 btrfs_end_transaction(trans);
7891                         if (ret)
7892                                 goto out;
7893                 }
7894
7895                 if (loop == LOOP_NO_EMPTY_SIZE) {
7896                         /*
7897                          * Don't loop again if we already have no empty_size and
7898                          * no empty_cluster.
7899                          */
7900                         if (empty_size == 0 &&
7901                             empty_cluster == 0) {
7902                                 ret = -ENOSPC;
7903                                 goto out;
7904                         }
7905                         empty_size = 0;
7906                         empty_cluster = 0;
7907                 }
7908
7909                 goto search;
7910         } else if (!ins->objectid) {
7911                 ret = -ENOSPC;
7912         } else if (ins->objectid) {
7913                 if (!use_cluster && last_ptr) {
7914                         spin_lock(&last_ptr->lock);
7915                         last_ptr->window_start = ins->objectid;
7916                         spin_unlock(&last_ptr->lock);
7917                 }
7918                 ret = 0;
7919         }
7920 out:
7921         if (ret == -ENOSPC) {
7922                 spin_lock(&space_info->lock);
7923                 space_info->max_extent_size = max_extent_size;
7924                 spin_unlock(&space_info->lock);
7925                 ins->offset = max_extent_size;
7926         }
7927         return ret;
7928 }
7929
7930 static void dump_space_info(struct btrfs_fs_info *fs_info,
7931                             struct btrfs_space_info *info, u64 bytes,
7932                             int dump_block_groups)
7933 {
7934         struct btrfs_block_group_cache *cache;
7935         int index = 0;
7936
7937         spin_lock(&info->lock);
7938         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7939                    info->flags,
7940                    info->total_bytes - btrfs_space_info_used(info, true),
7941                    info->full ? "" : "not ");
7942         btrfs_info(fs_info,
7943                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7944                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7945                 info->bytes_reserved, info->bytes_may_use,
7946                 info->bytes_readonly);
7947         spin_unlock(&info->lock);
7948
7949         if (!dump_block_groups)
7950                 return;
7951
7952         down_read(&info->groups_sem);
7953 again:
7954         list_for_each_entry(cache, &info->block_groups[index], list) {
7955                 spin_lock(&cache->lock);
7956                 btrfs_info(fs_info,
7957                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7958                         cache->key.objectid, cache->key.offset,
7959                         btrfs_block_group_used(&cache->item), cache->pinned,
7960                         cache->reserved, cache->ro ? "[readonly]" : "");
7961                 btrfs_dump_free_space(cache, bytes);
7962                 spin_unlock(&cache->lock);
7963         }
7964         if (++index < BTRFS_NR_RAID_TYPES)
7965                 goto again;
7966         up_read(&info->groups_sem);
7967 }
7968
7969 /*
7970  * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7971  *                        hole that is at least as big as @num_bytes.
7972  *
7973  * @root           -    The root that will contain this extent
7974  *
7975  * @ram_bytes      -    The amount of space in ram that @num_bytes take. This
7976  *                      is used for accounting purposes. This value differs
7977  *                      from @num_bytes only in the case of compressed extents.
7978  *
7979  * @num_bytes      -    Number of bytes to allocate on-disk.
7980  *
7981  * @min_alloc_size -    Indicates the minimum amount of space that the
7982  *                      allocator should try to satisfy. In some cases
7983  *                      @num_bytes may be larger than what is required and if
7984  *                      the filesystem is fragmented then allocation fails.
7985  *                      However, the presence of @min_alloc_size gives a
7986  *                      chance to try and satisfy the smaller allocation.
7987  *
7988  * @empty_size     -    A hint that you plan on doing more COW. This is the
7989  *                      size in bytes the allocator should try to find free
7990  *                      next to the block it returns.  This is just a hint and
7991  *                      may be ignored by the allocator.
7992  *
7993  * @hint_byte      -    Hint to the allocator to start searching above the byte
7994  *                      address passed. It might be ignored.
7995  *
7996  * @ins            -    This key is modified to record the found hole. It will
7997  *                      have the following values:
7998  *                      ins->objectid == start position
7999  *                      ins->flags = BTRFS_EXTENT_ITEM_KEY
8000  *                      ins->offset == the size of the hole.
8001  *
8002  * @is_data        -    Boolean flag indicating whether an extent is
8003  *                      allocated for data (true) or metadata (false)
8004  *
8005  * @delalloc       -    Boolean flag indicating whether this allocation is for
8006  *                      delalloc or not. If 'true' data_rwsem of block groups
8007  *                      is going to be acquired.
8008  *
8009  *
8010  * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
8011  * case -ENOSPC is returned then @ins->offset will contain the size of the
8012  * largest available hole the allocator managed to find.
8013  */
8014 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
8015                          u64 num_bytes, u64 min_alloc_size,
8016                          u64 empty_size, u64 hint_byte,
8017                          struct btrfs_key *ins, int is_data, int delalloc)
8018 {
8019         struct btrfs_fs_info *fs_info = root->fs_info;
8020         bool final_tried = num_bytes == min_alloc_size;
8021         u64 flags;
8022         int ret;
8023
8024         flags = get_alloc_profile_by_root(root, is_data);
8025 again:
8026         WARN_ON(num_bytes < fs_info->sectorsize);
8027         ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
8028                                hint_byte, ins, flags, delalloc);
8029         if (!ret && !is_data) {
8030                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
8031         } else if (ret == -ENOSPC) {
8032                 if (!final_tried && ins->offset) {
8033                         num_bytes = min(num_bytes >> 1, ins->offset);
8034                         num_bytes = round_down(num_bytes,
8035                                                fs_info->sectorsize);
8036                         num_bytes = max(num_bytes, min_alloc_size);
8037                         ram_bytes = num_bytes;
8038                         if (num_bytes == min_alloc_size)
8039                                 final_tried = true;
8040                         goto again;
8041                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8042                         struct btrfs_space_info *sinfo;
8043
8044                         sinfo = __find_space_info(fs_info, flags);
8045                         btrfs_err(fs_info,
8046                                   "allocation failed flags %llu, wanted %llu",
8047                                   flags, num_bytes);
8048                         if (sinfo)
8049                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
8050                 }
8051         }
8052
8053         return ret;
8054 }
8055
8056 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8057                                         u64 start, u64 len,
8058                                         int pin, int delalloc)
8059 {
8060         struct btrfs_block_group_cache *cache;
8061         int ret = 0;
8062
8063         cache = btrfs_lookup_block_group(fs_info, start);
8064         if (!cache) {
8065                 btrfs_err(fs_info, "Unable to find block group for %llu",
8066                           start);
8067                 return -ENOSPC;
8068         }
8069
8070         if (pin)
8071                 pin_down_extent(fs_info, cache, start, len, 1);
8072         else {
8073                 if (btrfs_test_opt(fs_info, DISCARD))
8074                         ret = btrfs_discard_extent(fs_info, start, len, NULL);
8075                 btrfs_add_free_space(cache, start, len);
8076                 btrfs_free_reserved_bytes(cache, len, delalloc);
8077                 trace_btrfs_reserved_extent_free(fs_info, start, len);
8078         }
8079
8080         btrfs_put_block_group(cache);
8081         return ret;
8082 }
8083
8084 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
8085                                u64 start, u64 len, int delalloc)
8086 {
8087         return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
8088 }
8089
8090 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
8091                                        u64 start, u64 len)
8092 {
8093         return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
8094 }
8095
8096 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8097                                       struct btrfs_fs_info *fs_info,
8098                                       u64 parent, u64 root_objectid,
8099                                       u64 flags, u64 owner, u64 offset,
8100                                       struct btrfs_key *ins, int ref_mod)
8101 {
8102         int ret;
8103         struct btrfs_extent_item *extent_item;
8104         struct btrfs_extent_inline_ref *iref;
8105         struct btrfs_path *path;
8106         struct extent_buffer *leaf;
8107         int type;
8108         u32 size;
8109
8110         if (parent > 0)
8111                 type = BTRFS_SHARED_DATA_REF_KEY;
8112         else
8113                 type = BTRFS_EXTENT_DATA_REF_KEY;
8114
8115         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8116
8117         path = btrfs_alloc_path();
8118         if (!path)
8119                 return -ENOMEM;
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                 return ret;
8127         }
8128
8129         leaf = path->nodes[0];
8130         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8131                                      struct btrfs_extent_item);
8132         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8133         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8134         btrfs_set_extent_flags(leaf, extent_item,
8135                                flags | BTRFS_EXTENT_FLAG_DATA);
8136
8137         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8138         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8139         if (parent > 0) {
8140                 struct btrfs_shared_data_ref *ref;
8141                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8142                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8143                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8144         } else {
8145                 struct btrfs_extent_data_ref *ref;
8146                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8147                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8148                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8149                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8150                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8151         }
8152
8153         btrfs_mark_buffer_dirty(path->nodes[0]);
8154         btrfs_free_path(path);
8155
8156         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8157                                           ins->offset);
8158         if (ret)
8159                 return ret;
8160
8161         ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8162         if (ret) { /* -ENOENT, logic error */
8163                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8164                         ins->objectid, ins->offset);
8165                 BUG();
8166         }
8167         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8168         return ret;
8169 }
8170
8171 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8172                                      struct btrfs_fs_info *fs_info,
8173                                      u64 parent, u64 root_objectid,
8174                                      u64 flags, struct btrfs_disk_key *key,
8175                                      int level, struct btrfs_key *ins)
8176 {
8177         int ret;
8178         struct btrfs_extent_item *extent_item;
8179         struct btrfs_tree_block_info *block_info;
8180         struct btrfs_extent_inline_ref *iref;
8181         struct btrfs_path *path;
8182         struct extent_buffer *leaf;
8183         u32 size = sizeof(*extent_item) + sizeof(*iref);
8184         u64 num_bytes = ins->offset;
8185         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8186
8187         if (!skinny_metadata)
8188                 size += sizeof(*block_info);
8189
8190         path = btrfs_alloc_path();
8191         if (!path) {
8192                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8193                                                    fs_info->nodesize);
8194                 return -ENOMEM;
8195         }
8196
8197         path->leave_spinning = 1;
8198         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8199                                       ins, size);
8200         if (ret) {
8201                 btrfs_free_path(path);
8202                 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid,
8203                                                    fs_info->nodesize);
8204                 return ret;
8205         }
8206
8207         leaf = path->nodes[0];
8208         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8209                                      struct btrfs_extent_item);
8210         btrfs_set_extent_refs(leaf, extent_item, 1);
8211         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8212         btrfs_set_extent_flags(leaf, extent_item,
8213                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8214
8215         if (skinny_metadata) {
8216                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8217                 num_bytes = fs_info->nodesize;
8218         } else {
8219                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8220                 btrfs_set_tree_block_key(leaf, block_info, key);
8221                 btrfs_set_tree_block_level(leaf, block_info, level);
8222                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8223         }
8224
8225         if (parent > 0) {
8226                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8227                 btrfs_set_extent_inline_ref_type(leaf, iref,
8228                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8229                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8230         } else {
8231                 btrfs_set_extent_inline_ref_type(leaf, iref,
8232                                                  BTRFS_TREE_BLOCK_REF_KEY);
8233                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8234         }
8235
8236         btrfs_mark_buffer_dirty(leaf);
8237         btrfs_free_path(path);
8238
8239         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8240                                           num_bytes);
8241         if (ret)
8242                 return ret;
8243
8244         ret = update_block_group(trans, fs_info, ins->objectid,
8245                                  fs_info->nodesize, 1);
8246         if (ret) { /* -ENOENT, logic error */
8247                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8248                         ins->objectid, ins->offset);
8249                 BUG();
8250         }
8251
8252         trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid,
8253                                           fs_info->nodesize);
8254         return ret;
8255 }
8256
8257 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8258                                      struct btrfs_root *root, u64 owner,
8259                                      u64 offset, u64 ram_bytes,
8260                                      struct btrfs_key *ins)
8261 {
8262         struct btrfs_fs_info *fs_info = root->fs_info;
8263         int ret;
8264
8265         BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8266
8267         btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8268                            root->root_key.objectid, owner, offset,
8269                            BTRFS_ADD_DELAYED_EXTENT);
8270
8271         ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid,
8272                                          ins->offset, 0,
8273                                          root->root_key.objectid, owner,
8274                                          offset, ram_bytes,
8275                                          BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8276         return ret;
8277 }
8278
8279 /*
8280  * this is used by the tree logging recovery code.  It records that
8281  * an extent has been allocated and makes sure to clear the free
8282  * space cache bits as well
8283  */
8284 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8285                                    struct btrfs_fs_info *fs_info,
8286                                    u64 root_objectid, u64 owner, u64 offset,
8287                                    struct btrfs_key *ins)
8288 {
8289         int ret;
8290         struct btrfs_block_group_cache *block_group;
8291         struct btrfs_space_info *space_info;
8292
8293         /*
8294          * Mixed block groups will exclude before processing the log so we only
8295          * need to do the exclude dance if this fs isn't mixed.
8296          */
8297         if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8298                 ret = __exclude_logged_extent(fs_info, ins->objectid,
8299                                               ins->offset);
8300                 if (ret)
8301                         return ret;
8302         }
8303
8304         block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8305         if (!block_group)
8306                 return -EINVAL;
8307
8308         space_info = block_group->space_info;
8309         spin_lock(&space_info->lock);
8310         spin_lock(&block_group->lock);
8311         space_info->bytes_reserved += ins->offset;
8312         block_group->reserved += ins->offset;
8313         spin_unlock(&block_group->lock);
8314         spin_unlock(&space_info->lock);
8315
8316         ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid,
8317                                          0, owner, offset, ins, 1);
8318         btrfs_put_block_group(block_group);
8319         return ret;
8320 }
8321
8322 static struct extent_buffer *
8323 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8324                       u64 bytenr, int level)
8325 {
8326         struct btrfs_fs_info *fs_info = root->fs_info;
8327         struct extent_buffer *buf;
8328
8329         buf = btrfs_find_create_tree_block(fs_info, bytenr);
8330         if (IS_ERR(buf))
8331                 return buf;
8332
8333         btrfs_set_header_generation(buf, trans->transid);
8334         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8335         btrfs_tree_lock(buf);
8336         clean_tree_block(fs_info, buf);
8337         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8338
8339         btrfs_set_lock_blocking(buf);
8340         set_extent_buffer_uptodate(buf);
8341
8342         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8343                 buf->log_index = root->log_transid % 2;
8344                 /*
8345                  * we allow two log transactions at a time, use different
8346                  * EXENT bit to differentiate dirty pages.
8347                  */
8348                 if (buf->log_index == 0)
8349                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8350                                         buf->start + buf->len - 1, GFP_NOFS);
8351                 else
8352                         set_extent_new(&root->dirty_log_pages, buf->start,
8353                                         buf->start + buf->len - 1);
8354         } else {
8355                 buf->log_index = -1;
8356                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8357                          buf->start + buf->len - 1, GFP_NOFS);
8358         }
8359         trans->dirty = true;
8360         /* this returns a buffer locked for blocking */
8361         return buf;
8362 }
8363
8364 static struct btrfs_block_rsv *
8365 use_block_rsv(struct btrfs_trans_handle *trans,
8366               struct btrfs_root *root, u32 blocksize)
8367 {
8368         struct btrfs_fs_info *fs_info = root->fs_info;
8369         struct btrfs_block_rsv *block_rsv;
8370         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8371         int ret;
8372         bool global_updated = false;
8373
8374         block_rsv = get_block_rsv(trans, root);
8375
8376         if (unlikely(block_rsv->size == 0))
8377                 goto try_reserve;
8378 again:
8379         ret = block_rsv_use_bytes(block_rsv, blocksize);
8380         if (!ret)
8381                 return block_rsv;
8382
8383         if (block_rsv->failfast)
8384                 return ERR_PTR(ret);
8385
8386         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8387                 global_updated = true;
8388                 update_global_block_rsv(fs_info);
8389                 goto again;
8390         }
8391
8392         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8393                 static DEFINE_RATELIMIT_STATE(_rs,
8394                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8395                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8396                 if (__ratelimit(&_rs))
8397                         WARN(1, KERN_DEBUG
8398                                 "BTRFS: block rsv returned %d\n", ret);
8399         }
8400 try_reserve:
8401         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8402                                      BTRFS_RESERVE_NO_FLUSH);
8403         if (!ret)
8404                 return block_rsv;
8405         /*
8406          * If we couldn't reserve metadata bytes try and use some from
8407          * the global reserve if its space type is the same as the global
8408          * reservation.
8409          */
8410         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8411             block_rsv->space_info == global_rsv->space_info) {
8412                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8413                 if (!ret)
8414                         return global_rsv;
8415         }
8416         return ERR_PTR(ret);
8417 }
8418
8419 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8420                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8421 {
8422         block_rsv_add_bytes(block_rsv, blocksize, 0);
8423         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8424 }
8425
8426 /*
8427  * finds a free extent and does all the dirty work required for allocation
8428  * returns the tree buffer or an ERR_PTR on error.
8429  */
8430 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8431                                              struct btrfs_root *root,
8432                                              u64 parent, u64 root_objectid,
8433                                              const struct btrfs_disk_key *key,
8434                                              int level, u64 hint,
8435                                              u64 empty_size)
8436 {
8437         struct btrfs_fs_info *fs_info = root->fs_info;
8438         struct btrfs_key ins;
8439         struct btrfs_block_rsv *block_rsv;
8440         struct extent_buffer *buf;
8441         struct btrfs_delayed_extent_op *extent_op;
8442         u64 flags = 0;
8443         int ret;
8444         u32 blocksize = fs_info->nodesize;
8445         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8446
8447 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8448         if (btrfs_is_testing(fs_info)) {
8449                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8450                                             level);
8451                 if (!IS_ERR(buf))
8452                         root->alloc_bytenr += blocksize;
8453                 return buf;
8454         }
8455 #endif
8456
8457         block_rsv = use_block_rsv(trans, root, blocksize);
8458         if (IS_ERR(block_rsv))
8459                 return ERR_CAST(block_rsv);
8460
8461         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8462                                    empty_size, hint, &ins, 0, 0);
8463         if (ret)
8464                 goto out_unuse;
8465
8466         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8467         if (IS_ERR(buf)) {
8468                 ret = PTR_ERR(buf);
8469                 goto out_free_reserved;
8470         }
8471
8472         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8473                 if (parent == 0)
8474                         parent = ins.objectid;
8475                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8476         } else
8477                 BUG_ON(parent > 0);
8478
8479         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8480                 extent_op = btrfs_alloc_delayed_extent_op();
8481                 if (!extent_op) {
8482                         ret = -ENOMEM;
8483                         goto out_free_buf;
8484                 }
8485                 if (key)
8486                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8487                 else
8488                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8489                 extent_op->flags_to_set = flags;
8490                 extent_op->update_key = skinny_metadata ? false : true;
8491                 extent_op->update_flags = true;
8492                 extent_op->is_data = false;
8493                 extent_op->level = level;
8494
8495                 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8496                                    root_objectid, level, 0,
8497                                    BTRFS_ADD_DELAYED_EXTENT);
8498                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, ins.objectid,
8499                                                  ins.offset, parent,
8500                                                  root_objectid, level,
8501                                                  BTRFS_ADD_DELAYED_EXTENT,
8502                                                  extent_op, NULL, NULL);
8503                 if (ret)
8504                         goto out_free_delayed;
8505         }
8506         return buf;
8507
8508 out_free_delayed:
8509         btrfs_free_delayed_extent_op(extent_op);
8510 out_free_buf:
8511         free_extent_buffer(buf);
8512 out_free_reserved:
8513         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8514 out_unuse:
8515         unuse_block_rsv(fs_info, block_rsv, blocksize);
8516         return ERR_PTR(ret);
8517 }
8518
8519 struct walk_control {
8520         u64 refs[BTRFS_MAX_LEVEL];
8521         u64 flags[BTRFS_MAX_LEVEL];
8522         struct btrfs_key update_progress;
8523         int stage;
8524         int level;
8525         int shared_level;
8526         int update_ref;
8527         int keep_locks;
8528         int reada_slot;
8529         int reada_count;
8530         int for_reloc;
8531 };
8532
8533 #define DROP_REFERENCE  1
8534 #define UPDATE_BACKREF  2
8535
8536 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8537                                      struct btrfs_root *root,
8538                                      struct walk_control *wc,
8539                                      struct btrfs_path *path)
8540 {
8541         struct btrfs_fs_info *fs_info = root->fs_info;
8542         u64 bytenr;
8543         u64 generation;
8544         u64 refs;
8545         u64 flags;
8546         u32 nritems;
8547         struct btrfs_key key;
8548         struct extent_buffer *eb;
8549         int ret;
8550         int slot;
8551         int nread = 0;
8552
8553         if (path->slots[wc->level] < wc->reada_slot) {
8554                 wc->reada_count = wc->reada_count * 2 / 3;
8555                 wc->reada_count = max(wc->reada_count, 2);
8556         } else {
8557                 wc->reada_count = wc->reada_count * 3 / 2;
8558                 wc->reada_count = min_t(int, wc->reada_count,
8559                                         BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8560         }
8561
8562         eb = path->nodes[wc->level];
8563         nritems = btrfs_header_nritems(eb);
8564
8565         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8566                 if (nread >= wc->reada_count)
8567                         break;
8568
8569                 cond_resched();
8570                 bytenr = btrfs_node_blockptr(eb, slot);
8571                 generation = btrfs_node_ptr_generation(eb, slot);
8572
8573                 if (slot == path->slots[wc->level])
8574                         goto reada;
8575
8576                 if (wc->stage == UPDATE_BACKREF &&
8577                     generation <= root->root_key.offset)
8578                         continue;
8579
8580                 /* We don't lock the tree block, it's OK to be racy here */
8581                 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8582                                                wc->level - 1, 1, &refs,
8583                                                &flags);
8584                 /* We don't care about errors in readahead. */
8585                 if (ret < 0)
8586                         continue;
8587                 BUG_ON(refs == 0);
8588
8589                 if (wc->stage == DROP_REFERENCE) {
8590                         if (refs == 1)
8591                                 goto reada;
8592
8593                         if (wc->level == 1 &&
8594                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8595                                 continue;
8596                         if (!wc->update_ref ||
8597                             generation <= root->root_key.offset)
8598                                 continue;
8599                         btrfs_node_key_to_cpu(eb, &key, slot);
8600                         ret = btrfs_comp_cpu_keys(&key,
8601                                                   &wc->update_progress);
8602                         if (ret < 0)
8603                                 continue;
8604                 } else {
8605                         if (wc->level == 1 &&
8606                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8607                                 continue;
8608                 }
8609 reada:
8610                 readahead_tree_block(fs_info, bytenr);
8611                 nread++;
8612         }
8613         wc->reada_slot = slot;
8614 }
8615
8616 /*
8617  * helper to process tree block while walking down the tree.
8618  *
8619  * when wc->stage == UPDATE_BACKREF, this function updates
8620  * back refs for pointers in the block.
8621  *
8622  * NOTE: return value 1 means we should stop walking down.
8623  */
8624 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8625                                    struct btrfs_root *root,
8626                                    struct btrfs_path *path,
8627                                    struct walk_control *wc, int lookup_info)
8628 {
8629         struct btrfs_fs_info *fs_info = root->fs_info;
8630         int level = wc->level;
8631         struct extent_buffer *eb = path->nodes[level];
8632         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8633         int ret;
8634
8635         if (wc->stage == UPDATE_BACKREF &&
8636             btrfs_header_owner(eb) != root->root_key.objectid)
8637                 return 1;
8638
8639         /*
8640          * when reference count of tree block is 1, it won't increase
8641          * again. once full backref flag is set, we never clear it.
8642          */
8643         if (lookup_info &&
8644             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8645              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8646                 BUG_ON(!path->locks[level]);
8647                 ret = btrfs_lookup_extent_info(trans, fs_info,
8648                                                eb->start, level, 1,
8649                                                &wc->refs[level],
8650                                                &wc->flags[level]);
8651                 BUG_ON(ret == -ENOMEM);
8652                 if (ret)
8653                         return ret;
8654                 BUG_ON(wc->refs[level] == 0);
8655         }
8656
8657         if (wc->stage == DROP_REFERENCE) {
8658                 if (wc->refs[level] > 1)
8659                         return 1;
8660
8661                 if (path->locks[level] && !wc->keep_locks) {
8662                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8663                         path->locks[level] = 0;
8664                 }
8665                 return 0;
8666         }
8667
8668         /* wc->stage == UPDATE_BACKREF */
8669         if (!(wc->flags[level] & flag)) {
8670                 BUG_ON(!path->locks[level]);
8671                 ret = btrfs_inc_ref(trans, root, eb, 1);
8672                 BUG_ON(ret); /* -ENOMEM */
8673                 ret = btrfs_dec_ref(trans, root, eb, 0);
8674                 BUG_ON(ret); /* -ENOMEM */
8675                 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8676                                                   eb->len, flag,
8677                                                   btrfs_header_level(eb), 0);
8678                 BUG_ON(ret); /* -ENOMEM */
8679                 wc->flags[level] |= flag;
8680         }
8681
8682         /*
8683          * the block is shared by multiple trees, so it's not good to
8684          * keep the tree lock
8685          */
8686         if (path->locks[level] && level > 0) {
8687                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8688                 path->locks[level] = 0;
8689         }
8690         return 0;
8691 }
8692
8693 /*
8694  * helper to process tree block pointer.
8695  *
8696  * when wc->stage == DROP_REFERENCE, this function checks
8697  * reference count of the block pointed to. if the block
8698  * is shared and we need update back refs for the subtree
8699  * rooted at the block, this function changes wc->stage to
8700  * UPDATE_BACKREF. if the block is shared and there is no
8701  * need to update back, this function drops the reference
8702  * to the block.
8703  *
8704  * NOTE: return value 1 means we should stop walking down.
8705  */
8706 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8707                                  struct btrfs_root *root,
8708                                  struct btrfs_path *path,
8709                                  struct walk_control *wc, int *lookup_info)
8710 {
8711         struct btrfs_fs_info *fs_info = root->fs_info;
8712         u64 bytenr;
8713         u64 generation;
8714         u64 parent;
8715         u32 blocksize;
8716         struct btrfs_key key;
8717         struct extent_buffer *next;
8718         int level = wc->level;
8719         int reada = 0;
8720         int ret = 0;
8721         bool need_account = false;
8722
8723         generation = btrfs_node_ptr_generation(path->nodes[level],
8724                                                path->slots[level]);
8725         /*
8726          * if the lower level block was created before the snapshot
8727          * was created, we know there is no need to update back refs
8728          * for the subtree
8729          */
8730         if (wc->stage == UPDATE_BACKREF &&
8731             generation <= root->root_key.offset) {
8732                 *lookup_info = 1;
8733                 return 1;
8734         }
8735
8736         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8737         blocksize = fs_info->nodesize;
8738
8739         next = find_extent_buffer(fs_info, bytenr);
8740         if (!next) {
8741                 next = btrfs_find_create_tree_block(fs_info, bytenr);
8742                 if (IS_ERR(next))
8743                         return PTR_ERR(next);
8744
8745                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8746                                                level - 1);
8747                 reada = 1;
8748         }
8749         btrfs_tree_lock(next);
8750         btrfs_set_lock_blocking(next);
8751
8752         ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8753                                        &wc->refs[level - 1],
8754                                        &wc->flags[level - 1]);
8755         if (ret < 0)
8756                 goto out_unlock;
8757
8758         if (unlikely(wc->refs[level - 1] == 0)) {
8759                 btrfs_err(fs_info, "Missing references.");
8760                 ret = -EIO;
8761                 goto out_unlock;
8762         }
8763         *lookup_info = 0;
8764
8765         if (wc->stage == DROP_REFERENCE) {
8766                 if (wc->refs[level - 1] > 1) {
8767                         need_account = true;
8768                         if (level == 1 &&
8769                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8770                                 goto skip;
8771
8772                         if (!wc->update_ref ||
8773                             generation <= root->root_key.offset)
8774                                 goto skip;
8775
8776                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8777                                               path->slots[level]);
8778                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8779                         if (ret < 0)
8780                                 goto skip;
8781
8782                         wc->stage = UPDATE_BACKREF;
8783                         wc->shared_level = level - 1;
8784                 }
8785         } else {
8786                 if (level == 1 &&
8787                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8788                         goto skip;
8789         }
8790
8791         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8792                 btrfs_tree_unlock(next);
8793                 free_extent_buffer(next);
8794                 next = NULL;
8795                 *lookup_info = 1;
8796         }
8797
8798         if (!next) {
8799                 if (reada && level == 1)
8800                         reada_walk_down(trans, root, wc, path);
8801                 next = read_tree_block(fs_info, bytenr, generation);
8802                 if (IS_ERR(next)) {
8803                         return PTR_ERR(next);
8804                 } else if (!extent_buffer_uptodate(next)) {
8805                         free_extent_buffer(next);
8806                         return -EIO;
8807                 }
8808                 btrfs_tree_lock(next);
8809                 btrfs_set_lock_blocking(next);
8810         }
8811
8812         level--;
8813         ASSERT(level == btrfs_header_level(next));
8814         if (level != btrfs_header_level(next)) {
8815                 btrfs_err(root->fs_info, "mismatched level");
8816                 ret = -EIO;
8817                 goto out_unlock;
8818         }
8819         path->nodes[level] = next;
8820         path->slots[level] = 0;
8821         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8822         wc->level = level;
8823         if (wc->level == 1)
8824                 wc->reada_slot = 0;
8825         return 0;
8826 skip:
8827         wc->refs[level - 1] = 0;
8828         wc->flags[level - 1] = 0;
8829         if (wc->stage == DROP_REFERENCE) {
8830                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8831                         parent = path->nodes[level]->start;
8832                 } else {
8833                         ASSERT(root->root_key.objectid ==
8834                                btrfs_header_owner(path->nodes[level]));
8835                         if (root->root_key.objectid !=
8836                             btrfs_header_owner(path->nodes[level])) {
8837                                 btrfs_err(root->fs_info,
8838                                                 "mismatched block owner");
8839                                 ret = -EIO;
8840                                 goto out_unlock;
8841                         }
8842                         parent = 0;
8843                 }
8844
8845                 if (need_account) {
8846                         ret = btrfs_qgroup_trace_subtree(trans, root, next,
8847                                                          generation, level - 1);
8848                         if (ret) {
8849                                 btrfs_err_rl(fs_info,
8850                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8851                                              ret);
8852                         }
8853                 }
8854                 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8855                                         parent, root->root_key.objectid,
8856                                         level - 1, 0);
8857                 if (ret)
8858                         goto out_unlock;
8859         }
8860
8861         *lookup_info = 1;
8862         ret = 1;
8863
8864 out_unlock:
8865         btrfs_tree_unlock(next);
8866         free_extent_buffer(next);
8867
8868         return ret;
8869 }
8870
8871 /*
8872  * helper to process tree block while walking up the tree.
8873  *
8874  * when wc->stage == DROP_REFERENCE, this function drops
8875  * reference count on the block.
8876  *
8877  * when wc->stage == UPDATE_BACKREF, this function changes
8878  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8879  * to UPDATE_BACKREF previously while processing the block.
8880  *
8881  * NOTE: return value 1 means we should stop walking up.
8882  */
8883 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8884                                  struct btrfs_root *root,
8885                                  struct btrfs_path *path,
8886                                  struct walk_control *wc)
8887 {
8888         struct btrfs_fs_info *fs_info = root->fs_info;
8889         int ret;
8890         int level = wc->level;
8891         struct extent_buffer *eb = path->nodes[level];
8892         u64 parent = 0;
8893
8894         if (wc->stage == UPDATE_BACKREF) {
8895                 BUG_ON(wc->shared_level < level);
8896                 if (level < wc->shared_level)
8897                         goto out;
8898
8899                 ret = find_next_key(path, level + 1, &wc->update_progress);
8900                 if (ret > 0)
8901                         wc->update_ref = 0;
8902
8903                 wc->stage = DROP_REFERENCE;
8904                 wc->shared_level = -1;
8905                 path->slots[level] = 0;
8906
8907                 /*
8908                  * check reference count again if the block isn't locked.
8909                  * we should start walking down the tree again if reference
8910                  * count is one.
8911                  */
8912                 if (!path->locks[level]) {
8913                         BUG_ON(level == 0);
8914                         btrfs_tree_lock(eb);
8915                         btrfs_set_lock_blocking(eb);
8916                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8917
8918                         ret = btrfs_lookup_extent_info(trans, fs_info,
8919                                                        eb->start, level, 1,
8920                                                        &wc->refs[level],
8921                                                        &wc->flags[level]);
8922                         if (ret < 0) {
8923                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8924                                 path->locks[level] = 0;
8925                                 return ret;
8926                         }
8927                         BUG_ON(wc->refs[level] == 0);
8928                         if (wc->refs[level] == 1) {
8929                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8930                                 path->locks[level] = 0;
8931                                 return 1;
8932                         }
8933                 }
8934         }
8935
8936         /* wc->stage == DROP_REFERENCE */
8937         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8938
8939         if (wc->refs[level] == 1) {
8940                 if (level == 0) {
8941                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8942                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8943                         else
8944                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8945                         BUG_ON(ret); /* -ENOMEM */
8946                         ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb);
8947                         if (ret) {
8948                                 btrfs_err_rl(fs_info,
8949                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
8950                                              ret);
8951                         }
8952                 }
8953                 /* make block locked assertion in clean_tree_block happy */
8954                 if (!path->locks[level] &&
8955                     btrfs_header_generation(eb) == trans->transid) {
8956                         btrfs_tree_lock(eb);
8957                         btrfs_set_lock_blocking(eb);
8958                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8959                 }
8960                 clean_tree_block(fs_info, eb);
8961         }
8962
8963         if (eb == root->node) {
8964                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8965                         parent = eb->start;
8966                 else
8967                         BUG_ON(root->root_key.objectid !=
8968                                btrfs_header_owner(eb));
8969         } else {
8970                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8971                         parent = path->nodes[level + 1]->start;
8972                 else
8973                         BUG_ON(root->root_key.objectid !=
8974                                btrfs_header_owner(path->nodes[level + 1]));
8975         }
8976
8977         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8978 out:
8979         wc->refs[level] = 0;
8980         wc->flags[level] = 0;
8981         return 0;
8982 }
8983
8984 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8985                                    struct btrfs_root *root,
8986                                    struct btrfs_path *path,
8987                                    struct walk_control *wc)
8988 {
8989         int level = wc->level;
8990         int lookup_info = 1;
8991         int ret;
8992
8993         while (level >= 0) {
8994                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8995                 if (ret > 0)
8996                         break;
8997
8998                 if (level == 0)
8999                         break;
9000
9001                 if (path->slots[level] >=
9002                     btrfs_header_nritems(path->nodes[level]))
9003                         break;
9004
9005                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9006                 if (ret > 0) {
9007                         path->slots[level]++;
9008                         continue;
9009                 } else if (ret < 0)
9010                         return ret;
9011                 level = wc->level;
9012         }
9013         return 0;
9014 }
9015
9016 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9017                                  struct btrfs_root *root,
9018                                  struct btrfs_path *path,
9019                                  struct walk_control *wc, int max_level)
9020 {
9021         int level = wc->level;
9022         int ret;
9023
9024         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9025         while (level < max_level && path->nodes[level]) {
9026                 wc->level = level;
9027                 if (path->slots[level] + 1 <
9028                     btrfs_header_nritems(path->nodes[level])) {
9029                         path->slots[level]++;
9030                         return 0;
9031                 } else {
9032                         ret = walk_up_proc(trans, root, path, wc);
9033                         if (ret > 0)
9034                                 return 0;
9035
9036                         if (path->locks[level]) {
9037                                 btrfs_tree_unlock_rw(path->nodes[level],
9038                                                      path->locks[level]);
9039                                 path->locks[level] = 0;
9040                         }
9041                         free_extent_buffer(path->nodes[level]);
9042                         path->nodes[level] = NULL;
9043                         level++;
9044                 }
9045         }
9046         return 1;
9047 }
9048
9049 /*
9050  * drop a subvolume tree.
9051  *
9052  * this function traverses the tree freeing any blocks that only
9053  * referenced by the tree.
9054  *
9055  * when a shared tree block is found. this function decreases its
9056  * reference count by one. if update_ref is true, this function
9057  * also make sure backrefs for the shared block and all lower level
9058  * blocks are properly updated.
9059  *
9060  * If called with for_reloc == 0, may exit early with -EAGAIN
9061  */
9062 int btrfs_drop_snapshot(struct btrfs_root *root,
9063                          struct btrfs_block_rsv *block_rsv, int update_ref,
9064                          int for_reloc)
9065 {
9066         struct btrfs_fs_info *fs_info = root->fs_info;
9067         struct btrfs_path *path;
9068         struct btrfs_trans_handle *trans;
9069         struct btrfs_root *tree_root = fs_info->tree_root;
9070         struct btrfs_root_item *root_item = &root->root_item;
9071         struct walk_control *wc;
9072         struct btrfs_key key;
9073         int err = 0;
9074         int ret;
9075         int level;
9076         bool root_dropped = false;
9077
9078         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9079
9080         path = btrfs_alloc_path();
9081         if (!path) {
9082                 err = -ENOMEM;
9083                 goto out;
9084         }
9085
9086         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9087         if (!wc) {
9088                 btrfs_free_path(path);
9089                 err = -ENOMEM;
9090                 goto out;
9091         }
9092
9093         trans = btrfs_start_transaction(tree_root, 0);
9094         if (IS_ERR(trans)) {
9095                 err = PTR_ERR(trans);
9096                 goto out_free;
9097         }
9098
9099         if (block_rsv)
9100                 trans->block_rsv = block_rsv;
9101
9102         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9103                 level = btrfs_header_level(root->node);
9104                 path->nodes[level] = btrfs_lock_root_node(root);
9105                 btrfs_set_lock_blocking(path->nodes[level]);
9106                 path->slots[level] = 0;
9107                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9108                 memset(&wc->update_progress, 0,
9109                        sizeof(wc->update_progress));
9110         } else {
9111                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9112                 memcpy(&wc->update_progress, &key,
9113                        sizeof(wc->update_progress));
9114
9115                 level = root_item->drop_level;
9116                 BUG_ON(level == 0);
9117                 path->lowest_level = level;
9118                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9119                 path->lowest_level = 0;
9120                 if (ret < 0) {
9121                         err = ret;
9122                         goto out_end_trans;
9123                 }
9124                 WARN_ON(ret > 0);
9125
9126                 /*
9127                  * unlock our path, this is safe because only this
9128                  * function is allowed to delete this snapshot
9129                  */
9130                 btrfs_unlock_up_safe(path, 0);
9131
9132                 level = btrfs_header_level(root->node);
9133                 while (1) {
9134                         btrfs_tree_lock(path->nodes[level]);
9135                         btrfs_set_lock_blocking(path->nodes[level]);
9136                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9137
9138                         ret = btrfs_lookup_extent_info(trans, fs_info,
9139                                                 path->nodes[level]->start,
9140                                                 level, 1, &wc->refs[level],
9141                                                 &wc->flags[level]);
9142                         if (ret < 0) {
9143                                 err = ret;
9144                                 goto out_end_trans;
9145                         }
9146                         BUG_ON(wc->refs[level] == 0);
9147
9148                         if (level == root_item->drop_level)
9149                                 break;
9150
9151                         btrfs_tree_unlock(path->nodes[level]);
9152                         path->locks[level] = 0;
9153                         WARN_ON(wc->refs[level] != 1);
9154                         level--;
9155                 }
9156         }
9157
9158         wc->level = level;
9159         wc->shared_level = -1;
9160         wc->stage = DROP_REFERENCE;
9161         wc->update_ref = update_ref;
9162         wc->keep_locks = 0;
9163         wc->for_reloc = for_reloc;
9164         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9165
9166         while (1) {
9167
9168                 ret = walk_down_tree(trans, root, path, wc);
9169                 if (ret < 0) {
9170                         err = ret;
9171                         break;
9172                 }
9173
9174                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9175                 if (ret < 0) {
9176                         err = ret;
9177                         break;
9178                 }
9179
9180                 if (ret > 0) {
9181                         BUG_ON(wc->stage != DROP_REFERENCE);
9182                         break;
9183                 }
9184
9185                 if (wc->stage == DROP_REFERENCE) {
9186                         level = wc->level;
9187                         btrfs_node_key(path->nodes[level],
9188                                        &root_item->drop_progress,
9189                                        path->slots[level]);
9190                         root_item->drop_level = level;
9191                 }
9192
9193                 BUG_ON(wc->level == 0);
9194                 if (btrfs_should_end_transaction(trans) ||
9195                     (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9196                         ret = btrfs_update_root(trans, tree_root,
9197                                                 &root->root_key,
9198                                                 root_item);
9199                         if (ret) {
9200                                 btrfs_abort_transaction(trans, ret);
9201                                 err = ret;
9202                                 goto out_end_trans;
9203                         }
9204
9205                         btrfs_end_transaction_throttle(trans);
9206                         if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9207                                 btrfs_debug(fs_info,
9208                                             "drop snapshot early exit");
9209                                 err = -EAGAIN;
9210                                 goto out_free;
9211                         }
9212
9213                         trans = btrfs_start_transaction(tree_root, 0);
9214                         if (IS_ERR(trans)) {
9215                                 err = PTR_ERR(trans);
9216                                 goto out_free;
9217                         }
9218                         if (block_rsv)
9219                                 trans->block_rsv = block_rsv;
9220                 }
9221         }
9222         btrfs_release_path(path);
9223         if (err)
9224                 goto out_end_trans;
9225
9226         ret = btrfs_del_root(trans, fs_info, &root->root_key);
9227         if (ret) {
9228                 btrfs_abort_transaction(trans, ret);
9229                 err = ret;
9230                 goto out_end_trans;
9231         }
9232
9233         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9234                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9235                                       NULL, NULL);
9236                 if (ret < 0) {
9237                         btrfs_abort_transaction(trans, ret);
9238                         err = ret;
9239                         goto out_end_trans;
9240                 } else if (ret > 0) {
9241                         /* if we fail to delete the orphan item this time
9242                          * around, it'll get picked up the next time.
9243                          *
9244                          * The most common failure here is just -ENOENT.
9245                          */
9246                         btrfs_del_orphan_item(trans, tree_root,
9247                                               root->root_key.objectid);
9248                 }
9249         }
9250
9251         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9252                 btrfs_add_dropped_root(trans, root);
9253         } else {
9254                 free_extent_buffer(root->node);
9255                 free_extent_buffer(root->commit_root);
9256                 btrfs_put_fs_root(root);
9257         }
9258         root_dropped = true;
9259 out_end_trans:
9260         btrfs_end_transaction_throttle(trans);
9261 out_free:
9262         kfree(wc);
9263         btrfs_free_path(path);
9264 out:
9265         /*
9266          * So if we need to stop dropping the snapshot for whatever reason we
9267          * need to make sure to add it back to the dead root list so that we
9268          * keep trying to do the work later.  This also cleans up roots if we
9269          * don't have it in the radix (like when we recover after a power fail
9270          * or unmount) so we don't leak memory.
9271          */
9272         if (!for_reloc && !root_dropped)
9273                 btrfs_add_dead_root(root);
9274         if (err && err != -EAGAIN)
9275                 btrfs_handle_fs_error(fs_info, err, NULL);
9276         return err;
9277 }
9278
9279 /*
9280  * drop subtree rooted at tree block 'node'.
9281  *
9282  * NOTE: this function will unlock and release tree block 'node'
9283  * only used by relocation code
9284  */
9285 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9286                         struct btrfs_root *root,
9287                         struct extent_buffer *node,
9288                         struct extent_buffer *parent)
9289 {
9290         struct btrfs_fs_info *fs_info = root->fs_info;
9291         struct btrfs_path *path;
9292         struct walk_control *wc;
9293         int level;
9294         int parent_level;
9295         int ret = 0;
9296         int wret;
9297
9298         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9299
9300         path = btrfs_alloc_path();
9301         if (!path)
9302                 return -ENOMEM;
9303
9304         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9305         if (!wc) {
9306                 btrfs_free_path(path);
9307                 return -ENOMEM;
9308         }
9309
9310         btrfs_assert_tree_locked(parent);
9311         parent_level = btrfs_header_level(parent);
9312         extent_buffer_get(parent);
9313         path->nodes[parent_level] = parent;
9314         path->slots[parent_level] = btrfs_header_nritems(parent);
9315
9316         btrfs_assert_tree_locked(node);
9317         level = btrfs_header_level(node);
9318         path->nodes[level] = node;
9319         path->slots[level] = 0;
9320         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9321
9322         wc->refs[parent_level] = 1;
9323         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9324         wc->level = level;
9325         wc->shared_level = -1;
9326         wc->stage = DROP_REFERENCE;
9327         wc->update_ref = 0;
9328         wc->keep_locks = 1;
9329         wc->for_reloc = 1;
9330         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9331
9332         while (1) {
9333                 wret = walk_down_tree(trans, root, path, wc);
9334                 if (wret < 0) {
9335                         ret = wret;
9336                         break;
9337                 }
9338
9339                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9340                 if (wret < 0)
9341                         ret = wret;
9342                 if (wret != 0)
9343                         break;
9344         }
9345
9346         kfree(wc);
9347         btrfs_free_path(path);
9348         return ret;
9349 }
9350
9351 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9352 {
9353         u64 num_devices;
9354         u64 stripped;
9355
9356         /*
9357          * if restripe for this chunk_type is on pick target profile and
9358          * return, otherwise do the usual balance
9359          */
9360         stripped = get_restripe_target(fs_info, flags);
9361         if (stripped)
9362                 return extended_to_chunk(stripped);
9363
9364         num_devices = fs_info->fs_devices->rw_devices;
9365
9366         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9367                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9368                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9369
9370         if (num_devices == 1) {
9371                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9372                 stripped = flags & ~stripped;
9373
9374                 /* turn raid0 into single device chunks */
9375                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9376                         return stripped;
9377
9378                 /* turn mirroring into duplication */
9379                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9380                              BTRFS_BLOCK_GROUP_RAID10))
9381                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9382         } else {
9383                 /* they already had raid on here, just return */
9384                 if (flags & stripped)
9385                         return flags;
9386
9387                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9388                 stripped = flags & ~stripped;
9389
9390                 /* switch duplicated blocks with raid1 */
9391                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9392                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9393
9394                 /* this is drive concat, leave it alone */
9395         }
9396
9397         return flags;
9398 }
9399
9400 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9401 {
9402         struct btrfs_space_info *sinfo = cache->space_info;
9403         u64 num_bytes;
9404         u64 min_allocable_bytes;
9405         int ret = -ENOSPC;
9406
9407         /*
9408          * We need some metadata space and system metadata space for
9409          * allocating chunks in some corner cases until we force to set
9410          * it to be readonly.
9411          */
9412         if ((sinfo->flags &
9413              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9414             !force)
9415                 min_allocable_bytes = SZ_1M;
9416         else
9417                 min_allocable_bytes = 0;
9418
9419         spin_lock(&sinfo->lock);
9420         spin_lock(&cache->lock);
9421
9422         if (cache->ro) {
9423                 cache->ro++;
9424                 ret = 0;
9425                 goto out;
9426         }
9427
9428         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9429                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9430
9431         if (btrfs_space_info_used(sinfo, true) + num_bytes +
9432             min_allocable_bytes <= sinfo->total_bytes) {
9433                 sinfo->bytes_readonly += num_bytes;
9434                 cache->ro++;
9435                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9436                 ret = 0;
9437         }
9438 out:
9439         spin_unlock(&cache->lock);
9440         spin_unlock(&sinfo->lock);
9441         return ret;
9442 }
9443
9444 int btrfs_inc_block_group_ro(struct btrfs_fs_info *fs_info,
9445                              struct btrfs_block_group_cache *cache)
9446
9447 {
9448         struct btrfs_trans_handle *trans;
9449         u64 alloc_flags;
9450         int ret;
9451
9452 again:
9453         trans = btrfs_join_transaction(fs_info->extent_root);
9454         if (IS_ERR(trans))
9455                 return PTR_ERR(trans);
9456
9457         /*
9458          * we're not allowed to set block groups readonly after the dirty
9459          * block groups cache has started writing.  If it already started,
9460          * back off and let this transaction commit
9461          */
9462         mutex_lock(&fs_info->ro_block_group_mutex);
9463         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9464                 u64 transid = trans->transid;
9465
9466                 mutex_unlock(&fs_info->ro_block_group_mutex);
9467                 btrfs_end_transaction(trans);
9468
9469                 ret = btrfs_wait_for_commit(fs_info, transid);
9470                 if (ret)
9471                         return ret;
9472                 goto again;
9473         }
9474
9475         /*
9476          * if we are changing raid levels, try to allocate a corresponding
9477          * block group with the new raid level.
9478          */
9479         alloc_flags = update_block_group_flags(fs_info, cache->flags);
9480         if (alloc_flags != cache->flags) {
9481                 ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9482                                      CHUNK_ALLOC_FORCE);
9483                 /*
9484                  * ENOSPC is allowed here, we may have enough space
9485                  * already allocated at the new raid level to
9486                  * carry on
9487                  */
9488                 if (ret == -ENOSPC)
9489                         ret = 0;
9490                 if (ret < 0)
9491                         goto out;
9492         }
9493
9494         ret = inc_block_group_ro(cache, 0);
9495         if (!ret)
9496                 goto out;
9497         alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9498         ret = do_chunk_alloc(trans, fs_info, alloc_flags,
9499                              CHUNK_ALLOC_FORCE);
9500         if (ret < 0)
9501                 goto out;
9502         ret = inc_block_group_ro(cache, 0);
9503 out:
9504         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9505                 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9506                 mutex_lock(&fs_info->chunk_mutex);
9507                 check_system_chunk(trans, fs_info, alloc_flags);
9508                 mutex_unlock(&fs_info->chunk_mutex);
9509         }
9510         mutex_unlock(&fs_info->ro_block_group_mutex);
9511
9512         btrfs_end_transaction(trans);
9513         return ret;
9514 }
9515
9516 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9517                             struct btrfs_fs_info *fs_info, u64 type)
9518 {
9519         u64 alloc_flags = get_alloc_profile(fs_info, type);
9520
9521         return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE);
9522 }
9523
9524 /*
9525  * helper to account the unused space of all the readonly block group in the
9526  * space_info. takes mirrors into account.
9527  */
9528 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9529 {
9530         struct btrfs_block_group_cache *block_group;
9531         u64 free_bytes = 0;
9532         int factor;
9533
9534         /* It's df, we don't care if it's racy */
9535         if (list_empty(&sinfo->ro_bgs))
9536                 return 0;
9537
9538         spin_lock(&sinfo->lock);
9539         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9540                 spin_lock(&block_group->lock);
9541
9542                 if (!block_group->ro) {
9543                         spin_unlock(&block_group->lock);
9544                         continue;
9545                 }
9546
9547                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9548                                           BTRFS_BLOCK_GROUP_RAID10 |
9549                                           BTRFS_BLOCK_GROUP_DUP))
9550                         factor = 2;
9551                 else
9552                         factor = 1;
9553
9554                 free_bytes += (block_group->key.offset -
9555                                btrfs_block_group_used(&block_group->item)) *
9556                                factor;
9557
9558                 spin_unlock(&block_group->lock);
9559         }
9560         spin_unlock(&sinfo->lock);
9561
9562         return free_bytes;
9563 }
9564
9565 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9566 {
9567         struct btrfs_space_info *sinfo = cache->space_info;
9568         u64 num_bytes;
9569
9570         BUG_ON(!cache->ro);
9571
9572         spin_lock(&sinfo->lock);
9573         spin_lock(&cache->lock);
9574         if (!--cache->ro) {
9575                 num_bytes = cache->key.offset - cache->reserved -
9576                             cache->pinned - cache->bytes_super -
9577                             btrfs_block_group_used(&cache->item);
9578                 sinfo->bytes_readonly -= num_bytes;
9579                 list_del_init(&cache->ro_list);
9580         }
9581         spin_unlock(&cache->lock);
9582         spin_unlock(&sinfo->lock);
9583 }
9584
9585 /*
9586  * checks to see if its even possible to relocate this block group.
9587  *
9588  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9589  * ok to go ahead and try.
9590  */
9591 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9592 {
9593         struct btrfs_root *root = fs_info->extent_root;
9594         struct btrfs_block_group_cache *block_group;
9595         struct btrfs_space_info *space_info;
9596         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9597         struct btrfs_device *device;
9598         struct btrfs_trans_handle *trans;
9599         u64 min_free;
9600         u64 dev_min = 1;
9601         u64 dev_nr = 0;
9602         u64 target;
9603         int debug;
9604         int index;
9605         int full = 0;
9606         int ret = 0;
9607
9608         debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9609
9610         block_group = btrfs_lookup_block_group(fs_info, bytenr);
9611
9612         /* odd, couldn't find the block group, leave it alone */
9613         if (!block_group) {
9614                 if (debug)
9615                         btrfs_warn(fs_info,
9616                                    "can't find block group for bytenr %llu",
9617                                    bytenr);
9618                 return -1;
9619         }
9620
9621         min_free = btrfs_block_group_used(&block_group->item);
9622
9623         /* no bytes used, we're good */
9624         if (!min_free)
9625                 goto out;
9626
9627         space_info = block_group->space_info;
9628         spin_lock(&space_info->lock);
9629
9630         full = space_info->full;
9631
9632         /*
9633          * if this is the last block group we have in this space, we can't
9634          * relocate it unless we're able to allocate a new chunk below.
9635          *
9636          * Otherwise, we need to make sure we have room in the space to handle
9637          * all of the extents from this block group.  If we can, we're good
9638          */
9639         if ((space_info->total_bytes != block_group->key.offset) &&
9640             (btrfs_space_info_used(space_info, false) + min_free <
9641              space_info->total_bytes)) {
9642                 spin_unlock(&space_info->lock);
9643                 goto out;
9644         }
9645         spin_unlock(&space_info->lock);
9646
9647         /*
9648          * ok we don't have enough space, but maybe we have free space on our
9649          * devices to allocate new chunks for relocation, so loop through our
9650          * alloc devices and guess if we have enough space.  if this block
9651          * group is going to be restriped, run checks against the target
9652          * profile instead of the current one.
9653          */
9654         ret = -1;
9655
9656         /*
9657          * index:
9658          *      0: raid10
9659          *      1: raid1
9660          *      2: dup
9661          *      3: raid0
9662          *      4: single
9663          */
9664         target = get_restripe_target(fs_info, block_group->flags);
9665         if (target) {
9666                 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9667         } else {
9668                 /*
9669                  * this is just a balance, so if we were marked as full
9670                  * we know there is no space for a new chunk
9671                  */
9672                 if (full) {
9673                         if (debug)
9674                                 btrfs_warn(fs_info,
9675                                            "no space to alloc new chunk for block group %llu",
9676                                            block_group->key.objectid);
9677                         goto out;
9678                 }
9679
9680                 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9681         }
9682
9683         if (index == BTRFS_RAID_RAID10) {
9684                 dev_min = 4;
9685                 /* Divide by 2 */
9686                 min_free >>= 1;
9687         } else if (index == BTRFS_RAID_RAID1) {
9688                 dev_min = 2;
9689         } else if (index == BTRFS_RAID_DUP) {
9690                 /* Multiply by 2 */
9691                 min_free <<= 1;
9692         } else if (index == BTRFS_RAID_RAID0) {
9693                 dev_min = fs_devices->rw_devices;
9694                 min_free = div64_u64(min_free, dev_min);
9695         }
9696
9697         /* We need to do this so that we can look at pending chunks */
9698         trans = btrfs_join_transaction(root);
9699         if (IS_ERR(trans)) {
9700                 ret = PTR_ERR(trans);
9701                 goto out;
9702         }
9703
9704         mutex_lock(&fs_info->chunk_mutex);
9705         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9706                 u64 dev_offset;
9707
9708                 /*
9709                  * check to make sure we can actually find a chunk with enough
9710                  * space to fit our block group in.
9711                  */
9712                 if (device->total_bytes > device->bytes_used + min_free &&
9713                     !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9714                         ret = find_free_dev_extent(trans, device, min_free,
9715                                                    &dev_offset, NULL);
9716                         if (!ret)
9717                                 dev_nr++;
9718
9719                         if (dev_nr >= dev_min)
9720                                 break;
9721
9722                         ret = -1;
9723                 }
9724         }
9725         if (debug && ret == -1)
9726                 btrfs_warn(fs_info,
9727                            "no space to allocate a new chunk for block group %llu",
9728                            block_group->key.objectid);
9729         mutex_unlock(&fs_info->chunk_mutex);
9730         btrfs_end_transaction(trans);
9731 out:
9732         btrfs_put_block_group(block_group);
9733         return ret;
9734 }
9735
9736 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9737                                   struct btrfs_path *path,
9738                                   struct btrfs_key *key)
9739 {
9740         struct btrfs_root *root = fs_info->extent_root;
9741         int ret = 0;
9742         struct btrfs_key found_key;
9743         struct extent_buffer *leaf;
9744         int slot;
9745
9746         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9747         if (ret < 0)
9748                 goto out;
9749
9750         while (1) {
9751                 slot = path->slots[0];
9752                 leaf = path->nodes[0];
9753                 if (slot >= btrfs_header_nritems(leaf)) {
9754                         ret = btrfs_next_leaf(root, path);
9755                         if (ret == 0)
9756                                 continue;
9757                         if (ret < 0)
9758                                 goto out;
9759                         break;
9760                 }
9761                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9762
9763                 if (found_key.objectid >= key->objectid &&
9764                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9765                         struct extent_map_tree *em_tree;
9766                         struct extent_map *em;
9767
9768                         em_tree = &root->fs_info->mapping_tree.map_tree;
9769                         read_lock(&em_tree->lock);
9770                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9771                                                    found_key.offset);
9772                         read_unlock(&em_tree->lock);
9773                         if (!em) {
9774                                 btrfs_err(fs_info,
9775                         "logical %llu len %llu found bg but no related chunk",
9776                                           found_key.objectid, found_key.offset);
9777                                 ret = -ENOENT;
9778                         } else {
9779                                 ret = 0;
9780                         }
9781                         free_extent_map(em);
9782                         goto out;
9783                 }
9784                 path->slots[0]++;
9785         }
9786 out:
9787         return ret;
9788 }
9789
9790 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9791 {
9792         struct btrfs_block_group_cache *block_group;
9793         u64 last = 0;
9794
9795         while (1) {
9796                 struct inode *inode;
9797
9798                 block_group = btrfs_lookup_first_block_group(info, last);
9799                 while (block_group) {
9800                         spin_lock(&block_group->lock);
9801                         if (block_group->iref)
9802                                 break;
9803                         spin_unlock(&block_group->lock);
9804                         block_group = next_block_group(info, block_group);
9805                 }
9806                 if (!block_group) {
9807                         if (last == 0)
9808                                 break;
9809                         last = 0;
9810                         continue;
9811                 }
9812
9813                 inode = block_group->inode;
9814                 block_group->iref = 0;
9815                 block_group->inode = NULL;
9816                 spin_unlock(&block_group->lock);
9817                 ASSERT(block_group->io_ctl.inode == NULL);
9818                 iput(inode);
9819                 last = block_group->key.objectid + block_group->key.offset;
9820                 btrfs_put_block_group(block_group);
9821         }
9822 }
9823
9824 /*
9825  * Must be called only after stopping all workers, since we could have block
9826  * group caching kthreads running, and therefore they could race with us if we
9827  * freed the block groups before stopping them.
9828  */
9829 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9830 {
9831         struct btrfs_block_group_cache *block_group;
9832         struct btrfs_space_info *space_info;
9833         struct btrfs_caching_control *caching_ctl;
9834         struct rb_node *n;
9835
9836         down_write(&info->commit_root_sem);
9837         while (!list_empty(&info->caching_block_groups)) {
9838                 caching_ctl = list_entry(info->caching_block_groups.next,
9839                                          struct btrfs_caching_control, list);
9840                 list_del(&caching_ctl->list);
9841                 put_caching_control(caching_ctl);
9842         }
9843         up_write(&info->commit_root_sem);
9844
9845         spin_lock(&info->unused_bgs_lock);
9846         while (!list_empty(&info->unused_bgs)) {
9847                 block_group = list_first_entry(&info->unused_bgs,
9848                                                struct btrfs_block_group_cache,
9849                                                bg_list);
9850                 list_del_init(&block_group->bg_list);
9851                 btrfs_put_block_group(block_group);
9852         }
9853         spin_unlock(&info->unused_bgs_lock);
9854
9855         spin_lock(&info->block_group_cache_lock);
9856         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9857                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9858                                        cache_node);
9859                 rb_erase(&block_group->cache_node,
9860                          &info->block_group_cache_tree);
9861                 RB_CLEAR_NODE(&block_group->cache_node);
9862                 spin_unlock(&info->block_group_cache_lock);
9863
9864                 down_write(&block_group->space_info->groups_sem);
9865                 list_del(&block_group->list);
9866                 up_write(&block_group->space_info->groups_sem);
9867
9868                 /*
9869                  * We haven't cached this block group, which means we could
9870                  * possibly have excluded extents on this block group.
9871                  */
9872                 if (block_group->cached == BTRFS_CACHE_NO ||
9873                     block_group->cached == BTRFS_CACHE_ERROR)
9874                         free_excluded_extents(info, block_group);
9875
9876                 btrfs_remove_free_space_cache(block_group);
9877                 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9878                 ASSERT(list_empty(&block_group->dirty_list));
9879                 ASSERT(list_empty(&block_group->io_list));
9880                 ASSERT(list_empty(&block_group->bg_list));
9881                 ASSERT(atomic_read(&block_group->count) == 1);
9882                 btrfs_put_block_group(block_group);
9883
9884                 spin_lock(&info->block_group_cache_lock);
9885         }
9886         spin_unlock(&info->block_group_cache_lock);
9887
9888         /* now that all the block groups are freed, go through and
9889          * free all the space_info structs.  This is only called during
9890          * the final stages of unmount, and so we know nobody is
9891          * using them.  We call synchronize_rcu() once before we start,
9892          * just to be on the safe side.
9893          */
9894         synchronize_rcu();
9895
9896         release_global_block_rsv(info);
9897
9898         while (!list_empty(&info->space_info)) {
9899                 int i;
9900
9901                 space_info = list_entry(info->space_info.next,
9902                                         struct btrfs_space_info,
9903                                         list);
9904
9905                 /*
9906                  * Do not hide this behind enospc_debug, this is actually
9907                  * important and indicates a real bug if this happens.
9908                  */
9909                 if (WARN_ON(space_info->bytes_pinned > 0 ||
9910                             space_info->bytes_reserved > 0 ||
9911                             space_info->bytes_may_use > 0))
9912                         dump_space_info(info, space_info, 0, 0);
9913                 list_del(&space_info->list);
9914                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9915                         struct kobject *kobj;
9916                         kobj = space_info->block_group_kobjs[i];
9917                         space_info->block_group_kobjs[i] = NULL;
9918                         if (kobj) {
9919                                 kobject_del(kobj);
9920                                 kobject_put(kobj);
9921                         }
9922                 }
9923                 kobject_del(&space_info->kobj);
9924                 kobject_put(&space_info->kobj);
9925         }
9926         return 0;
9927 }
9928
9929 static void link_block_group(struct btrfs_block_group_cache *cache)
9930 {
9931         struct btrfs_space_info *space_info = cache->space_info;
9932         int index = btrfs_bg_flags_to_raid_index(cache->flags);
9933         bool first = false;
9934
9935         down_write(&space_info->groups_sem);
9936         if (list_empty(&space_info->block_groups[index]))
9937                 first = true;
9938         list_add_tail(&cache->list, &space_info->block_groups[index]);
9939         up_write(&space_info->groups_sem);
9940
9941         if (first) {
9942                 struct raid_kobject *rkobj;
9943                 int ret;
9944
9945                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9946                 if (!rkobj)
9947                         goto out_err;
9948                 rkobj->raid_type = index;
9949                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9950                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9951                                   "%s", get_raid_name(index));
9952                 if (ret) {
9953                         kobject_put(&rkobj->kobj);
9954                         goto out_err;
9955                 }
9956                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9957         }
9958
9959         return;
9960 out_err:
9961         btrfs_warn(cache->fs_info,
9962                    "failed to add kobject for block cache, ignoring");
9963 }
9964
9965 static struct btrfs_block_group_cache *
9966 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9967                                u64 start, u64 size)
9968 {
9969         struct btrfs_block_group_cache *cache;
9970
9971         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9972         if (!cache)
9973                 return NULL;
9974
9975         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9976                                         GFP_NOFS);
9977         if (!cache->free_space_ctl) {
9978                 kfree(cache);
9979                 return NULL;
9980         }
9981
9982         cache->key.objectid = start;
9983         cache->key.offset = size;
9984         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9985
9986         cache->fs_info = fs_info;
9987         cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
9988         set_free_space_tree_thresholds(cache);
9989
9990         atomic_set(&cache->count, 1);
9991         spin_lock_init(&cache->lock);
9992         init_rwsem(&cache->data_rwsem);
9993         INIT_LIST_HEAD(&cache->list);
9994         INIT_LIST_HEAD(&cache->cluster_list);
9995         INIT_LIST_HEAD(&cache->bg_list);
9996         INIT_LIST_HEAD(&cache->ro_list);
9997         INIT_LIST_HEAD(&cache->dirty_list);
9998         INIT_LIST_HEAD(&cache->io_list);
9999         btrfs_init_free_space_ctl(cache);
10000         atomic_set(&cache->trimming, 0);
10001         mutex_init(&cache->free_space_lock);
10002         btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
10003
10004         return cache;
10005 }
10006
10007 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10008 {
10009         struct btrfs_path *path;
10010         int ret;
10011         struct btrfs_block_group_cache *cache;
10012         struct btrfs_space_info *space_info;
10013         struct btrfs_key key;
10014         struct btrfs_key found_key;
10015         struct extent_buffer *leaf;
10016         int need_clear = 0;
10017         u64 cache_gen;
10018         u64 feature;
10019         int mixed;
10020
10021         feature = btrfs_super_incompat_flags(info->super_copy);
10022         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10023
10024         key.objectid = 0;
10025         key.offset = 0;
10026         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10027         path = btrfs_alloc_path();
10028         if (!path)
10029                 return -ENOMEM;
10030         path->reada = READA_FORWARD;
10031
10032         cache_gen = btrfs_super_cache_generation(info->super_copy);
10033         if (btrfs_test_opt(info, SPACE_CACHE) &&
10034             btrfs_super_generation(info->super_copy) != cache_gen)
10035                 need_clear = 1;
10036         if (btrfs_test_opt(info, CLEAR_CACHE))
10037                 need_clear = 1;
10038
10039         while (1) {
10040                 ret = find_first_block_group(info, path, &key);
10041                 if (ret > 0)
10042                         break;
10043                 if (ret != 0)
10044                         goto error;
10045
10046                 leaf = path->nodes[0];
10047                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10048
10049                 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10050                                                        found_key.offset);
10051                 if (!cache) {
10052                         ret = -ENOMEM;
10053                         goto error;
10054                 }
10055
10056                 if (need_clear) {
10057                         /*
10058                          * When we mount with old space cache, we need to
10059                          * set BTRFS_DC_CLEAR and set dirty flag.
10060                          *
10061                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10062                          *    truncate the old free space cache inode and
10063                          *    setup a new one.
10064                          * b) Setting 'dirty flag' makes sure that we flush
10065                          *    the new space cache info onto disk.
10066                          */
10067                         if (btrfs_test_opt(info, SPACE_CACHE))
10068                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10069                 }
10070
10071                 read_extent_buffer(leaf, &cache->item,
10072                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10073                                    sizeof(cache->item));
10074                 cache->flags = btrfs_block_group_flags(&cache->item);
10075                 if (!mixed &&
10076                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10077                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10078                         btrfs_err(info,
10079 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10080                                   cache->key.objectid);
10081                         ret = -EINVAL;
10082                         goto error;
10083                 }
10084
10085                 key.objectid = found_key.objectid + found_key.offset;
10086                 btrfs_release_path(path);
10087
10088                 /*
10089                  * We need to exclude the super stripes now so that the space
10090                  * info has super bytes accounted for, otherwise we'll think
10091                  * we have more space than we actually do.
10092                  */
10093                 ret = exclude_super_stripes(info, cache);
10094                 if (ret) {
10095                         /*
10096                          * We may have excluded something, so call this just in
10097                          * case.
10098                          */
10099                         free_excluded_extents(info, cache);
10100                         btrfs_put_block_group(cache);
10101                         goto error;
10102                 }
10103
10104                 /*
10105                  * check for two cases, either we are full, and therefore
10106                  * don't need to bother with the caching work since we won't
10107                  * find any space, or we are empty, and we can just add all
10108                  * the space in and be done with it.  This saves us _alot_ of
10109                  * time, particularly in the full case.
10110                  */
10111                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10112                         cache->last_byte_to_unpin = (u64)-1;
10113                         cache->cached = BTRFS_CACHE_FINISHED;
10114                         free_excluded_extents(info, cache);
10115                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10116                         cache->last_byte_to_unpin = (u64)-1;
10117                         cache->cached = BTRFS_CACHE_FINISHED;
10118                         add_new_free_space(cache, info,
10119                                            found_key.objectid,
10120                                            found_key.objectid +
10121                                            found_key.offset);
10122                         free_excluded_extents(info, cache);
10123                 }
10124
10125                 ret = btrfs_add_block_group_cache(info, cache);
10126                 if (ret) {
10127                         btrfs_remove_free_space_cache(cache);
10128                         btrfs_put_block_group(cache);
10129                         goto error;
10130                 }
10131
10132                 trace_btrfs_add_block_group(info, cache, 0);
10133                 update_space_info(info, cache->flags, found_key.offset,
10134                                   btrfs_block_group_used(&cache->item),
10135                                   cache->bytes_super, &space_info);
10136
10137                 cache->space_info = space_info;
10138
10139                 link_block_group(cache);
10140
10141                 set_avail_alloc_bits(info, cache->flags);
10142                 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10143                         inc_block_group_ro(cache, 1);
10144                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10145                         spin_lock(&info->unused_bgs_lock);
10146                         /* Should always be true but just in case. */
10147                         if (list_empty(&cache->bg_list)) {
10148                                 btrfs_get_block_group(cache);
10149                                 list_add_tail(&cache->bg_list,
10150                                               &info->unused_bgs);
10151                         }
10152                         spin_unlock(&info->unused_bgs_lock);
10153                 }
10154         }
10155
10156         list_for_each_entry_rcu(space_info, &info->space_info, list) {
10157                 if (!(get_alloc_profile(info, space_info->flags) &
10158                       (BTRFS_BLOCK_GROUP_RAID10 |
10159                        BTRFS_BLOCK_GROUP_RAID1 |
10160                        BTRFS_BLOCK_GROUP_RAID5 |
10161                        BTRFS_BLOCK_GROUP_RAID6 |
10162                        BTRFS_BLOCK_GROUP_DUP)))
10163                         continue;
10164                 /*
10165                  * avoid allocating from un-mirrored block group if there are
10166                  * mirrored block groups.
10167                  */
10168                 list_for_each_entry(cache,
10169                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10170                                 list)
10171                         inc_block_group_ro(cache, 1);
10172                 list_for_each_entry(cache,
10173                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10174                                 list)
10175                         inc_block_group_ro(cache, 1);
10176         }
10177
10178         init_global_block_rsv(info);
10179         ret = 0;
10180 error:
10181         btrfs_free_path(path);
10182         return ret;
10183 }
10184
10185 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10186 {
10187         struct btrfs_fs_info *fs_info = trans->fs_info;
10188         struct btrfs_block_group_cache *block_group, *tmp;
10189         struct btrfs_root *extent_root = fs_info->extent_root;
10190         struct btrfs_block_group_item item;
10191         struct btrfs_key key;
10192         int ret = 0;
10193         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10194
10195         trans->can_flush_pending_bgs = false;
10196         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10197                 if (ret)
10198                         goto next;
10199
10200                 spin_lock(&block_group->lock);
10201                 memcpy(&item, &block_group->item, sizeof(item));
10202                 memcpy(&key, &block_group->key, sizeof(key));
10203                 spin_unlock(&block_group->lock);
10204
10205                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10206                                         sizeof(item));
10207                 if (ret)
10208                         btrfs_abort_transaction(trans, ret);
10209                 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid,
10210                                                key.offset);
10211                 if (ret)
10212                         btrfs_abort_transaction(trans, ret);
10213                 add_block_group_free_space(trans, fs_info, block_group);
10214                 /* already aborted the transaction if it failed. */
10215 next:
10216                 list_del_init(&block_group->bg_list);
10217         }
10218         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10219 }
10220
10221 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10222                            struct btrfs_fs_info *fs_info, u64 bytes_used,
10223                            u64 type, u64 chunk_offset, u64 size)
10224 {
10225         struct btrfs_block_group_cache *cache;
10226         int ret;
10227
10228         btrfs_set_log_full_commit(fs_info, trans);
10229
10230         cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10231         if (!cache)
10232                 return -ENOMEM;
10233
10234         btrfs_set_block_group_used(&cache->item, bytes_used);
10235         btrfs_set_block_group_chunk_objectid(&cache->item,
10236                                              BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10237         btrfs_set_block_group_flags(&cache->item, type);
10238
10239         cache->flags = type;
10240         cache->last_byte_to_unpin = (u64)-1;
10241         cache->cached = BTRFS_CACHE_FINISHED;
10242         cache->needs_free_space = 1;
10243         ret = exclude_super_stripes(fs_info, cache);
10244         if (ret) {
10245                 /*
10246                  * We may have excluded something, so call this just in
10247                  * case.
10248                  */
10249                 free_excluded_extents(fs_info, cache);
10250                 btrfs_put_block_group(cache);
10251                 return ret;
10252         }
10253
10254         add_new_free_space(cache, fs_info, chunk_offset, chunk_offset + size);
10255
10256         free_excluded_extents(fs_info, cache);
10257
10258 #ifdef CONFIG_BTRFS_DEBUG
10259         if (btrfs_should_fragment_free_space(cache)) {
10260                 u64 new_bytes_used = size - bytes_used;
10261
10262                 bytes_used += new_bytes_used >> 1;
10263                 fragment_free_space(cache);
10264         }
10265 #endif
10266         /*
10267          * Ensure the corresponding space_info object is created and
10268          * assigned to our block group. We want our bg to be added to the rbtree
10269          * with its ->space_info set.
10270          */
10271         cache->space_info = __find_space_info(fs_info, cache->flags);
10272         if (!cache->space_info) {
10273                 ret = create_space_info(fs_info, cache->flags,
10274                                        &cache->space_info);
10275                 if (ret) {
10276                         btrfs_remove_free_space_cache(cache);
10277                         btrfs_put_block_group(cache);
10278                         return ret;
10279                 }
10280         }
10281
10282         ret = btrfs_add_block_group_cache(fs_info, cache);
10283         if (ret) {
10284                 btrfs_remove_free_space_cache(cache);
10285                 btrfs_put_block_group(cache);
10286                 return ret;
10287         }
10288
10289         /*
10290          * Now that our block group has its ->space_info set and is inserted in
10291          * the rbtree, update the space info's counters.
10292          */
10293         trace_btrfs_add_block_group(fs_info, cache, 1);
10294         update_space_info(fs_info, cache->flags, size, bytes_used,
10295                                 cache->bytes_super, &cache->space_info);
10296         update_global_block_rsv(fs_info);
10297
10298         link_block_group(cache);
10299
10300         list_add_tail(&cache->bg_list, &trans->new_bgs);
10301
10302         set_avail_alloc_bits(fs_info, type);
10303         return 0;
10304 }
10305
10306 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10307 {
10308         u64 extra_flags = chunk_to_extended(flags) &
10309                                 BTRFS_EXTENDED_PROFILE_MASK;
10310
10311         write_seqlock(&fs_info->profiles_lock);
10312         if (flags & BTRFS_BLOCK_GROUP_DATA)
10313                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10314         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10315                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10316         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10317                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10318         write_sequnlock(&fs_info->profiles_lock);
10319 }
10320
10321 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10322                              struct btrfs_fs_info *fs_info, u64 group_start,
10323                              struct extent_map *em)
10324 {
10325         struct btrfs_root *root = fs_info->extent_root;
10326         struct btrfs_path *path;
10327         struct btrfs_block_group_cache *block_group;
10328         struct btrfs_free_cluster *cluster;
10329         struct btrfs_root *tree_root = fs_info->tree_root;
10330         struct btrfs_key key;
10331         struct inode *inode;
10332         struct kobject *kobj = NULL;
10333         int ret;
10334         int index;
10335         int factor;
10336         struct btrfs_caching_control *caching_ctl = NULL;
10337         bool remove_em;
10338
10339         block_group = btrfs_lookup_block_group(fs_info, group_start);
10340         BUG_ON(!block_group);
10341         BUG_ON(!block_group->ro);
10342
10343         /*
10344          * Free the reserved super bytes from this block group before
10345          * remove it.
10346          */
10347         free_excluded_extents(fs_info, block_group);
10348         btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10349                                   block_group->key.offset);
10350
10351         memcpy(&key, &block_group->key, sizeof(key));
10352         index = btrfs_bg_flags_to_raid_index(block_group->flags);
10353         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10354                                   BTRFS_BLOCK_GROUP_RAID1 |
10355                                   BTRFS_BLOCK_GROUP_RAID10))
10356                 factor = 2;
10357         else
10358                 factor = 1;
10359
10360         /* make sure this block group isn't part of an allocation cluster */
10361         cluster = &fs_info->data_alloc_cluster;
10362         spin_lock(&cluster->refill_lock);
10363         btrfs_return_cluster_to_free_space(block_group, cluster);
10364         spin_unlock(&cluster->refill_lock);
10365
10366         /*
10367          * make sure this block group isn't part of a metadata
10368          * allocation cluster
10369          */
10370         cluster = &fs_info->meta_alloc_cluster;
10371         spin_lock(&cluster->refill_lock);
10372         btrfs_return_cluster_to_free_space(block_group, cluster);
10373         spin_unlock(&cluster->refill_lock);
10374
10375         path = btrfs_alloc_path();
10376         if (!path) {
10377                 ret = -ENOMEM;
10378                 goto out;
10379         }
10380
10381         /*
10382          * get the inode first so any iput calls done for the io_list
10383          * aren't the final iput (no unlinks allowed now)
10384          */
10385         inode = lookup_free_space_inode(fs_info, block_group, path);
10386
10387         mutex_lock(&trans->transaction->cache_write_mutex);
10388         /*
10389          * make sure our free spache cache IO is done before remove the
10390          * free space inode
10391          */
10392         spin_lock(&trans->transaction->dirty_bgs_lock);
10393         if (!list_empty(&block_group->io_list)) {
10394                 list_del_init(&block_group->io_list);
10395
10396                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10397
10398                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10399                 btrfs_wait_cache_io(trans, block_group, path);
10400                 btrfs_put_block_group(block_group);
10401                 spin_lock(&trans->transaction->dirty_bgs_lock);
10402         }
10403
10404         if (!list_empty(&block_group->dirty_list)) {
10405                 list_del_init(&block_group->dirty_list);
10406                 btrfs_put_block_group(block_group);
10407         }
10408         spin_unlock(&trans->transaction->dirty_bgs_lock);
10409         mutex_unlock(&trans->transaction->cache_write_mutex);
10410
10411         if (!IS_ERR(inode)) {
10412                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10413                 if (ret) {
10414                         btrfs_add_delayed_iput(inode);
10415                         goto out;
10416                 }
10417                 clear_nlink(inode);
10418                 /* One for the block groups ref */
10419                 spin_lock(&block_group->lock);
10420                 if (block_group->iref) {
10421                         block_group->iref = 0;
10422                         block_group->inode = NULL;
10423                         spin_unlock(&block_group->lock);
10424                         iput(inode);
10425                 } else {
10426                         spin_unlock(&block_group->lock);
10427                 }
10428                 /* One for our lookup ref */
10429                 btrfs_add_delayed_iput(inode);
10430         }
10431
10432         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10433         key.offset = block_group->key.objectid;
10434         key.type = 0;
10435
10436         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10437         if (ret < 0)
10438                 goto out;
10439         if (ret > 0)
10440                 btrfs_release_path(path);
10441         if (ret == 0) {
10442                 ret = btrfs_del_item(trans, tree_root, path);
10443                 if (ret)
10444                         goto out;
10445                 btrfs_release_path(path);
10446         }
10447
10448         spin_lock(&fs_info->block_group_cache_lock);
10449         rb_erase(&block_group->cache_node,
10450                  &fs_info->block_group_cache_tree);
10451         RB_CLEAR_NODE(&block_group->cache_node);
10452
10453         if (fs_info->first_logical_byte == block_group->key.objectid)
10454                 fs_info->first_logical_byte = (u64)-1;
10455         spin_unlock(&fs_info->block_group_cache_lock);
10456
10457         down_write(&block_group->space_info->groups_sem);
10458         /*
10459          * we must use list_del_init so people can check to see if they
10460          * are still on the list after taking the semaphore
10461          */
10462         list_del_init(&block_group->list);
10463         if (list_empty(&block_group->space_info->block_groups[index])) {
10464                 kobj = block_group->space_info->block_group_kobjs[index];
10465                 block_group->space_info->block_group_kobjs[index] = NULL;
10466                 clear_avail_alloc_bits(fs_info, block_group->flags);
10467         }
10468         up_write(&block_group->space_info->groups_sem);
10469         if (kobj) {
10470                 kobject_del(kobj);
10471                 kobject_put(kobj);
10472         }
10473
10474         if (block_group->has_caching_ctl)
10475                 caching_ctl = get_caching_control(block_group);
10476         if (block_group->cached == BTRFS_CACHE_STARTED)
10477                 wait_block_group_cache_done(block_group);
10478         if (block_group->has_caching_ctl) {
10479                 down_write(&fs_info->commit_root_sem);
10480                 if (!caching_ctl) {
10481                         struct btrfs_caching_control *ctl;
10482
10483                         list_for_each_entry(ctl,
10484                                     &fs_info->caching_block_groups, list)
10485                                 if (ctl->block_group == block_group) {
10486                                         caching_ctl = ctl;
10487                                         refcount_inc(&caching_ctl->count);
10488                                         break;
10489                                 }
10490                 }
10491                 if (caching_ctl)
10492                         list_del_init(&caching_ctl->list);
10493                 up_write(&fs_info->commit_root_sem);
10494                 if (caching_ctl) {
10495                         /* Once for the caching bgs list and once for us. */
10496                         put_caching_control(caching_ctl);
10497                         put_caching_control(caching_ctl);
10498                 }
10499         }
10500
10501         spin_lock(&trans->transaction->dirty_bgs_lock);
10502         if (!list_empty(&block_group->dirty_list)) {
10503                 WARN_ON(1);
10504         }
10505         if (!list_empty(&block_group->io_list)) {
10506                 WARN_ON(1);
10507         }
10508         spin_unlock(&trans->transaction->dirty_bgs_lock);
10509         btrfs_remove_free_space_cache(block_group);
10510
10511         spin_lock(&block_group->space_info->lock);
10512         list_del_init(&block_group->ro_list);
10513
10514         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10515                 WARN_ON(block_group->space_info->total_bytes
10516                         < block_group->key.offset);
10517                 WARN_ON(block_group->space_info->bytes_readonly
10518                         < block_group->key.offset);
10519                 WARN_ON(block_group->space_info->disk_total
10520                         < block_group->key.offset * factor);
10521         }
10522         block_group->space_info->total_bytes -= block_group->key.offset;
10523         block_group->space_info->bytes_readonly -= block_group->key.offset;
10524         block_group->space_info->disk_total -= block_group->key.offset * factor;
10525
10526         spin_unlock(&block_group->space_info->lock);
10527
10528         memcpy(&key, &block_group->key, sizeof(key));
10529
10530         mutex_lock(&fs_info->chunk_mutex);
10531         if (!list_empty(&em->list)) {
10532                 /* We're in the transaction->pending_chunks list. */
10533                 free_extent_map(em);
10534         }
10535         spin_lock(&block_group->lock);
10536         block_group->removed = 1;
10537         /*
10538          * At this point trimming can't start on this block group, because we
10539          * removed the block group from the tree fs_info->block_group_cache_tree
10540          * so no one can't find it anymore and even if someone already got this
10541          * block group before we removed it from the rbtree, they have already
10542          * incremented block_group->trimming - if they didn't, they won't find
10543          * any free space entries because we already removed them all when we
10544          * called btrfs_remove_free_space_cache().
10545          *
10546          * And we must not remove the extent map from the fs_info->mapping_tree
10547          * to prevent the same logical address range and physical device space
10548          * ranges from being reused for a new block group. This is because our
10549          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10550          * completely transactionless, so while it is trimming a range the
10551          * currently running transaction might finish and a new one start,
10552          * allowing for new block groups to be created that can reuse the same
10553          * physical device locations unless we take this special care.
10554          *
10555          * There may also be an implicit trim operation if the file system
10556          * is mounted with -odiscard. The same protections must remain
10557          * in place until the extents have been discarded completely when
10558          * the transaction commit has completed.
10559          */
10560         remove_em = (atomic_read(&block_group->trimming) == 0);
10561         /*
10562          * Make sure a trimmer task always sees the em in the pinned_chunks list
10563          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10564          * before checking block_group->removed).
10565          */
10566         if (!remove_em) {
10567                 /*
10568                  * Our em might be in trans->transaction->pending_chunks which
10569                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10570                  * and so is the fs_info->pinned_chunks list.
10571                  *
10572                  * So at this point we must be holding the chunk_mutex to avoid
10573                  * any races with chunk allocation (more specifically at
10574                  * volumes.c:contains_pending_extent()), to ensure it always
10575                  * sees the em, either in the pending_chunks list or in the
10576                  * pinned_chunks list.
10577                  */
10578                 list_move_tail(&em->list, &fs_info->pinned_chunks);
10579         }
10580         spin_unlock(&block_group->lock);
10581
10582         if (remove_em) {
10583                 struct extent_map_tree *em_tree;
10584
10585                 em_tree = &fs_info->mapping_tree.map_tree;
10586                 write_lock(&em_tree->lock);
10587                 /*
10588                  * The em might be in the pending_chunks list, so make sure the
10589                  * chunk mutex is locked, since remove_extent_mapping() will
10590                  * delete us from that list.
10591                  */
10592                 remove_extent_mapping(em_tree, em);
10593                 write_unlock(&em_tree->lock);
10594                 /* once for the tree */
10595                 free_extent_map(em);
10596         }
10597
10598         mutex_unlock(&fs_info->chunk_mutex);
10599
10600         ret = remove_block_group_free_space(trans, fs_info, block_group);
10601         if (ret)
10602                 goto out;
10603
10604         btrfs_put_block_group(block_group);
10605         btrfs_put_block_group(block_group);
10606
10607         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10608         if (ret > 0)
10609                 ret = -EIO;
10610         if (ret < 0)
10611                 goto out;
10612
10613         ret = btrfs_del_item(trans, root, path);
10614 out:
10615         btrfs_free_path(path);
10616         return ret;
10617 }
10618
10619 struct btrfs_trans_handle *
10620 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10621                                      const u64 chunk_offset)
10622 {
10623         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10624         struct extent_map *em;
10625         struct map_lookup *map;
10626         unsigned int num_items;
10627
10628         read_lock(&em_tree->lock);
10629         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10630         read_unlock(&em_tree->lock);
10631         ASSERT(em && em->start == chunk_offset);
10632
10633         /*
10634          * We need to reserve 3 + N units from the metadata space info in order
10635          * to remove a block group (done at btrfs_remove_chunk() and at
10636          * btrfs_remove_block_group()), which are used for:
10637          *
10638          * 1 unit for adding the free space inode's orphan (located in the tree
10639          * of tree roots).
10640          * 1 unit for deleting the block group item (located in the extent
10641          * tree).
10642          * 1 unit for deleting the free space item (located in tree of tree
10643          * roots).
10644          * N units for deleting N device extent items corresponding to each
10645          * stripe (located in the device tree).
10646          *
10647          * In order to remove a block group we also need to reserve units in the
10648          * system space info in order to update the chunk tree (update one or
10649          * more device items and remove one chunk item), but this is done at
10650          * btrfs_remove_chunk() through a call to check_system_chunk().
10651          */
10652         map = em->map_lookup;
10653         num_items = 3 + map->num_stripes;
10654         free_extent_map(em);
10655
10656         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10657                                                            num_items, 1);
10658 }
10659
10660 /*
10661  * Process the unused_bgs list and remove any that don't have any allocated
10662  * space inside of them.
10663  */
10664 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10665 {
10666         struct btrfs_block_group_cache *block_group;
10667         struct btrfs_space_info *space_info;
10668         struct btrfs_trans_handle *trans;
10669         int ret = 0;
10670
10671         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10672                 return;
10673
10674         spin_lock(&fs_info->unused_bgs_lock);
10675         while (!list_empty(&fs_info->unused_bgs)) {
10676                 u64 start, end;
10677                 int trimming;
10678
10679                 block_group = list_first_entry(&fs_info->unused_bgs,
10680                                                struct btrfs_block_group_cache,
10681                                                bg_list);
10682                 list_del_init(&block_group->bg_list);
10683
10684                 space_info = block_group->space_info;
10685
10686                 if (ret || btrfs_mixed_space_info(space_info)) {
10687                         btrfs_put_block_group(block_group);
10688                         continue;
10689                 }
10690                 spin_unlock(&fs_info->unused_bgs_lock);
10691
10692                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10693
10694                 /* Don't want to race with allocators so take the groups_sem */
10695                 down_write(&space_info->groups_sem);
10696                 spin_lock(&block_group->lock);
10697                 if (block_group->reserved ||
10698                     btrfs_block_group_used(&block_group->item) ||
10699                     block_group->ro ||
10700                     list_is_singular(&block_group->list)) {
10701                         /*
10702                          * We want to bail if we made new allocations or have
10703                          * outstanding allocations in this block group.  We do
10704                          * the ro check in case balance is currently acting on
10705                          * this block group.
10706                          */
10707                         spin_unlock(&block_group->lock);
10708                         up_write(&space_info->groups_sem);
10709                         goto next;
10710                 }
10711                 spin_unlock(&block_group->lock);
10712
10713                 /* We don't want to force the issue, only flip if it's ok. */
10714                 ret = inc_block_group_ro(block_group, 0);
10715                 up_write(&space_info->groups_sem);
10716                 if (ret < 0) {
10717                         ret = 0;
10718                         goto next;
10719                 }
10720
10721                 /*
10722                  * Want to do this before we do anything else so we can recover
10723                  * properly if we fail to join the transaction.
10724                  */
10725                 trans = btrfs_start_trans_remove_block_group(fs_info,
10726                                                      block_group->key.objectid);
10727                 if (IS_ERR(trans)) {
10728                         btrfs_dec_block_group_ro(block_group);
10729                         ret = PTR_ERR(trans);
10730                         goto next;
10731                 }
10732
10733                 /*
10734                  * We could have pending pinned extents for this block group,
10735                  * just delete them, we don't care about them anymore.
10736                  */
10737                 start = block_group->key.objectid;
10738                 end = start + block_group->key.offset - 1;
10739                 /*
10740                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10741                  * btrfs_finish_extent_commit(). If we are at transaction N,
10742                  * another task might be running finish_extent_commit() for the
10743                  * previous transaction N - 1, and have seen a range belonging
10744                  * to the block group in freed_extents[] before we were able to
10745                  * clear the whole block group range from freed_extents[]. This
10746                  * means that task can lookup for the block group after we
10747                  * unpinned it from freed_extents[] and removed it, leading to
10748                  * a BUG_ON() at btrfs_unpin_extent_range().
10749                  */
10750                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10751                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10752                                   EXTENT_DIRTY);
10753                 if (ret) {
10754                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10755                         btrfs_dec_block_group_ro(block_group);
10756                         goto end_trans;
10757                 }
10758                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10759                                   EXTENT_DIRTY);
10760                 if (ret) {
10761                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10762                         btrfs_dec_block_group_ro(block_group);
10763                         goto end_trans;
10764                 }
10765                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10766
10767                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10768                 spin_lock(&space_info->lock);
10769                 spin_lock(&block_group->lock);
10770
10771                 space_info->bytes_pinned -= block_group->pinned;
10772                 space_info->bytes_readonly += block_group->pinned;
10773                 percpu_counter_add(&space_info->total_bytes_pinned,
10774                                    -block_group->pinned);
10775                 block_group->pinned = 0;
10776
10777                 spin_unlock(&block_group->lock);
10778                 spin_unlock(&space_info->lock);
10779
10780                 /* DISCARD can flip during remount */
10781                 trimming = btrfs_test_opt(fs_info, DISCARD);
10782
10783                 /* Implicit trim during transaction commit. */
10784                 if (trimming)
10785                         btrfs_get_block_group_trimming(block_group);
10786
10787                 /*
10788                  * Btrfs_remove_chunk will abort the transaction if things go
10789                  * horribly wrong.
10790                  */
10791                 ret = btrfs_remove_chunk(trans, fs_info,
10792                                          block_group->key.objectid);
10793
10794                 if (ret) {
10795                         if (trimming)
10796                                 btrfs_put_block_group_trimming(block_group);
10797                         goto end_trans;
10798                 }
10799
10800                 /*
10801                  * If we're not mounted with -odiscard, we can just forget
10802                  * about this block group. Otherwise we'll need to wait
10803                  * until transaction commit to do the actual discard.
10804                  */
10805                 if (trimming) {
10806                         spin_lock(&fs_info->unused_bgs_lock);
10807                         /*
10808                          * A concurrent scrub might have added us to the list
10809                          * fs_info->unused_bgs, so use a list_move operation
10810                          * to add the block group to the deleted_bgs list.
10811                          */
10812                         list_move(&block_group->bg_list,
10813                                   &trans->transaction->deleted_bgs);
10814                         spin_unlock(&fs_info->unused_bgs_lock);
10815                         btrfs_get_block_group(block_group);
10816                 }
10817 end_trans:
10818                 btrfs_end_transaction(trans);
10819 next:
10820                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10821                 btrfs_put_block_group(block_group);
10822                 spin_lock(&fs_info->unused_bgs_lock);
10823         }
10824         spin_unlock(&fs_info->unused_bgs_lock);
10825 }
10826
10827 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10828 {
10829         struct btrfs_space_info *space_info;
10830         struct btrfs_super_block *disk_super;
10831         u64 features;
10832         u64 flags;
10833         int mixed = 0;
10834         int ret;
10835
10836         disk_super = fs_info->super_copy;
10837         if (!btrfs_super_root(disk_super))
10838                 return -EINVAL;
10839
10840         features = btrfs_super_incompat_flags(disk_super);
10841         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10842                 mixed = 1;
10843
10844         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10845         ret = create_space_info(fs_info, flags, &space_info);
10846         if (ret)
10847                 goto out;
10848
10849         if (mixed) {
10850                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10851                 ret = create_space_info(fs_info, flags, &space_info);
10852         } else {
10853                 flags = BTRFS_BLOCK_GROUP_METADATA;
10854                 ret = create_space_info(fs_info, flags, &space_info);
10855                 if (ret)
10856                         goto out;
10857
10858                 flags = BTRFS_BLOCK_GROUP_DATA;
10859                 ret = create_space_info(fs_info, flags, &space_info);
10860         }
10861 out:
10862         return ret;
10863 }
10864
10865 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10866                                    u64 start, u64 end)
10867 {
10868         return unpin_extent_range(fs_info, start, end, false);
10869 }
10870
10871 /*
10872  * It used to be that old block groups would be left around forever.
10873  * Iterating over them would be enough to trim unused space.  Since we
10874  * now automatically remove them, we also need to iterate over unallocated
10875  * space.
10876  *
10877  * We don't want a transaction for this since the discard may take a
10878  * substantial amount of time.  We don't require that a transaction be
10879  * running, but we do need to take a running transaction into account
10880  * to ensure that we're not discarding chunks that were released in
10881  * the current transaction.
10882  *
10883  * Holding the chunks lock will prevent other threads from allocating
10884  * or releasing chunks, but it won't prevent a running transaction
10885  * from committing and releasing the memory that the pending chunks
10886  * list head uses.  For that, we need to take a reference to the
10887  * transaction.
10888  */
10889 static int btrfs_trim_free_extents(struct btrfs_device *device,
10890                                    u64 minlen, u64 *trimmed)
10891 {
10892         u64 start = 0, len = 0;
10893         int ret;
10894
10895         *trimmed = 0;
10896
10897         /* Not writeable = nothing to do. */
10898         if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10899                 return 0;
10900
10901         /* No free space = nothing to do. */
10902         if (device->total_bytes <= device->bytes_used)
10903                 return 0;
10904
10905         ret = 0;
10906
10907         while (1) {
10908                 struct btrfs_fs_info *fs_info = device->fs_info;
10909                 struct btrfs_transaction *trans;
10910                 u64 bytes;
10911
10912                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10913                 if (ret)
10914                         return ret;
10915
10916                 down_read(&fs_info->commit_root_sem);
10917
10918                 spin_lock(&fs_info->trans_lock);
10919                 trans = fs_info->running_transaction;
10920                 if (trans)
10921                         refcount_inc(&trans->use_count);
10922                 spin_unlock(&fs_info->trans_lock);
10923
10924                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10925                                                  &start, &len);
10926                 if (trans)
10927                         btrfs_put_transaction(trans);
10928
10929                 if (ret) {
10930                         up_read(&fs_info->commit_root_sem);
10931                         mutex_unlock(&fs_info->chunk_mutex);
10932                         if (ret == -ENOSPC)
10933                                 ret = 0;
10934                         break;
10935                 }
10936
10937                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10938                 up_read(&fs_info->commit_root_sem);
10939                 mutex_unlock(&fs_info->chunk_mutex);
10940
10941                 if (ret)
10942                         break;
10943
10944                 start += len;
10945                 *trimmed += bytes;
10946
10947                 if (fatal_signal_pending(current)) {
10948                         ret = -ERESTARTSYS;
10949                         break;
10950                 }
10951
10952                 cond_resched();
10953         }
10954
10955         return ret;
10956 }
10957
10958 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10959 {
10960         struct btrfs_block_group_cache *cache = NULL;
10961         struct btrfs_device *device;
10962         struct list_head *devices;
10963         u64 group_trimmed;
10964         u64 start;
10965         u64 end;
10966         u64 trimmed = 0;
10967         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10968         int ret = 0;
10969
10970         /*
10971          * try to trim all FS space, our block group may start from non-zero.
10972          */
10973         if (range->len == total_bytes)
10974                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10975         else
10976                 cache = btrfs_lookup_block_group(fs_info, range->start);
10977
10978         while (cache) {
10979                 if (cache->key.objectid >= (range->start + range->len)) {
10980                         btrfs_put_block_group(cache);
10981                         break;
10982                 }
10983
10984                 start = max(range->start, cache->key.objectid);
10985                 end = min(range->start + range->len,
10986                                 cache->key.objectid + cache->key.offset);
10987
10988                 if (end - start >= range->minlen) {
10989                         if (!block_group_cache_done(cache)) {
10990                                 ret = cache_block_group(cache, 0);
10991                                 if (ret) {
10992                                         btrfs_put_block_group(cache);
10993                                         break;
10994                                 }
10995                                 ret = wait_block_group_cache_done(cache);
10996                                 if (ret) {
10997                                         btrfs_put_block_group(cache);
10998                                         break;
10999                                 }
11000                         }
11001                         ret = btrfs_trim_block_group(cache,
11002                                                      &group_trimmed,
11003                                                      start,
11004                                                      end,
11005                                                      range->minlen);
11006
11007                         trimmed += group_trimmed;
11008                         if (ret) {
11009                                 btrfs_put_block_group(cache);
11010                                 break;
11011                         }
11012                 }
11013
11014                 cache = next_block_group(fs_info, cache);
11015         }
11016
11017         mutex_lock(&fs_info->fs_devices->device_list_mutex);
11018         devices = &fs_info->fs_devices->alloc_list;
11019         list_for_each_entry(device, devices, dev_alloc_list) {
11020                 ret = btrfs_trim_free_extents(device, range->minlen,
11021                                               &group_trimmed);
11022                 if (ret)
11023                         break;
11024
11025                 trimmed += group_trimmed;
11026         }
11027         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11028
11029         range->len = trimmed;
11030         return ret;
11031 }
11032
11033 /*
11034  * btrfs_{start,end}_write_no_snapshotting() are similar to
11035  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11036  * data into the page cache through nocow before the subvolume is snapshoted,
11037  * but flush the data into disk after the snapshot creation, or to prevent
11038  * operations while snapshotting is ongoing and that cause the snapshot to be
11039  * inconsistent (writes followed by expanding truncates for example).
11040  */
11041 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11042 {
11043         percpu_counter_dec(&root->subv_writers->counter);
11044         /*
11045          * Make sure counter is updated before we wake up waiters.
11046          */
11047         smp_mb();
11048         if (waitqueue_active(&root->subv_writers->wait))
11049                 wake_up(&root->subv_writers->wait);
11050 }
11051
11052 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11053 {
11054         if (atomic_read(&root->will_be_snapshotted))
11055                 return 0;
11056
11057         percpu_counter_inc(&root->subv_writers->counter);
11058         /*
11059          * Make sure counter is updated before we check for snapshot creation.
11060          */
11061         smp_mb();
11062         if (atomic_read(&root->will_be_snapshotted)) {
11063                 btrfs_end_write_no_snapshotting(root);
11064                 return 0;
11065         }
11066         return 1;
11067 }
11068
11069 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11070 {
11071         while (true) {
11072                 int ret;
11073
11074                 ret = btrfs_start_write_no_snapshotting(root);
11075                 if (ret)
11076                         break;
11077                 wait_on_atomic_t(&root->will_be_snapshotted, atomic_t_wait,
11078                                  TASK_UNINTERRUPTIBLE);
11079         }
11080 }