fs/btrfs/extent-tree.c

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